Tech Guides - Artificial Intelligence

Unity has undoubtedly been one of the leaders when it comes to developing cross-platform products - going from strengths to strengths in developing visually stimulating 2D as well as 3D games and simulations.With Artificial Intelligence revolutionizing the way games are being developed these days, Unity have identified the power of Machine Learning and introduced Unity Machine Learning Agents. With this, they plan on empowering the game developers and researchers in their quest to develop intelligent games, robotics and simulations. What are Unity Machine Learning Agents? Traditionally, game developers have been hard-coding the behaviour of the game agents. Although effective, this is a tedious task and it also limits the intelligence of the agents. Simply put, the agents are not smart enough. To overcome this obstacle, Unity have simplified the training process for the game developers and researchers by introducing Unity Machine Learning Agents (ML-Agents, in short). Through just a simple Python API, the game agents can be now trained to use deep reinforcement learning, an advanced form of machine learning, to learn from their actions and modify their behaviour accordingly. These agents can then be used to dynamically modify the difficulty of the game. How do they work? As mentioned earlier, the Unity ML-Agents are designed to work based on the concept of deep reinforcement learning, a branch of machine learning where the agents are trained to learn from their own actions. Here is a simple flowchart to demonstrate how reinforcement learning works: The reinforcement learning training cycle The learning environment to be configured for the ML-agents consists of 3 primary objects: Agent: Every agent has a unique set of states, observations and actions within the environment, and is assigned rewards for particular events. Brain: A brain decides what action any agent is supposed to take in a particular scenario. Think of it as a regular human brain, which basically controls the bodily functions. Academy: This object contains all the brains within the environment To train the agents, a variety of scenarios are made possible by varying the connection of different components (explained above) of the environment. Some are single agents, some simultaneous single agents, and others could be co-operative and competitive multi-agents and more. You can read more about these possibilities on the official Unity blog. Apart from the way these agents are trained, Unity are also adding some cool new features in these ML-agents. Some of these are: Monitoring the agents’ decision-making to make it more accurate Incorporating curriculum learning, by which the complexity of the tasks can eventually be increased to aid more effective learning Imitation learning is a newly-introduced feature wherein the agents simply mimic the actions we want them to perform, rather than they learning on their own. What next for Unity Machine Learning Agents? Unity recently announced the release of v0.3 beta SDK of the ML-agents, and have been making significant progress in this domain to develop smarter, more intelligent game agents which can be used with the Unity game engine. Still very much in the research phase, these agents can also be used as an example by academic researchers to study the complex behaviour of trained models in different environments and scenarios where the variables associated with the in-game physics and visual appearance can be altered. Going forward, these agents can also be used by enterprises for large scale simulations, in robotics and also in the development of autonomous vehicles. These are interesting times for game developers, and Unity in particular, in their quest for developing smarter, cutting-edge games. Inclusion of machine learning in their game development strategy is a terrific move, although it will take some time for this to be perfected and incorporated seamlessly. Nonetheless, all the research and innovation being put into this direction certainly seems well worth it!

Games these days are closer to reality than ever. Life-like graphics, smart gameplay and realistic machine-human interactions have led to major game studios up the ante when it comes to adopting the latest and most up to date tech for developing games. In fact, not so long ago, we shared with you a few interesting ways in which Artificial Intelligence is transforming the gaming industry. Inclusion of deep learning in games has emerged as one popular solution to make the games smarter. Deep learning can be used to enhance the realism and excitement in games by teaching the game agents how to behave more accurately, and in a more life-like manner. We recently came up with this interesting implementation of deep learning to to play the game of FIFA 18, and we were quite impressed! Using just 2 layers of neural networks and with a limited amount of training, the bot that was developed managed to learn the basic rules of football (soccer). Not just that, it was also able to perform the basic movements and tasks in the game correctly. To achieve this, 2 neural networks were developed - a Convolutional Neural Network to detect objects within the game, and a second layer of LSTM (Long Short Term Memory) network to specify the movements accordingly. The same user also managed to leverage deep learning to improve the in-game graphics of the FIFA 18 game. Using the deepfakes algorithm, he managed to swap the in-game face of one of the players with the player’s real-life face. The reason? The in-game faces, although quite realistic, could be better and more realistic. The experiment ended up being near perfect, as the resultant face that was created was as good as perfect. How did he do it? After gathering some training data which was basically some images of players scraped off Google, the user managed to train two autoencoders which learnt the distinction between the in-game face and the real-world face. Then, using the deepfakes algorithm, the inputs were reversed, recreating the real-world face in the game itself. The difference is quite astonishing: Apart from improving the gameplay and the in-game character graphics, deep learning can also be used to enhance the way the opponents/adversaries interact with the player in the game. If we take the example of the FIFA game mentioned before, deep learning can be used to enhance the behaviour and appearance of the in-game crowd, who can react or cheer their team better as per their team’s performance. How can Deep Learning benefit video games? The following are some of the clear advantages of implementing deep learning techniques in games: Highly accurate results can be achieved with more and more training data Manual intervention is minimal Game developers can focus on effective storytelling than on the in-game graphics Another obvious question comes to mind at this stage, however. What are the drawbacks of implementing deep learning for games? A few come to mind immediately: Complexity of the training models can be quite high Images in games need to be generated in real-time which is quite a challenge The computation time can be quite significant The training dataset for achieving accurate results can be quite humongous With advancements in technology and better, faster hardware, many of the current limitations in developing smarter games  can be overcome. Fast generative models can look into the real-time generation of images, while faster graphic cards can take care of the model computation issue. All in all, dabbling with deep learning in games seems worth the punt which game studios should definitely think of taking. What do you think? Is incorporating deep learning techniques in games a scalable idea?

The FAT* 2018 Conference on Fairness, Accountability, and Transparency is a first-of-its-kind international and interdisciplinary peer-reviewed conference that seeks to publish and present work examining the fairness, accountability, and transparency of algorithmic systems. This article covers research papers dedicated to 1st Session on Online discrimination and Privacy. FAT*  hosted the presentation of research work from a wide variety of disciplines, including computer science, statistics, the social sciences, and law. It took place on February 23 and 24, 2018, at the New York University Law School, in cooperation with its Technology Law and Policy Clinic. The conference brought together over 450 attendees from academic researchers, policymakers, and Machine learning practitioners. It witnessed 17 research papers, 6 tutorials, and 2 keynote presentations from leading experts in the field.  Session 1 explored ways in which online discrimination can happen and privacy could be compromised. The papers presented look for novel and practical solutions to some of the problems identified. We attempt to introduce our readers to the papers that will be presented at FAT* 2018 in this area thereby summarising the key challenges and questions explored by leading minds on the topic and their proposed potential answers to those issues. Session Chair: Joshua Kroll (University of California, Berkeley) Paper 1: Potential for Discrimination in Online Targeted Advertising Problems identified in the paper: Much recent work has focused on detecting instances of discrimination in online services ranging from discriminatory pricing on e-commerce and travel sites like Staples (Mikians et al., 2012) and Hotels.com (Hannák et al., 2014) to discriminatory prioritization of service requests and offerings from certain users over others in crowdsourcing and social networking sites like TaskRabbit (Hannák et al., 2017). In this paper, we focus on the potential for discrimination in online advertising, which underpins much of the Internet’s economy. Specifically, we focus on targeted advertising, where ads are shown only to a subset of users that have attributes (features) selected by the advertiser. Key Takeaways: A malicious advertiser can create highly discriminatory ads without using sensitive attributes such as gender or race. The current methods used to counter the problem are insufficient. The potential for discrimination in targeted advertising arises from the ability of an advertiser to use the extensive personal (demographic, behavioral, and interests) data that ad platforms gather about their users to target their ads. Different targeting methods offered by Facebook: attribute-based targeting, PII-based (custom audience) targeting and Look-alike audience targeting Three basic approaches to quantifying discrimination and their tradeoffs: Based on advertiser’s intent Based on ad targeting process Based on the targeted audience (outcomes) Paper 2: Discrimination in Online Personalization: A Multidisciplinary Inquiry The authors explore ways in which discrimination may arise in the targeting of job-related advertising, noting the potential for multiple parties to contribute to its occurrence. They then examine the statutes and case law interpreting the prohibition on advertisements that indicate a preference based on protected class and consider its application to online advertising. This paper provides a legal analysis of a real case, which found that simulated users selecting a gender in Google’s Ad Settings produces employment-related advertisements differing rates along gender lines despite identical web browsing patterns. Key Takeaways: The authors’ analysis of existing case law concludes that Section 230 may not immunize advertising platforms from liability under the FHA for algorithmic targeting of advertisements that indicate a preference for or against a protected class. Possible causes of ad targeting: Targeting was fully a product of the advertiser selecting gender segmentation. Targeting was fully a product of machine learning—Google alone selects gender. Targeting was fully a product of the advertiser selecting keywords. Targeting was fully the product of the advertiser being outbid for women. Given the limited scope of Title VII the authors conclude that Google is unlikely to face liability on the facts presented by Datta et al. Thus, the advertising prohibition of Title VII, like the prohibitions on discriminatory employment practices, is ill-equipped to advance the aims of equal treatment in a world where algorithms play an increasing role in decision making. Paper 3: Privacy for All: Ensuring Fair and Equitable Privacy Protections In this position paper, the authors argue for applying recent research on ensuring sociotechnical systems are fair and non-discriminatory to the privacy protections those systems may provide. Just as algorithmic decision-making systems may have discriminatory outcomes even without explicit or deliberate discrimination, so also privacy regimes may disproportionately fail to protect vulnerable members of their target population, resulting in disparate impact with respect to the effectiveness of privacy protections. Key Takeaways: Research questions posed: Are technical or non-technical privacy protection schemes fair? When and how do privacy protection technologies or policies improve or impede the fairness of systems they affect? When and how do fairness-enhancing technologies or policies enhance or reduce the privacy protections of the people involved? Data linking can lead to deanonymization; live recommenders can also be attacked to leak information The authors propose a new definition for a fair privacy scheme: a privacy scheme is (group-)fair if the probability of failure and expected risk are statistically independent of the subject’s membership in a protected class.   If you have missed Session 2, Session 3, Session 4 and Session 5 of the FAT* 2018 Conference, we have got you covered.

