Conversational AI with Transformer Models - Building Blocks and Optimization Techniques

Explore the world of conversational AI and transformer models in this 36-minute video from Databricks. Delve into the building blocks of conversational agents and natural language understanding engines, focusing on the advantages of transformer models over traditional RNN/LSTM approaches. Learn about knowledge distillation and model compression techniques for deploying parameter-heavy models in resource-limited production environments. Gain insights into the flow of conversational agents, the benefits of transformer-based models, and various model compression techniques including quantization. Discover practical applications with sample code in PyTorch and TensorFlow 2, and understand key concepts such as BERT, masked language models, and next sentence prediction. By the end of this talk, you'll have a comprehensive understanding of how to build and optimize conversational AI systems using cutting-edge transformer models.

Intro
Why Conversational Al/chatbots?
Chatbot Conversation Framework
Use-case in hand
Chatbot Flow Diagram
Components of NLU Engine
Transformers for Intent Classification
BERT: Bidirectional Encoder Representations from Transformers
Masked Language Model
Next Sentence Prediction
BERT: CLS token for classification
Different models with accuracy and size over time
Use-case data summary
Model Training
Efficient Model Inference
Knowledge Distillation
Quantization
No padding
Productizing BERT for CPU Inference
Ensembling LUIS and DistilBERT
Team behind the project