Performance Optimization for Real-Time Inference

Machine Learning (ML) and Deep Learning (DL) models are used in almost every industry, such as e-commerce, manufacturing, life sciences, and finance. Due to this, there have been meaningful innovations to improve the performance of these models. Since the introduction of transformer-based models in 2018, which were initially developed for Natural Language Processing (NLP) applications, the size of the models and the datasets required to train the models has grown exponentially. Transformer-based models are now used for forecasting as well as computer vision applications, in addition to NLP.

Let’s travel back in time a little to understand the growth in size of these models. Embeddings from Language Models (ELMo), which was introduced in 2018, had 93.6 million parameters, while the Generative Pretrained Transformer model (also known as GPT-3), in 2020, had 175 billion parameters. Today, we have DL models such as Switch Transformers...