Tommy’s a brilliant young man, who loves programming. He’s so occupied with computers that he hardly has any time for friends. Tommy programs a very intelligent robot called Polly, using Artificial Intelligence, so that he has someone to talk to. One day, Tommy gets hurt real bad about something and needs someone to talk to. He rushes home to talk to Polly and pours out his emotions to her. To his disappointment, Polly starts giving him advice like she does for any other thing. She doesn’t understand that he needs someone to “feel” what he’s feeling rather than rant away on what he should or shouldn’t be doing. He naturally feels disconnected from Polly. My Friend doesn’t get me Have you ever wondered what it would be like to have a robot as a friend? I’m thinking something along the lines of Siri. Siri’s pretty good at holding conversations and is quick witted too. But Siri can’t understand your feelings or emotions, neither can “she” feel anything herself. Are we missing that “personality” from the artificial beings that we’re creating? Even if you talk about chatbots, although we gain through convenience, we lose the emotional aspect, especially at a time when expressive communication is the most important. Do we really need it? I remember watching the Terminator, where Arnie asks John, “Why do you cry?” John finds it difficult to explain to him, why humans cry. The fact is though, that the machine actually understood there was something wrong with the human, thanks to the visual effects associated with crying. We’ve also seen some instances of robots or AI analysing sentiment through text processing as well. But how accurate is this? How would a machine know when a human is actually using sarcasm? What if John was faking it and could cry at the drop of a hat or he just happened to be chopping onions? That’s food for thought now. On the contrary, you might wonder though, do we really want our machines to start analysing our emotions? What if they take advantage of our emotional state? Well, that’s a bit of a far fetched thought and what we need to understand is that it’s necessary for robots to gauge a bit of our emotions to enhance the experience of interacting with them. There are several wonderful applications for such a technology. For instance, Marketing organisations could use applications that detect users facial expressions when they look at a new commercial to gauge their “interest”. It could also be used by law enforcement as a replacement to the polygraph. Another interesting use case would be to help autism affected individuals understand the emotions of others better. The combination of AI and EI could find a tonne of applications right from cars that can sense if the driver is tired or sleepy and prevent an accident by pulling over, to a fridge that can detect if you’re stressed and lock itself, to prevent you from binge eating! Recent Developments in Emotional Intelligence There are several developments happening from the past few years, in terms of building systems that understand emotions. Pepper, a Japanese robot, for instance, can tell feelings such as joy, sadness and anger, and respond by playing you a song. A couple of years ago, Microsoft released a tool, the Emotion API, that could breakdown a person’s emotions based only on their picture. Physiologists, Neurologists and Psychologists, have collaborated with engineers to find measurable indicators of human emotion that can be taught to computers to look out for. There are projects that have attempted to decode facial expressions, the pitch of our voices, biometric data such as heart rate and even our body language and muscle movements. Bronwyn van der Merwe, General Manager of Fjord in the Asia Pacific region revealed that big companies like Amazon, Google and Microsoft are hiring comedians and script writers in order to harness the human-like aspect of AI by inducing personality into their technologies. Jerry, Ellen, Chris, Russell...are you all listening? How it works Almost 40% of our emotions are conveyed through tone of voice and the rest is read through facial expressions and gestures we make. An enormous amount of data is collected from media content and other sources and is used as training data for algorithms to learn human facial expressions and speech. One type of learning used is Active Learning or human-assisted machine learning. This is a kind of supervised learning, where the learning algorithm is able to interactively query the user to obtain new data points or an output. Situations might exist where unlabeled data is plentiful but manually labeling the data is expensive. In such a scenario, learning algorithms can query the user for labels. Since the algorithm chooses the examples, the number of examples to learn a concept turns out to be lower than what is required for usual supervised learning. Another approach is to use Transfer Learning, a method that focuses on storing the knowledge that’s gained while solving one problem and then applying it to a different but related problem. For example, knowledge gained while learning to recognize fruits could apply when trying to recognize vegetables. This works by analysing a video for facial expressions and then transfering that learning to label speech modality. What’s under the hood of these machines? Powerful robots that are capable of understanding emotions would most certainly be running Neural Nets under the hood. Complementing the power of these Neural Nets are beefy CPUs and GPUs on the likes of the Nvidia Titan X GPU and Intel Nervana CPU chip. Last year at NIPS, amongst controversial body shots and loads of humour filled interactions, Kory Mathewson and Piotr Mirowski entertained audiences with A.L.Ex and Pyggy, two AI robots that have played alongside humans in over 30 shows. These robots introduce audiences to the “comedy of speech recognition errors” by blabbering away to each other as well as to humans. Built around a Recurrent Neural Network that’s trained on dialogue from thousands of films, A.L.Ex. communicates with human performers, audience participants, and spectators through speech recognition, voice synthesis, and video projection. A.L.E.x is written in Torch and Lua code and has a word vocabulary of 50,000 words that have been extracted from 102,916 movies and it is built on an RNN with Long-Short Term Memory architecture and 512 dimensional layers. The unconquered challenges today The way I see it, there are broadly 3 challenge areas that AI powered robots face in this day: Rationality and emotions: AI robots need to be fed with initial logic by humans, failing which, they cannot learn on their own. They may never have the level of rationality or the breadth of emotions to take decisions the way humans do. Intuition, strategic thinking and emotions: Machines are incapable of thinking into the future and taking decisions the way humans can. For example, not very far into the future, we might have an AI powered dating application that measures a subscriber’s interest level while chatting with someone. It might just rate the interest level lower, if the person is in a bad mood due to some other reason. It wouldn’t consider the reason behind the emotion and whether it was actually linked to the ongoing conversation. Spontaneity, empathy and emotions: It may be years before robots are capable of coming up with a plan B, the way humans do. Having a contingency plan and implementing it in an emotional crisis is something that AI fails at accomplishing. For example, if you’re angry at something and just want to be left alone, your companion robot might just follow what you say without understanding your underlying emotion, while an actual human would instantly empathise with your situation and rather try to be there for you. Bronwyn van der Merwe said, "As human beings, we have contextual understanding and we have empathy, and right now there isn't a lot of that built into AI. We do believe that in the future, the companies that are going to succeed will be those that can build into their technology that kind of an understanding". What’s in store for the future If you ask me, right now we’re on the highway to something really great. Yes, there are several aspects that are unclear about AI and robots making our lives easier vs disrupting them, but as time passes, science is fitting the pieces of the puzzle together to bring about positive changes in our lives. AI is improving on the emotional front as I write, although there are clearly miles to go. Companies like Affectiva are pioneering emotion recognition technology and are working hard to improve the way AI understands human emotions. Biggies like Microsoft had been working on bringing in emotional intelligence into their AI since before 2015 and have come a long way since then. Perhaps, in the next Terminator movie, Arnie might just comfort a weeping Sarah Connor, saying, “Don’t cry, Sarah dear, he’s not worth it”, or something of the sort. As a parting note and just for funsies, here’s a final question for you, “Can you imagine a point in the future when robots have such high levels of EQ, that some of us might consider choosing them as a partner over humans?”

In the paper, “MaskGAN: Better Text Generation via Filling in the ______”, Ian Goodfellow, along with William Fedus and Andrew M. Dai have proposed a way to improve sample quality using Generative Adversarial Networks (GANs), which explicitly trains the generator to produce high quality samples and have also shown a lot of success in image generation.  Ian Goodfellow is a Research scientist at Google Brain. His research interests lies in the fields of deep learning, machine learning security and privacy, and particularly in generative models. Ian Goodfellow is known as the father of Generative Adversarial Networks. He runs the Self-Organizing Conference on Machine Learning, which was founded at OpenAI in 2016. Generative Adversarial Networks (GANs) is an architecture for training generative models in an adversarial setup, with a generator generating images that is trying to fool a discriminator that is trained to discriminate between real and synthetic images. GANs have had a lot of success in producing more realistic images than other approaches but they have only seen limited use for text sequences. They were originally designed to output differentiable values, as such discrete language generation is challenging for them. The team of researchers, introduce an actor-critic conditional GAN that fills in missing text conditioned on the surrounding context. The paper also shows that this GAN produces more realistic text samples compared to a maximum likelihood trained model. MaskGAN: Better Text Generation via Filling in the _______ What problem is the paper attempting to solve? This paper highlights how text classification was traditionally done using Recurrent Neural Network models by sampling from a distribution that is conditioned on the previous word and a hidden state that consists of a representation of the words generated so far. These are typically trained with maximum likelihood in an approach known as teacher forcing. However, this method causes problems when, during sample generation, the model is often forced to condition on sequences that were never conditioned on at training time, which leads to unpredictable dynamics in the hidden state of the RNN. Also, methods such as Professor Forcing and Scheduled Sampling have been proposed to solve this issue, which work indirectly by either causing the hidden state dynamics to become predictable (Professor Forcing) or by randomly conditioning on sampled words at training time, however, they do not directly specify a cost function on the output of the RNN encouraging high sample quality. The method proposed in the paper is trying to solve problem of text generation with GANs, by a sensible combination of novel approaches. MaskGANs Paper summary This paper proposes to improve sample quality using Generative Adversarial Network (GANs), which explicitly trains the generator to produce high quality samples. The model is trained on a text fill-in-the-blank or in-filling task. In this task, portions of a body of text are deleted or redacted. The goal of the model is to then infill the missing portions of text so that it is indistinguishable from the original data. While in-filling text, the model operates autoregressively over the tokens it has thus far filled in, as in standard language modeling, while conditioning on the true known context. If the entire body of text is redacted, then this reduces to language modeling. The paper also shows qualitatively and quantitatively, evidence that this new proposed method produces more realistic text samples compared to a maximum likelihood trained model. Key Takeaways One can have a hold about what MaskGANs are, as this paper introduces a text generation model trained on in-filling (MaskGAN). The paper considers the actor-critic architecture in extremely large action spaces, new evaluation metrics, and the generation of synthetic training data. The proposed contiguous in-filling task i.e. MASKGAN, is a good approach to reduce mode collapse and help with training stability for textual GANs. The paper shows that MaskGAN samples on a larger dataset (IMDB reviews) is significantly better than the corresponding tuned MaskMLE model as shown by human evaluation. One can produce high-quality samples despite the MaskGAN model having much higher perplexity on the ground-truth test set Reviewer feedback summary/takeaways Overall Score: 21/30 Average Score: 7/10. Reviewers liked the overall idea behind the paper. They appreciated the benefits they received from context (left context and right context) by solving a "fill-in-the-blank" task at training time and translating this into text generation at test time. A reviewer also stated that experiments were well carried through and very thorough. A reviewer also commented that the importance of the MaskGAN mechanism has been highlighted and the description of the reinforcement learning training part has been clarified. However, with pros, the paper has also received some cons stating, There is a lot of pre-training required for the proposed architecture Generated texts are generally locally valid but not always valid globally It was not made very clear whether the discriminator also conditions on the unmasked sequence. A reviewer also stated that there were some unanswered questions such as Was pre-training done for the baseline as well? How was the masking done? How did you decide on the words to mask? Was this at random? Is it actually usable in place of ordinary LSTM (or RNN)-based generation?

