Prof. Rowel Atienza discusses the intuition behind deep learning, advances in GANs & techniques to create cutting-edge AI models

In recent years, deep learning has made unprecedented progress in vision, speech, natural language processing and understanding, and other areas of data science. Developments in deep learning techniques, including GANs, variational autoencoders and deep reinforcement learning, are creating impressive AI results. For example, DeepMind's AlphaGo Zero became a game changer in AI research when it beat world champions in the game of Go.

In this interview, Professor Rowel Atienza, author of the book Advanced Deep Learning with Keras talks about the recent developments in the field of deep learning. This book is a comprehensive guide to the advanced deep learning techniques available today, so you can create your own cutting-edge AI. This book strikes a balance between advanced concepts in deep learning and practical implementations with Keras.

Key takeaways from the interview

1. The intuition of deep learning is built on the fact that the deeper the network gets, the more feature representations the network learns in order to solve complex real-world problems.
2. The objective of deep learning is to enable agents to be more robust to unforeseen events and to lessen the dependency on huge data.
3. Advances in GANs enable us to generate high-dimensional fake data such as high-resolution images or videos that look very convincing.
4. Deep learning tackles the curse of dimensionality by finding efficient data structures and layers that could represent complex data in the most efficient manner.

The interview in detail

What is the intuition behind deep learning? What are the recent developments in deep learning?

Rowel Atienza: Deep learning is built on the intuition that the deeper the network gets, the more feature representations the network learns in order to solve complex real-world problems. Unlike machine learning, deep learning learns these features automatically from data in different degrees of supervision.

There are many recent developments in deep learning. There are advances on graph neural networks because people are realizing the limits of NLP (Natural Language Processing), CNN (Convolution Neural Networks), and RNN (Recurrent Neural Networks) in representing more complex data structures such as social network, 3D shapes, molecular structures, etc. Implementing the causality in reasoning on data is another area of strong interest. Deep learning is strong on correlation not on discovering causality in data. Meta learning or learning to learn is also another area of interest. The objective is to enable agents to be more robust to unforeseen events and to lessen the dependency on huge data.

What are different deep learning techniques to create successful AI?

RA: A successful AI is dependent on two things: 1) deep domain knowledge and 2) deep understanding of state of the art techniques that will work on the domain problem. Domain knowledge comes from someone who is very familiar with the domain problem. This person is not necessarily knowledgeable in AI. This domain knowledge is then modelled in AI to automate the process of problem solving.

How deep learning tackles the curse of dimensionality?

RA: One of the goals of deep learning is to keep on finding efficient data structures and layers that could represent complex data in the most efficient manner. For example, geometric deep learning is able to circumvent the limitations of representing and learning from 3D data by avoiding inefficient 3D convolutions. There is still so much to be done in this space.

What is autoencoders? What is the need of autoencoders in deep learning? How do you create an autoencoder?

RA: Autoencoders compress high dimensionality data into low dimensionality code without losing important information. Low-dimensional code is suitable for further processing by other deep learning models such as in generative models like GANs and VAEs. Autoencoder can easily be implemented using two networks, an encoder and decoder. The depth, width, and type of layers are dependent on the original data to be encoded.

Why are GANs so innovative?

RA: GANs are innovative since they are good in generating fake data that look real. It is something that is hard to accomplish using other generative models. The advances in GANs enable us to generate high-dimensional fake data such as high resolution image or video that look very convincing.

Tell us a little bit about this book? What makes this book necessary? What gap does it fill?

RA: Advanced Deep Learning with Keras focuses on recent advances on deep learning It starts with a quick review of deep learning concepts (NLP, CNN, RNN). The discussions on deep neural networks, autoencoders, generative adversarial network (GAN), variational autoencoders (VAE), and deep reinforcement learning (DRL) follow.

The book is important for everyone who would like to understand advanced concepts on deep learning and their corresponding implementation in Keras. The current version has in depth focus on generative models (autoencoders, GANs, VAEs) that could be used in-practical setting. The DRL explains the core concepts of value-based and policy-based methods in reinforcement learning and the corresponding working implementations in Keras which are difficult to make them right.

About the Book

Advanced Deep Learning with Keras is a comprehensive guide to the advanced deep learning techniques available today, so you can create your own cutting-edge AI. Using Keras as an open-source deep learning library, you'll find hands-on projects throughout that show you how to create more effective AI with the latest techniques.

About the Author

Rowel Atienza is an Associate Professor at the Electrical and Electronics Engineering Institute of the University of the Philippines, Diliman. He holds the Dado and Maria Banatao Institute Professorial Chair in Artificial Intelligence. Rowel has been fascinated with intelligent robots since he graduated from the University of the Philippines. He received his MEng from the National University of Singapore for his work on an AI-enhanced four-legged robot. He finished his Ph.D. at The Australian National University for his contribution to the field of active gaze tracking for human-robot interaction.

Deep learning models have massive carbon footprints, can photonic chips help reduce power consumption?

Machine learning experts on how we can use machine learning to mitigate and adapt to the changing climate

Google launches beta version of Deep Learning Containers for developing, testing and deploying ML applications