The paper “Twin Networks: Matching the Future for Sequence Generation”, is written by Yoshua Bengio in collaboration with Dmitriy Serdyuk, Nan Rosemary Ke, Alessandro Sordoni, Adam Trischler, and Chris Pal. This paper proposes a simple technique for encouraging generative RNNs to plan ahead. To achieve this, the authors have presented a simple RNN model, which further has two separate networks--a forward and a backward--that run in opposite directions during training. The main motive to train them in opposite direction is the hypothesis that the states of the forward model should be able to predict the entire future sequence. Yoshua Bengio is a Canadian computer scientist.He is known for his work on artificial neural networks and deep learning. His main research ambition is to understand principles of learning that yield intelligence. Yoshua has been co-organizing the Learning Workshop with Yann Le Cun, with whom he has also created the International Conference on Representation Learning (ICLR), since the year 1999. Yoshua has also organized or co-organized numerous other events, principally the deep learning workshops and symposia at NIPS and ICML since 2007. The article talks about TwinNet i.e. Twin Networks, a method for training RNN architectures to better model the future in its internal state, which are supervised by another RNN modelling the future in reverse order. Twin Networks: Matching the Future for Sequence Generation What problem is the paper attempting to solve? Recurrent Neural Networks (RNNs) are the basis of state-of-art models for generating sequential data such as text and speech and are usually trained by teacher forcing. This corresponds to optimizing one-step ahead prediction. At present, there is no explicit bias toward planning in the training objective, the model may prefer to focus on the most recent tokens instead of capturing subtle long-term dependencies, which could contribute to global coherence. Local correlations are usually stronger than long-term dependencies and thus end up dominating the learning signal. This results obtained from this are, samples from RNNs tend to exhibit local coherence but lack meaningful global structure. Recent efforts to address this problem have involved augmenting RNNs with external memory, with unitary or hierarchical architectures, or with explicit planning mechanisms. Parallel efforts aim to prevent overfitting on strong local correlations by regularizing the states of the network, by applying dropout or penalizing various statistics. To solve this, the paper proposes TwinNet, a simple method for regularizing a recurrent neural network that encourages modeling those aspects of the past that are predictive of the long-term future. Paper summary This paper presents a simple technique which enables generative recurrent neural networks to plan ahead. A backward RNN is trained to generate a given sequence in a reverse order. The states of the forward model are implicitly forced to predict cotemporal states of the backward model. The paper empirically shows that this approach achieves 9% relative improvement for a speech recognition task, and also achieves significant improvement on a COCO caption generation task. Overall, the model is driven by the intuition: (a) The backward hidden states contain a summary of the future of the sequence (b) To predict the future more accurately, the model will have to form a better representation of the past. The paper also demonstrates the success of the TwinNet approach experimentally, through several conditional and unconditional generation tasks that include speech recognition, image captioning, language modelling, and sequential image generation. Key Takeaways The paper introduces a simple method for training generative recurrent networks that regularizes the hidden states of the network to anticipate future states The paper also provides an extensive evaluation of the proposed model on multiple tasks and concludes that it helps training and regularization for conditioned generation (speech recognition, image captioning) and for the unconditioned case (sequential MNIST, language modelling) As a deeper analysis, the paper includes a visualization, which includes the introduced cost and observes that it negatively correlates with the word frequency. Reviewer feedback summary Overall Score: 21/30 Average Score: 7/10 The reviewers stated that the paper presents a novel approach to regularize RNNs and give results on different datasets indicating wide range of application. However, based on our results, they said that further experimentation and extensive hyperparameter search is needed. Overall, the paper is detailed, simple to implement and positive empirical results support the described approach. The reviewers have also pointed out a few limitations which include: Major downside of the approach is the cost in terms of resources. The twin model requires large memory and takes longer to train (~ 2-4 times) while providing little improvement over the baseline. During evaluation we found that the attention twin model gives results like “a woman at table a with cake a”, where it forces the model to look like a sentence from the back side too. This might be the reason for low metric values observed in soft attention twin net model. The effect of twin net as a regularizer can be examined against other regularization strategies for comparison purposes.

This ICLR 2018 accepted paper, Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments, addresses the use of meta-learning to operate in non-stationary environments, represented as a Markov chain of distinct tasks. This paper is authored by Pieter Abbeel, Maruan Al-Shedivat, Trapit Bansal, Yura Burda, Ilya Sutskever, and Igor Mordatch. Pieter Abbeel is a professor at UC Berkeley since 2008. He was also a Research Scientist at OpenAI (2016-2017). His current research focuses on robotics and machine learning with particular focus on meta-learning and deep reinforcement learning. One of the other authors of this paper, Ilya Sutskever is the co-founder and Research Director of OpenAI. He was also a Research Scientist at the Google Brain Team for 3 years. Meta-Learning, or alternatively learning to learn, typically uses metadata to understand how automatic learning can become flexible in solving learning problems, i.e. to learn the learning algorithm itself. Continuous adaptation in real-world environments is quite essential for any learning agent and meta-learning approach is an appropriate choice for this task. This article will talk about one of the top accepted research papers in the field of meta-learning at the 6th annual ICLR conference scheduled to happen between April 30 - May 03, 2018. Using a gradient-based meta-learning algorithm for Nonstationary Environments What problem is the paper attempting to solve? Reinforcement Learning algorithms, although achieving impressive results ranging from playing games to applications in dialogue systems to robotics, are only limited to solving tasks in stationary environments. On the other hand, the real-world is often nonstationary either due to complexity, changes in the dynamics in the environment over the lifetime of a system, or presence of multiple learning actors. Nonstationarity breaks the standard assumptions and requires agents to continuously adapt, both at training and execution time, in order to succeed. The classical approaches to dealing with nonstationarity are usually based on context detection and tracking i.e., reacting to the already happened changes in the environment by continuously fine-tuning the policy. However, nonstationarity allows only for limited interaction before the properties of the environment change. Thus, it immediately puts learning into the few-shot regime and often renders simple fine-tuning methods impractical. In order to continuously learn and adapt from limited experience in nonstationary environments, the authors of this paper propose the learning-to-learn (or meta-learning) approach. Paper summary This paper proposes a gradient-based meta-learning algorithm suitable for continuous adaptation of RL agents in nonstationary environments. The agents meta-learn to anticipate the changes in the environment and update their policies accordingly. This method builds upon the previous work on gradient-based model-agnostic meta-learning (MAML) that has been shown successful in the few shot settings. Their algorithm re-derive MAML for multi-task reinforcement learning from a probabilistic perspective, and then extends it to dynamically changing tasks. This paper also considers the problem of continuous adaptation to a learning opponent in a competitive multi-agent setting and have designed RoboSumo—a 3D environment with simulated physics that allows pairs of agents to compete against each other. The paper answers the following questions: What is the behavior of different adaptation methods (in nonstationary locomotion and competitive multi-agent environments) when the interaction with the environment is strictly limited to one or very few episodes before it changes? What is the sample complexity of different methods, i.e., how many episodes are required for a method to successfully adapt to the changes? Additionally, it answers the following questions specific to the competitive multi-agent setting: Given a diverse population of agents that have been trained under the same curriculum, how do different adaptation methods rank in a competition versus each other? When the population of agents is evolved for several generations, what happens with the proportions of different agents in the population? Key Takeaways This work proposes a simple gradient-based meta-learning approach suitable for continuous adaptation in nonstationary environments. This method was applied to nonstationary locomotion and within a competitive multi-agent setting—the RoboSumo environment. The key idea of the method is to regard nonstationarity as a sequence of stationary tasks and train agents to exploit the dependencies between consecutive tasks such that they can handle similar nonstationarities at execution time. In both cases, i.e meta-learning algorithm and the multi-agent setting,  meta-learned adaptation rules were more efficient than the baselines in the few-shot regime. Additionally, agents that meta-learned to adapt, demonstrated the highest level of skill when competing in iterated games against each other. Reviewer feedback summary Overall Score: 24/30 Average Score: 8 The paper was termed as a great contribution to ICLR. According to the reviewers, the paper addressed a very important problem for general AI and was well-written. They also appreciated the careful experiment designs, and thorough comparisons making the results convincing. They found that editorial rigor and image quality could be better. However, there was no content related improvements suggested. The paper was appreciated for being dense and rich on rapid meta-learning.

The paper, Parameter space noise for exploration proposes parameter space noise as an efficient solution for exploration, a big problem for deep reinforcement learning. This paper is authored by Pieter Abbeel, Matthias Plappert, Rein Houthooft, Prafulla Dhariwal, Szymon Sidor, Richard Y. Chen, Xi Chen, Tamim Asfour, and Marcin Andrychowicz. Pieter Abbeel is currently a professor at UC Berkeley since 2008. He was also a Research Scientist at OpenAI (2016-2017). Pieter is one of the pioneers of deep reinforcement learning for robotics, including learning locomotion and visuomotor skills. His current research focuses on robotics and machine learning with particular focus on deep reinforcement learning, meta-learning, and AI safety. Deep reinforcement learning is the combination of deep learning with reinforcement learning to create artificial agents to achieve human-level performance across many challenging domains. This article will talk about one of Pieter’s top accepted research papers in the field of deep reinforcement learning at the 6th annual ICLR conference scheduled to happen between April 30 - May 03, 2018. Improving the exploratory behavior of Deep RL algorithms with Parameter Space Noise What problem is the paper attempting to solve? This paper is about the exploration challenge in deep reinforcement learning (RL) algorithms. The main purpose of exploration is to ensure that the agent’s behavior does not converge prematurely to a local optimum. Enabling efficient and effective exploration is difficult since it is not directed by the reward function of the underlying Markov decision process (MDP). A large number of methods have been proposed to tackle this challenge in high-dimensional and/or continuous-action MDPs. These methods increase the exploratory nature of these algorithms through the addition of temporally-correlated noise or through the addition of parameter noise. The main limitation of these methods is that they are either only proposed and evaluated for the on-policy setting with relatively small and shallow function approximators or disregard all temporal structure and gradient information. Paper summary This paper proposes adding noise to the parameters (parameter space noise) of a deep network when taking actions in deep reinforcement learning to encourage exploration. The effectiveness of this approach is demonstrated through empirical analysis across a variety of reinforcement learning tasks (i.e.DQN, DDPG, and TRPO). It answers the following questions: Do existing state-of-the-art RL algorithms benefit from incorporating parameter space noise? Does parameter space noise aid in exploring sparse reward environments more effectively? How does parameter space noise exploration compare against evolution strategies for deep policies with respect to sample efficiency? Key Takeaways The paper describes a method which proves parameter space noise as a conceptually simple yet effective replacement for traditional action space noise like -greedy and additive Gaussian noise. This work shows that parameter perturbations can successfully be combined with contemporary on- and off-policy deep RL algorithms such as DQN, DDPG, and TRPO and often results in improved performance compared to action noise. The paper attempts to prove with experiments that using parameter noise allows solving environments with very sparse rewards, in which action noise is unlikely to succeed. Parameter space noise is a viable and interesting alternative to action space noise, which is still the effective standard in most reinforcement learning applications. Reviewer feedback summary Overall Score: 20/30 Average Score: 6.66 The reviewers were pleased with the paper. They termed it as a simple strategy for exploration that is effective empirically. The paper was found to be clear and well written with thorough experiments across deep RL domains.  The authors have also released open-source code along with their paper for reproducibility, which was appreciated by the reviewers. However, a common trend among the reviews was that the authors overstated their claims and contributions.  The reviewers called out some statements in particular (e.g. the discussion of ES and RL). They also felt that the paper lacked a strong justification for the method other than it being empirically effective and intuitive.

Okay, you’re probably here because you’ve got just a few months to graduate and the projects section of your resume is blank. Or you’re just an inquisitive little nerd scraping the WWW for ways to crack that dream job. Either way, you’re not alone and there are ten thousand others trying to build a great Data Science portfolio to land them a good job. Look no further, we’ll try our best to help you on how to make a portfolio that catches the recruiter’s eye! David “Trent” Salazar‘s portfolio is a great example of a wholesome one and Sajal Sharma’s, is a good example of how one can display their Data Science Portfolios on a platform like Github. Companies are on the lookout for employees who can add value to the business. To showcase this on your resume effectively, the first step is to understand the different ways in which you can add value. 4 things you need to show in a data science portfolio Data science can be broken down into 4 broad areas: Obtaining insights from data and presenting them to the business leaders Designing an application that directly benefits the customer Designing an application or system that directly benefits other teams in the organisation Sharing expertise on data science with other teams You’ll need to ensure that your portfolio portrays all or at least most of the above, in order to easily make it through a job selection. So let’s see what we can do to make a great portfolio. Demonstrate that you know what you're doing So the idea is to show the recruiter that you’re capable of performing the critical aspects of Data Science, i.e. import a data set, clean the data, extract useful information from the data using various techniques, and finally visualise the findings and communicate them. Apart from the technical skills, there are a few soft skills that are expected as well. For instance, the ability to communicate and collaborate with others, the ability to reason and take the initiative when required. If your project is actually able to communicate these things, you’re in! Stay focused and be specific You might know a lot, but rather than throwing all your skills, projects and knowledge in the employer’s face, it’s always better to be focused on doing something and doing it right. Just as you’d do in your resume, keeping things short and sweet, you can implement this while building your portfolio too. Always remember, the interviewer is looking for specific skills. Research the data science job market Find 5-6 jobs, probably from Linkedin or Indeed, that interest you and go through their descriptions thoroughly. Understand what kind of skills the employer is looking for. For example, it could be classification, machine learning, statistical modeling or regression. Pick up the tools that are required for the job - for example, Python, R, TensorFlow, Hadoop, or whatever might get the job done. If you don’t know how to use that tool, you’ll want to skill-up as you work your way through the projects. Also, identify the kind of data that they would like you to be working on, like text or numerical, etc. Now, once you have this information at hand, start building your project around these skills and tools. Be a problem solver Working on projects that are not actual ‘problems’ that you’re solving, won’t stand out in your portfolio. The closer your projects are to the real-world, the easier it will be for the recruiter to make their decision to choose you. This will also showcase your analytical skills and how you’ve applied data science to solve a prevailing problem. Put at least 3 diverse projects in your data science portfolio A nice way to create a portfolio is to list 3 good projects that are diverse in nature. Here are some interesting projects to get you started on your portfolio: Data Cleaning and wrangling Data Cleaning is one of the most critical tasks that a data scientist performs. By taking a group of diverse data sets, consolidating and making sense of them, you’re giving the recruiter confidence that you know how to prep them for analysis. For example, you can take Twitter or Whatsapp data and clean it for analysis. The process is pretty simple; you first find a “dirty” data set, then spot an interesting angle to approach the data from, clean it up and perform analysis on it, and finally present your findings. Data storytelling Storytelling showcases not only your ability to draw insight from raw data, but it also reveals how well you’re able to convey the insights to others and persuade them. For example, you can use data from the bus system in your country and gather insights to identify which stops incur the most delays. This could be fixed by changing their route. Make sure your analysis is descriptive and your code and logic can be followed. Here’s what you do; first you find a good dataset, then you explore the data and spot correlations in the data. Then you visualize it before you start writing up your narrative. Tackle the data from various angles and pick up the most interesting one. If it’s interesting to you, it will most probably be interesting to anyone else who’s reviewing it. Break down and explain each step in detail, each code snippet, as if you were describing it to a friend. The idea is to teach the reviewer something new as you run through the analysis. End to end data science If you’re more into Machine Learning, or algorithm writing, you should do an end-to-end data science project. The project should be capable of taking in data, processing it and finally learning from it, every step of the way. For example, you can pick up fuel pricing data for your city or maybe stock market data. The data needs to be dynamic and updated regularly. The trick for this one is to keep the code simple so that it’s easy to set up and run. You first need to identify a good topic. Understand here that we will not be working with a single dataset, rather you will need to import and parse all the data and bring it under a single dataset yourself. Next, get the training and test data ready to make predictions. Document your code and other findings and you’re good to go. Prove you have the data science skill set If you want to get that job, you’ve got to have the appropriate tools to get the job done. Here’s a list of some of the most popular tools with a link to the right material for you to skill-up: Data science languages There's a number of key languages in data science that are essential. It might seem obvious, but making sure they're on your resume and demonstrated in your portfolio is incredibly important. Include things like: Python R Java Scala SQL Big Data tools If you're applying for big data roles, demonstrating your experience with the key technologies is a must. It not only proves you have the skills, but also shows that you have an awareness of what tools can be used to build a big data solution or project. You'll need: Hadoop, Spark Hive Machine learning frameworks With machine learning so in demand, if you can prove you've used a number of machine learning frameworks, you've already done a lot to impress. Remember, many organizations won't actually know as much about machine learning as you think. In fact, they might even be hiring you with a view to building out this capability. Remember to include: TensorFlow Caffe2 Keras PyTorch Data visualisation tools Data visualization is a crucial component of any data science project. If you can visualize and communicate data effectively, you're immediately demonstrating you're able to collaborate with others and make your insights accessible and useful to the wider business. Include tools like these in your resume and portfolio:  D3.js Excel chart  Tableau  ggplot2 So there you have it. You know what to do to build a decent data science portfolio. It’s really worth attending competitions and challenges. It will not only help you keep up to data and well oiled with your skills, but also give you a broader picture of what people are actually working on and with what tools they’re able to solve problems.

Python has become a big hit in the Data Science community over the last five years. So much so that it is slowly taking over R - the ‘lingua franca of statistics’ - as the preferred choice of tool for many. The recently published Stack Overflow Developer Survey 2018 suggests Python is the next big programming language, and its adoption in the industry is only going to increase. Python’s rise has been staggering, but not really surprising. Its general-purpose nature, coupled with the efficiency and ease of use make it easier for you to build your data science solutions without any hassle. You also have a rich suite of Python libraries available at your disposal for all your Data Science-related tasks - from basic web scraping to something as complex as training deep learning models. In this article, we take a look at some of the most popular and widely used Python libraries and their application areas. Web Scraping Web scraping is a popular information extraction technique from the web using the HTTP protocol, with the help of a web browser. The two most commonly used tools for web scraping are, unsurprisingly, Python-based. 1.Beautiful Soup Beautiful Soup is a popular Python library for extracting information out of the HTML and XML files. It provides a unique, easy way to navigate, search and modify the parsed data, potentially saving you hours of needless work. It works with both the versions of Python, i.e. 2.7 and 3.x and is very easy to use. Check out our latest tutorial on how to scrape web page using the Beautiful Soup. [box type="info" align="" class="" width=""]Editor's Tip: If you’re new to the concept of web scraping, Beautiful Soup should be your go-to library. You can learn more about how to use this library more efficiently in our book Python Web Scraping Cookbook [/box] 2.Scrapy Scrapy is a free, open source framework written in Python. Although developed for web scraping, it can also be used as a general web crawler and extract data using different APIs. Following the ‘Don’t Repeat Yourself’ philosophy of frameworks such as Django, Scrapy includes a set of self-contained crawlers, with each of them following specific instructions with a specific objective. [box type="info" align="" class="" width=""]Editor’s tip: To learn how to use Scrapy for your scraping projects, our book Python Web Scraping, Second Edition is definitely worth checking out. [/box] Scientific Computation and Data Analysis Arguably the most common data science tasks, Python proves to be of great worth to data scientists by providing unique libraries for data manipulation and analysis, as well as mathematical computation. 3. NumPy NumPy is the most popular library for scientific computing in Python and is a part of the larger Python stack for scientific computation called SciPy (discussed below). Apart from its uses in linear algebra and other mathematical functions, it can also be used as a multi-dimensional container, or array, of generic data with arbitrary data types. NumPy integrates seamlessly languages such as C/C++ and because of its support for multiple data types, it works well with a variety of databases as well. 4. SciPy SciPy is a Python-based framework containing open source libraries for mathematics, scientific computation and data analysis.  The SciPy library is a collection of algorithms and tools for advanced mathematical computations, statistics and much more. The SciPy stack consists of the following libraries: NumPy - Python package for numerical computation SciPy - One of the core packages of the SciPy stack for signal processing, optimization and advanced statistics matplotlib - Popular Python library for data visualization SymPy - Library for symbolic mathematics and algebra pandas - Python library for data manipulation and analysis iPython -  Interactive console to run Python-based code 5. pandas pandas is a widely used Python package providing data structures and tools for effective data manipulation and analysis. It is a popularly used tool for Quantitative Analysis and finds a lot of application in algorithmic trading and risk analysis. With a large community of dedicated users, pandas is regularly updated to get new API changes, performance updates and bug fixes. This is one library you definitely need to work with to truly realize its power. [box type="info" align="" class="" width=""]Editor's Tip: To get a more hands-on understanding of how to effectively use pandas for data analysis, make sure you check out our highly popular title pandas Cookbook.[/box] Machine Learning and Deep Learning Python trumps all other languages when it comes to implementing efficient machine learning and deep learning models, simply by virtue of its diverse, effective and easy to use set of libraries. It is worth having a look at the experts’ take on why Python is great for machine learning and Artificial Intelligence. In this section, we see some of the most popular and commonly used Python libraries for machine learning and deep learning: 6. Scikit-learn scikit-learn is the most popular Python library for data mining, analysis and machine learning. It is built using the capabilities of NumPy, SciPy and matplotlib, and is commercially usable. You can implement a variety of machine learning techniques such as classification, regression, clustering and more, using scikit-learn. It is very easy to install and has a clean, slick documentation for anyone looking to get started with it. [box type="info" align="" class="" width=""]Editor’s tip: To understand how to use scikit-learn in your machine learning projects, our bestselling book Python Machine Learning, Second Edition is all you need. If you’re looking to specifically master scikit-learn, Mastering Machine Learning with scikit-learn will prove to be a very useful resource. Check it out! [/box] 7. Tensorflow Tensorflow is the popular machine learning library everyone seems to be talking about today. It is a Python-based framework for effective machine learning and deep learning using multiple CPUs or GPUs. Backed by Google, it was initially developed by the research team of Google Brain, and is the widely used framework in the world for machine intelligence. It enjoys the support of a large community of active users and is finding widespread application for advanced machine learning across a multitude of industrial domains - from manufacturing and retail to healthcare and smart cars. If you are interested to know more about Tensorflow, you can quickly check out the tutorial here. [box type="info" align="" class="" width=""]Editor's Tip: Tensorflow being the most popular framework for machine learning and deep learning, it is one library you should definitely master. Check out the following books to skill up quickly! Machine Learning with TensorFlow 1.x TensorFlow Machine Learning Cookbook Deep Learning with TensorFlow Tensorflow 1.x Deep Learning Cookbook Mastering Tensorflow 1.x [/box] 8. Keras Keras is a Python-based neural networks API, and offers a simplified interface to train and deploy your deep learning models with ease. It has support for a variety of deep learning frameworks such as Tensorflow, Deeplearning4j and CNTK. Keras is very user-friendly, follows a modular approach and supports both CPU and GPU-based computations. If you want to make the deep learning process simpler and effective, this library is definitely worth checking out! [box type="info" align="" class="" width=""]Editor's Tip: If you’re looking for a resource that teaches you how to use Keras effectively, our trending book Deep Learning with Keras will be of great help to you! [/box] 9. PyTorch One of the more recent additions to Python deep learning family is PyTorch, a neural network modeling library with strong GPU support. Although still in a beta stage, this project is backed by bigwigs such as Facebook and Twitter. PyTorch builds on the architecture of Torch, another popular deep library, to enable more efficient tensor computation and implementation of dynamic neural networks. [box type="info" align="" class="" width=""]Editor's Tip: Here is Deep Learning with PyTorch to get you started with this amazing tool. [/box] Natural Language Processing Natural Language Processing pertains to designing of systems that process, interpret and analyze human language, spoken or written. Python offers unique libraries for performing a variety of tasks such as working with structured and unstructured text, predictive analytics and much more. 10. NLTK NLTK is a popular Python library for language processing. It offers easy to use interfaces for a variety of NLP tasks such as text classification, tokenization, text parsing, semantic reasoning and much more. It is an open source, community-driven project, and has support for both Python 2 and Python 3. 11. SpaCy SpaCy is another library for advanced natural language processing, based on Python and Cython. It has an extensive support for various deep learning libraries and frameworks such as Tensorflow and PyTorch. With SpaCy, you can build complex statistical models for NLP with relative ease. SpaCy is easy to install and use, and proves to be of great help when it comes to large-scale extracting and analyzing of textual information. [box type="info" align="" class="" width=""]Editor's Tip: To know more about how these libraries are used for natural language processing, make sure you check out the book Natural Language Processing with Python Cookbook [/box] Data Visualization Data visualization is a popularly used Data Science technique for visually analysing and communicating information and valuable business insights through graphs, charts, dashboards and reports. Python offers a lot of popular libraries for effective data storytelling. Some of them are listed below: 12. matplotlib matplotlib is the most popular Python library for data visualization which allows for enterprise-grade 2D and 3D plotting. With matplotlib, you can build different kinds of visualizations such as histograms, bar charts, scatter plots and much more, with just a few lines of code. The popularity of matplotlib rivals that of R’s highly acclaimed ggplot2, and deciding which library is better has been a hot topic for debate, for many years now. Matplotlib runs seamlessly on all Python consoles, including iPython and Jupyter notebooks, giving you all the necessary tools to create and share your data visualizations with others. [box type="info" align="" class="" width=""]Editor's Tip: Get started with matplotlib today, with the help of Matplotlib 2.x By Example [/box] 13. Seaborn Seaborn is a Python-based data visualization library, which finds its roots in matplotlib. Apart from offering attractive and insightful data visualizations, seaborn also offers strong support for other Python libraries such as NumPy and pandas. Per the official seaborn page: “If matplotlib “tries to make easy things easy and hard things possible”, seaborn tries to make a well-defined set of hard things easy too.” 14. Bokeh Bokeh is an interactive data visualization library based on Python. It aims to provide D3.js style elegant graphics and visualizations and runs primarily on modern web browsers. Apart from the ability to create a wide variety of visualizations, Bokeh also supports large-scale interactivity and visualizations of real-time datasets. 15. Plotly Plotly is a popularly used Python library which is used across the world for making publication-quality plots and graphs. With Plotly, you can build interactive dashboards, scatter plots, histograms, candlestick charts, heat maps, and a whole host of other data visualizations with ease. With superior interactivity, deployment and publication capabilities, Plotly is used across different domains, majorly finance and geospatial industries for effective data storytelling. So there you have it! Python has an extensive suite of libraries for every data science related task, each equipped with unique features to make the task fast and hassle-free. While there are a lot more Python libraries out there, we cherry-picked these 15 libraries based on their popularity, usefulness and the value they bring to the table. Also, the extensive community support for Python means you can get help for any kind of problem you might come across while using these tools. It's time now for you to go out there and crunch some data with some of these Python powered libraries!

According to an International Data Corporation (IDC) report, Artificial intelligence (AI) has the potential to impact many areas of customer relationship management (CRM). AI as an armor will ease out mundane tasks for the CRM teams, which implies they will be able to address more customer queries through an automated approach. An AI-based expert CRM offers highly predictive and intuitive ways to customer problems, thus grabbing maximum customer attention. With AI, CRM platforms within different departments such as sales, finance, marketing etc. do not limit themselves to getting service feedback from their customers. But they can also gain information based on the data that customers generate online i.e the social media or IoT devices. With such massive amount of data hailing from various channels, it becomes a bit tricky for organizations to keep a track of its customers. Not only this, but to extract detailed insights from huge amount of data becomes all the more difficult. And here is the gap where, organizations feel the need to bring in an AI-based optimized approach for their CRM platform. The AI-enabled platform can assist CRM teams to gain insights from the large aggregation of customer data, while also paving a way for seamless customer interactions. Organizations can not only provide customers with helpful suggestions, but also recommend products to boost their business profitability. AI-infused CRM platforms can take over straightforward tasks such as client feedback, that otherwise is time consuming. It allows businesses to focus on customers that provide higher business value, which might have got neglected previously. It also acts as a guide for executive level employees via a virtual assistant, allowing them to tackle customer queries without any assistance from senior executives. AI techniques such as Natural language processing(NLP) and predictive analytics are used within the CRM domain, to gain intelligent insights in order to enhance human decision making. NLP interprets incoming emails, categorizes them on the basis of intent, and automatically drafts responses by identifying the priority level. Predictive Analytics helps in detecting the optimal time for solving customer queries, and the mode of communication that will best fit to engage with the customer. With such functionalities, a smarter move towards digitizing organizational solutions can be achieved  reaping huge profits for organizations who wish to leverage it. How AI is transforming CRM Businesses aim to satisfy customers who utilize their services. This is because, keeping a customer happy can lead to further incrementation in revenue generation. Organizations can achieve this rapidly with the help of AI. Salesforce, the market leader in the CRM space, integrated an AI assistant which is popularly known as Einstein. Einstein makes CRM an easy-to-use platform by simply allowing customers to import their data on Salesforce and automatically provides ready-to-crunch data driven insights across different channels. Other organizations such as SAP and Oracle are implementing AI-based technologies for their CRM platforms to provide an improvised customer experience. Let’s explore how AI benefits within an organization: Steering Sales With AI, the sales team can shift their focus from the mundane administrative tasks and get to know their customers better. Sales CRM team leverages novel scoring techniques, which help in prioritizing quality leads, thus generating maximum revenue for the organization. Sales leaders, with help of AI can work towards improving sales productivity. After analyzing company’s historical data and employee activities, the AI-fused CRM software can present a performance report of the top sales representatives. Such a feature helps sales leaders to strategize what the bottom line representatives should learn from the top representatives to drive conversations with their customers that show a likelihood for sales generation. People.ai, a sales management platform, utilize AI to deliver performance analytics, personalized coaching, and provide reviews for their sales pipeline. This can assist sales leaders get a complete view of sales activities going on within their organizations. Marketing it better To trigger a customer sale requires extensive push marketing strategies.With Artificial Intelligence enabled marketing, customers are driven into a predictive journey, which ensures each journey to end up into a sale or a subscription. Both ways it is a win-win situation for the organizations. Predictive scoring can intelligently determine the likelihood of a customer to subscribe to a newsletter or trigger a purchase. AI can also analyze images across various social media sources such as Pinterest, Facebook, and can provide suggestions for visuals of an upcoming advertising campaign. Also, by carrying out sentiment analysis on product reviews and customer feedback, the marketing team can take into account, user’s sentiment about a particular brand or product. This helps brands to announce discount offers in case of a decreased sale, or increase the production of a product in demand. Marketo, a marketing automation platform includes a software which aids different CRM platforms to gain rich behavioral insights of their customers and to drive business strategies. 24*7 customer support Whenever a customer query arises within a CRM, AI anticipates the likely issues and resolves them before it results into a problem. Different customer cases are classified and directed to the right service agent to address with the help of predictive analytics techniques. Also, NLP-based digital assistants known as chatbots are used to analyze the written content within e-mails. A chatbot efficiently responds to customer e-mails; in most rare cases, it directs the e-mail to a service agent. Chatbots can even notify a customer about an early-bird offer to purchase a product, which they are likely to buy. It can also issue meetings and notify the same by scheduling reminders­‑given the era of push notifications and smart wearables. Hence, with AI into CRM, organizations can not only offer customers better services but also provide 24*7 support. Agent.ai, an AI-based customer service platform, allows organizations to provide a 24*7*365 customer support including holidays, weekends, and non-staffed hours. Application development no more a developer’s play Building an application has become an important milestone to achieve for any organization. If the application has a seamless and user-friendly interface, it is favoured by many customers and thus, the organization gets more customer traction. Building an application was considered as ‘a developers job only’ as it involves coding. However, due to the rise in platforms that help build an application with lesser coding or in fact no-coding, any non-coder can easily develop an application. CRM platforms helps businesses to build applications, which provides insight driven predictions and recommendation solutions to their customers. Salesforce assures their customers that each application built on their platform includes intelligent data modeling, tracking, and monitoring. Business users, data scientists, or any non-developer, can now build applications without learning to code. This helps them to create prediction-based applications their way; without the IT hassle. Challenges & Limitations AI implementations are becoming common with an increased number of organizations adopting it both on a small and a large scale. Many businesses are moving towards a smart customer management by infusing AI within their organizations. AI undoubtedly brings in an ease of work, but there are challenges that the CRM platform can face, which if unaddressed may cause revenue declination for businesses. Below are the challenges which organizations might face while setting up AI in their CRM platform: Outdated data: Organizations collect a huge amount of data during various business activities to drive meaningful insights about sales, customer preferences, etc. This data is a treasure trove for the marketing team, to plan strategies in order to attract more new customers and retain the existing ones. On the contrary, if the data provided is not updated,  CRM teams may find it difficult to understand the current customer relationship status. To avoid this, a comprehensive data cleanup project is essential to maintain better quality of data. Partially automated: AI creates an optimized environment  for the CRM with the use of  predictive analytics and natural language processing for better customer engagement. This eases out the mundane elements for the CRM team, and they can focus on other strategic outcomes. This does not imply that AI is completely replacing humans. Instead, a human touch is required to monitor if the solutions given by the AI benefits the customer and how they can tweak it to a much more smarter AI. Intricacies of language: An AI is trained on data which includes various set of phrases and questions, and also the desired output that it should give. If the query input by the customer is not phrased in a correct manner, the AI is unable to provide them with correct solutions. Hence, customers have to take precautions while asking their queries and phrase it in the correct manner, else the machine would not understand what the customer aims to ask. Infusing AI into CRM has multiple benefits, but the three most important ones include predictive scoring, forecasting, and recommendations. These benefits empower CRM to outsmart its traditional counterpart by helping organizations to serve its customers with state-of-the-art results. Customers appreciate when their query is addressed in lesser time,leaving a positive remark on the organization. Additionally we have digital assistants to assist firms in solving customer query quickly.

Many organizations rely on machine learning techniques in their day-today workflow, to cut down on the time required to do a job. The reason why these techniques are robust is because they undergo various tests in order to carry out correct predictions about any data fed into them. During this phase, there are also certain errors generated, which can lead to an inconsistent ML model. Two common errors that we are going to look at in this article are that of bias and Variance, and how a trade-off can be achieved between the two in order to generate a successful ML model.  Let’s first have a look at what creates these kind of errors. Machine learning techniques or more precisely supervised learning techniques involve training, often the most important stage in the ML workflow. The machine learning model is trained using the training data. How is this training data prepared? This is done by using a dataset for which the output of the algorithm is known. During the training stage, the algorithm analyzes the training data that is fed and produces patterns which are captured within an inferred function. This inferred function, which is derived after analysis of the training dataset, is the model that would be further used to map new examples. An ideal model generated from this training data should be able to generalize well. This means, it should learn from the training data and should correctly predict or classify data within any new problem instance. In general, the more complex the model is, the better it classifies the training data. However, if the model is too complex i.e it will pick up random features i.e. noise in the training data, this is the case of overfitting i.e. the model is said to overfit . On the other hand, if the model is not so complex, or missing out on important dynamics present within the data, then it is a case of underfitting. Both overfitting and underfitting are basically errors in the ML models or algorithms. Also, it is generally impossible to minimize both these errors at the same time and this leads to a condition called as the Bias-Variance Tradeoff. Before getting into knowing how to achieve the trade-off, lets simply understand how bias and variance errors occur. The Bias and Variance Error Let’s understand each error with the help of an example. Suppose you have 3 training datasets say T1, T2, and T3, and you pass these datasets through a supervised learning algorithm. The algorithm generates three different models say M1, M2, and M3 from each of the training dataset. Now let’s say you have a new input A. The whole idea is to apply each model on this new input A. Here, there can be two types of errors that can occur. If the output generated by each model on the input A is different(B1, B2, B3), the algorithm is said to have a high Variance Error. On the other hand, if the output from all the three models is same (B) but incorrect, the algorithm is said to have a high Bias Error. High Variance also means that the algorithm produces a model that is too specific to the training data, which is a typical case of Overfitting. On the other hand, high bias means that the algorithm has not picked up defining patterns from the dataset, this is a case of Underfitting. Some examples of high-bias ML algorithms are: Linear Regression, Linear Discriminant Analysis and Logistic Regression Examples of high-variance Ml algorithms are: Decision Trees, k-Nearest Neighbors and Support Vector Machines.  How to achieve a Bias-Variance Trade-off? For any supervised algorithm, having a high bias error usually means it has low variance error and vise versa. To be more specific, parametric or linear ML algorithms often have a high bias but low variance. On the other hand, non-parametric or non-linear algorithms have vice versa. The goal of any ML model is to obtain a low variance and a low bias state, which is often a task due to the parametrization of machine learning algorithms. So how can we achieve a trade-off between the two? Following are some ways to achieve the Bias-Variance Tradeoff: By minimizing the total error: The optimum location for any model is the level of complexity at which the increase in bias is equivalent to the reduction in variance. Practically, there is no analytical method to find the optimal level. One should use an accurate measure for error prediction and explore different levels of model complexity, and then choose the complexity level that reduces the overall error. Generally resampling based measures such as cross-validation should be preferred over theoretical measures such as Aikake's Information Criteria. Source: http://scott.fortmann-roe.com/docs/BiasVariance.html (The irreducible error is the noise that cannot be reduced by algorithms but can be reduced with better data cleaning.) Using Bagging and Resampling techniques: These can be used to reduce the variance in model predictions. In bagging (Bootstrap Aggregating), several replicas of the original dataset are created using random selection with replacement. One modeling algorithm that makes use of bagging is Random Forests. In Random Forest algorithm, the bias of the full model is equivalent to the bias of a single decision tree--which itself has high variance. By creating many of these trees, in effect a "forest", and then averaging them the variance of the final model can be greatly reduced over that of a single tree. Adjusting minor values in algorithms: Both the k-nearest algorithms and Support Vector Machines(SVM) algorithms have low bias and high variance. But the trade-offs in both these cases can be changed. In the K-nearest algorithm, the value of k can be increased, which would simultaneously increase the number of neighbors that contribute to the prediction. This in turn would increase the bias of the model. Whereas, in the SVM algorithm, the trade-off can be changed by an increase in the C parameter that would influence the violations of the margin allowed in the training data. This will increase the bias but decrease the variance. Using a proper Machine learning workflow: This means you have to ensure proper training by: Maintaining separate training and test sets - Splitting the dataset into training (50%), testing(25%), and validation sets ( 25%). The training set is to build the model, test set is to check the accuracy of the model, and the validation set is to evaluate the performance of your model hyperparameters. Optimizing your model by using systematic cross-validation - A cross-validation technique is a must to fine tune the model parameters, especially for unknown instances. In supervised machine learning, validation or cross-validation is used to find out the predictive accuracy within various models of varying complexity, in order to find the best model.For instance, one can use the k-fold cross validation method. Here, the dataset is divided into k folds. For each fold, train the algorithm on k-1 folds iteratively, using the remaining fold(also called as 'holdout fold')as the test set. Repeat this process until each k has acted as a test set. The average of the k recorded errors is called as the cross validation error and can serve as the performance metric for the model.   Trying out appropriate algorithms - Before relying on any model we need to first ensure that the model works best for our assumptions. One can make use of the No Free Lunch theorem, which states that one model can not work for only one problem. For instance, while using No Free lunch theorem, a random search will do the same as any of the heuristic optimization algorithms.   Tuning the hyperparameters that can give an impactful performance - Any machine learning model requires different hyperparameters such as constraints, weights or learning rates for generalizing different data patterns. Tuning these hyperparameters is necessary so that the model can optimally solve machine learning problems. Grid search and randomized search are two such methods practiced for hyperparameter tuning. So, we have listed some of the ways where you can achieve trade-off between the two. Both bias and variance are related to each other, if you increase one the other decreases and vice versa. By a trade-off, there is an optimal balance in the bias and variance which gives us a model that is neither underfit nor overfit. And finally, the ultimate goal of any supervised machine algorithm lies in isolating the signal from the dataset, and making sure that it eliminates the noise.  

[box type="note" align="" class="" width=""]We have come to you with another guest post by Indra den Bakker, an experienced deep learning engineer and a mentor on Udacity for many budding data scientists. Indra has also written one of our best selling titles, Python Deep Learning Cookbook which covers solutions to various problems in modeling deep neural networks.[/box] In 2014, we took a significant step in AI with the introduction of Generative Adversarial Networks -better known as GANs- by Ian Goodfellow, amongst others. The real breakthrough of GANs didn’t follow until 2016, however, the original paper includes many novel ideas that would be exploited in the years to come. Previously, deep learning had already revolutionized many industries by achieving above human performance. However, many critics argued that these deep learning models couldn’t compete with human creativity. With the introduction to GANs, Ian showed that these critics could be wrong. Figure 1: example of style transfer with deep learning The idea behind GANs is to create new examples based on a training set - for example to demonstrate the ability to create new paintings or new handwritten digits. In GANs two competing deep learning models are trained simultaneously. These networks compete against each other: one model tries to generate new realistic examples, this network is also called the generator. The other network tries to classify if an example originates from the training set or from the generator, also called as discriminator. In other words, the generator tries to mislead the discriminator by generating new examples. In the figure below we can see the general structure of GANs. Figure 2: GAN structure with X as training examples and Z as noise input. GANs are fundamentally different from other machine learning applications. The task of a GAN is unsupervised: we try to extract patterns and structure from data without additional information. Therefore, we don’t have a truth label. GANs shouldn’t be confused with autoencoder networks. With autoencoders we know what the output should be: the same as the input. But in case of GANs we try to create new examples that look like the training examples but are different. It’s a new way of teaching an agent to learn complex tasks by imitating an “expert”. If the generator is able to fool the discriminator one could argue that the agent mastered the task - think about the Turing test. Best way to explain GANs is to use images as an example. The resulting output of GANs can be fascinating. The most used dataset for GANs is the popular MNIST dataset. This dataset has been used in many deep learning papers, including the original Generative Adversarial Nets paper. Figure 3: example of MNIST training images Let’s say as input we have a bunch of handwritten digits. We want our model to be able to take these examples and create new handwritten digits. We want our model to learn how to write digits in such a way that it looks like handwritten digits. Note, that we don’t care which digits the model creates as long as it looks like one of the digits from 0 to 9. As you may suspect, there is a thin line between generating examples that are exact copies of the training set and newly created images. We need to make sure that the generator generates new images that follow the distribution of the training examples but are slightly different. This is where the creativity needs to come in. In Figure 2, we’ve showed that the generator uses noise -random values- as input. This noise is random, to make sure that the generator creates different output each time. Now that we know what we need and what we want to achieve, let’s have a closer look at both model architectures. Let’s start with the generator. We will feed the generator with random noise: a vector of 100 values randomly drawn between -1 and 1. Next, we stack multiple fully connected layers with Leaky ReLU activation function. Our training images are in grayscale and are sized as 28x28. Which means, flattened we need an output of 784 units for the final layer of our generator - the output of the generator should match the size of the training images. As activation function for our final layer we will be using TanH to make sure the resulting values are squeezed between -1 and 1. The final model architecture of our generator looks as follows: Figure 4: model architecture of the generator Next, we define our discriminator model. Most common is to use a mirrored version of the generator, where we have as input 784 values and as final layer a fully connected layer with 1 hidden neuron and sigmoid activation function for binary classification. Keep in mind that both the generator and discriminator are trained at the same time. The model looks like this: Figure 5: model architecture of the discriminator In general, generating new images is a harder task. Therefore, sometimes it can be beneficial to train the generator twice for each step. Whereas the discriminator will only be trained once. Another option is to set the learning rate for the discriminator a bit smaller than the learning rate for the generator. Tracking the performance of GANs can be tricky. Sometimes a lower loss doesn’t represent a better output. That’s why it’s a good idea to output the generated images during the training process. In the following figure we can see the digits generated by a GAN after 20 epochs. Figure 6: example output of generated MNIST images As we have stated in the introduction, GANs didn’t get much traction until 2016. GANs were mostly unstable and hard to train. Small adjustments in the model or training parameter resulted in unsatisfying results. Advancements in model architecture and other improvements fixed some of the previous limitations and unlocked the real potential of GANs. An important improvement was introduced by Deep Convolutional GANs (DCGANs). DCGANs is a network architecture, where in both the discriminator and generator are fully convolutional. The output is more stable - for datasets with higher translation invariance, like the Fashion MNIST dataset. Figure 7: example of Fashion MNIST images generated by a Deep Convolutional Generative Adversarial Network (DCGAN) There is so much more to discover with GANs and there is huge potential still to be unlocked. According to Yann LeCun - one of the fathers of deep learning - GANs are the most important advancement in machine learning in the last 20 years. GANs can be used for many different applications, ranging from 3D face generation to upscaling resolution of images and text-to-image. GANs might be the stepping stone we have been waiting for to add creativity to machines. [author title="Author's Bio"]Indra den Bakker is an experienced deep learning engineer and mentor on Udacity. He is the founder of 23insights, a part of NVIDIA's Inception program—a machine learning start-up building solutions that transform the world’s most important industries. For Udacity, he mentors students pursuing a Nanodegree in deep learning and related fields, and he is also responsible for reviewing student projects. Indra has a background in computational intelligence and has worked for several years as a data scientist for IPG Mediabrands and Screen6 before founding 23insights. [/author]      

At the wee hours of the night, on January 4 this year, over 9.8 million people experienced a magnitude 4.4 earthquake that rumbled across the San Francisco Bay Area. This was followed by a magnitude 7.6 earthquake in the Caribbean sea on January 9, following which a tsunami advisory was in effect for Puerto Rico and the U.S. and British Virgin Islands. In the past 6 months, the United States alone has witnessed four back-to-back storms from one brutal hurricane season, and a massive wildfire with almost 2 million acres of land ablaze. Natural Disasters across the globe are increasingly becoming more damaging and frequent. Since 1970, the number of disasters worldwide has more than quadrupled to around 400 a year. These series of natural disasters have strained the emergency services and disaster relief operations beyond capacity. We now need to look for newer ways to assist the affected people and automate the recovery process. Artificial intelligence and machine learning have advanced to the state where they are highly proficient in making predictions, and in identification and classification tasks. These use cases of AI can also be applied to prevent disasters or respond quickly, in case of an emergency.   Here are 5 ways how AI can lend a helping hand during emergency situations. 1. Machine Learning for targeted disaster relief management In case of any disaster, the first step is to formulate a critical response team to help those in distress. Before the team goes into action, it is important to analyze and assess the extent of damage and to ensure that the right aid goes first to those who need it the most. AI techniques such as image recognition and classification can be quite helpful in assessing the damage as they can analyze and observe images from the satellites. They can immediately and efficiently filter these images, which would have required months to be sorted manually. AI can identify objects and features such as damaged buildings, flooding, blocked roads from these images. They can also identify temporary settlements which may indicate that people are homeless, and so the first care could be directed towards them.   Artificial intelligence and machine learning tools can also aggregate and crunch data from multiple resources such as crowd-sourced mapping materials or Google maps. Machine learning approaches then combine all this data together, remove unreliable data, and identify informative sources to generate heat maps. These heat maps can identify areas in need of urgent assistance and direct relief efforts to those areas. Heat maps are also helpful for government and other humanitarian agencies in deciding where to conduct aerial assessments. DigitalGlobe provides space imagery and geospatial content. Their Open Data Program is a special program for disaster response. The software learns how to recognize buildings on satellite photos by learning from the crowd. DigitalGlobe releases pre- and post-event imagery for select natural disasters each year, and their crowdsourcing platform, Tomnod, will prioritize micro-tasking to accelerate damage assessments. Following the Nepal Earthquakes in 2015, Rescue Global and academicians from the Orchid Project used machine learning to carry out rescue activities. They took pre and post-disaster imagery and utilized crowd-sourced data analysis and machine learning to identify locations affected by the quakes that had not yet been assessed or received aid. This information was then shared with relief workforces to facilitate their activities. 2. Next Generation 911 911 is the first source of contact during any emergency situation. 911 dispatch centers are already overloaded with calls on a regular day. In case of a disaster or calamity, the number gets quadrupled, or even more. This calls for augmenting traditional 911 emergency centers with newer technologies for better management. Traditional 911 centers rely on voice-based calls alone. Next-gen dispatch services are upgrading their emergency dispatch technology with machine learning to receive more types of data. So now they can ingest the data from not just calls but also from text, video, audio, and pictures, to analyze them to make quick assessments. The insights gained from all this information can be passed on to the emergency response teams out in the field to efficiently carry out critical tasks. The Association of Public-Safety Communications (APCO) have employed IBM’s Watson to listen to 911 calls. This initiative is to help emergency call centers improve operations and public safety by using Watson’s speech-to-text and analytics programs. Using Watson's speech-to-text function, the context of each call is fed into the AI's analytics program allowing improvements in how call centers respond to emergencies. It also helps in reducing call times, provide accurate information, and help accelerate time-sensitive emergency services. 3. Sentiment analysis on social media data for disaster management and recovery Social media channels are a major source of news in present times. Some of the most actionable information, during a disaster, comes from social media users. Real-time images and comments from Facebook, Twitter, Instagram, and YouTube can be analyzed and validated by AI to filter real information from fake ones. These vital stats can help on-the-ground aid workers to reach the point of crisis sooner and direct their efforts to the needy. This data can also help rescue workers in reducing the time needed to find victims. In addition, AI and predictive analytics software can analyze digital content from Twitter, Facebook, and Youtube to provide early warnings, ground-level location data, and real-time report verification. In fact, AI could also be used to view the unstructured data and background of pictures and videos posted to social channels and compare them to find missing people. AI-powered chatbots can help residents affected by a calamity. The chatbot can interact with the victim, or other citizens in the vicinity via popular social media channels and ask them to upload information such as location, a photo, and some description. The AI can then validate and check this information from other sources and pass on the relevant details to the disaster relief committee. This type of information can assist them with assessing damage in real time and help prioritize response efforts. AI for Digital Response (AIDR) is a free and open platform which uses machine intelligence to automatically filter and classify social media messages related to emergencies, disasters, and humanitarian crises. For this, it uses a Collector and a Tagger. The Collector helps in collecting and filtering tweets using keywords and hashtags such as "cyclone" and "#Irma," for example. The Collector works as a word-filter. The Tagger is a topic-filter which classifies tweets by topics of interest, such as "Infrastructure Damage," and "Donations," for example. The Tagger automatically applies the classifier to incoming tweets collected in real-time using the Collector. 4. AI answers distress and help-calls Emergency relief services are flooded with distress and help calls in the event of any emergency situation. Managing such a huge amount of calls is time-consuming and expensive when done manually. The chances of a critical information being lost or unobserved is also a possibility. In such cases, AI can work as a 24/7 dispatcher. AI systems and voice assistants can analyze massive amounts of calls, determine what type of incident occurred and verify the location. They can not only interact with callers naturally and process those calls, but can also instantly transcribe and translate languages. AI systems can analyze the tone of voice for urgency, filtering redundant or less urgent calls and prioritizing them based on the emergency. Blueworx is a powerful IVR platform which uses AI to replace call center officials. Using AI technology is especially useful when unexpected events such as natural disasters drive up call volume. Their AI engine is well suited to respond to emergency calls as unlike a call center agent, it can know who a customer is even before they call. It also provides intelligent call routing, proactive outbound notifications, unified messaging, and Interactive Voice Response. 5. Predictive analytics for proactive disaster management Machine learning and other data science approaches are not limited to assisting the on-ground relief teams or assisting only after the actual emergency. Machine learning approaches such as predictive analytics can also analyze past events to identify and extract patterns and populations vulnerable to natural calamities. A large number of supervised and unsupervised learning approaches are used to identify at-risk areas and improve predictions of future events. For instance, clustering algorithms can classify disaster data on the basis of severity. They can identify and segregate climatic patterns which may cause local storms with the cloud conditions which may lead to a widespread cyclone. Predictive machine learning models can also help officials distribute supplies to where people are going, rather than where they were by analyzing real-time behavior and movement of people. In addition, predictive analytics techniques can also provide insight for understanding the economic and human impact of natural calamities. Artificial neural networks take in information such as region, country, and natural disaster type to predict the potential monetary impact of natural disasters. Recent advances in cloud technologies and numerous open source tools have enabled predictive analytics with almost no initial infrastructure investment. So agencies with limited resources can also build systems based on data science and develop more sophisticated models to analyze disasters.   Optima Predict, a suite of software by Intermedix collects and reads information about disasters such as viral outbreaks or criminal activity in real time. The software spots geographical clusters of reported incidents before humans notice the trend and then alerts key officials about it. The data can also be synced with FirstWatch, which is an online dashboard for an EMS (Emergency Medical Services) personnel. Thanks to the multiple benefits of AI, government agencies and NGOs can start utilizing machine learning to deal with disasters. As AI and allied fields like robotics further develop and expand, we may see a fleet of drone services, equipped with sophisticated machine learning. These advanced drones could expedite access to real-time information at disaster sites using video capturing capabilities and also deliver lightweight physical goods to hard to reach areas. As with every progressing technology, AI will also build on its existing capabilities. It has the potential to eliminate outages before they are detected and give disaster response leaders an informed, clearer picture of the disaster area, ultimately saving lives.

Machine learning has been the talk of the town! With implementations in large number of organizations to carry out prediction and classification tasks. Machine learning is cutting edge in  identifying data and processing it to generate meaningful insights based on predictive analytics. But to leverage machine learning, huge computational resources are required. While many may think of it as rocket science, Google has simplified machine learning access to everyone through Deeplearn.js - an initiative that allows ML to run entirely on a web browser. Deeplearn.js is an open source WebGL- accelerated JS library. This Google PAIR’s initiative (to study and redesign human interactions with ML) aims to make ML available for everyone. This implies that it will not be restricted to specific groups of people such as developers or any businesses implementing it. Deeplearn.js + browser: A perfect match? We can say browsers such as Chrome, Internet explorer, Safari, etc are an integral part of our life as it connects us with the world. Their accessibility feature is visible in PWAs’(Progressive Web Apps) wherein applications can run on browsers without the need to download them. In a similar way, machine learning can be carried out within browsers without the fuss of downloading or installing any computational resources. Wonder how? With Deeplearn.js! Deeplearn.js specifically written in Javascript, is exclusively tailored for machine learning to function on web browsers. It offers an interactive client-side platform which helps them carry out rapid prototyping and visualizations. Machine learning involves rapid computations with huge CPU requirements and is a complete  mismatch for Javascript because of its speed limit. Deeplearn.js is a work-around that allows ML to be implemented using Javascript via the WebGL Javascript API. Additionally, you can use hardware accelerators such as GPUs via the webGL to perform faster and excellent computations with 2D and 3D graphics. Basic Mechanism - The structure of Deeplearn.js is a blend of Tensorflow and NumPy, which are Python-based packages for scientific computing. The NumPy acts as a quick execution model and the TensorFlow API provides a delayed execution model. Though TensorFlow is a fast and scalable framework widely used by researchers and developers. However, creating web applications on the browser with TensorFlow is difficult as it lacks runtime support to create web applications. Deeplearn.js allows TensorFlow model capabilities to be imported on the browser. By using the tools within Deeplearn.js, weights from the TensorFlow model can be exported. Opportunities for business - Traditional businesses shy away from using latest ML tools as computational resources are expensive and complicated. Also, due to the complexities in ML, there is a need to hire a technical expert. Through Deeplearn.js, firms can now easily access advanced ML tools and resources. It can not only help them solve data centric business problems but also additionally provide them with innovative strategies, increased competition and improved advantages to stay ahead of their competitors. Differentiating factor - Deeplearn.js is not the only inbrowser ML library. There are other competing frameworks such as ConvNetJS and Tensorfire, a much recent and almost identical framework to deeplearn.js. A unique feature that differentiates deeplearn.js is its capability to perform faster inference, along with full back propagation. Implementations with Deeplearn.js Performance RNN aids in generating music with expressive timing and dynamics. It has been successfully ported into the browser using the Deeplearn.js environment after being trained in TensorFlow. The training data used was the Yamaha e-Piano Competition dataset, which includes MIDI captures of ~1400 performances by skilled pianists. Teachable Machine is built using Deeplearn.js library. It allows users to teach a machine via a camera with live teaching and without any requirement to code.    Faster Neural Style Transfer algorithm allows in-browser image style transfer. It transfers the style of an image into the content of another image. To explore other practical projects on Deeplearn.js, you may visit the GitHub repository here. Deeplearn.js, with the fusion of Machine learning has opened new opportunities and focus areas for businesses and non-developers. SME’s (Subject Matter Expertise) within a business can now grasp deeper insights on how to achieve desired results with Machine learning. The browser is home for many developments which are yet to be revealed in the future. Deeplearn.js truly is a milestone in bringing the web and ML a step closer. However being at the early stage, it would be exciting to see how it unfolds ML for anyone on the planet.