Introduction to Attention-Based Neural Networks

Course details

Attention-based models allow neural networks to focus on the most important features of the input, thus producing better results at the output. In this course, Janani Ravi explains how recurrent neural networks work and builds and trains two image captioning models one without attention and another using attention models and compares their results. If you have some experience and understanding of how neural networks work and want to see what attention-based models can do for you, check out this course.

Concepts

Introduction

Prerequisites
What are attention-based models
Attention in language generation and translation models

Recurrent Neural Networks to Learn Sequential Data

Feed forward networks and their limitations
Recurrent neural networks for sequential data
The need for long memory cells
LSTM and GRU cells
Types of RRNNS

Encoder-Decoder Networks for Language Models

Language generation models
Sequence to sequence models for language translation

Attention-Based Neural Networks

The role of attention in sequence to sequence models
Attention mechanism in sequence to sequence models
Alignment weights in attention models
Bahdanau attention
Attention models for image captioning
Encoder decoder structure for image captioning

Image Captioning Model without Attention

Setting up Colab and Google Drive
Loading in the Flickr8k dataset
Constructing the vocabulary
Setting up the dataset class
Implementing utility functions for training data
Building the encoder CNN
Building the decoder RNN
Setting up the sequence to sequence model
Training the image captioning model

Image Captioning Model Using Attention

Loading the dataset and setting up utility functions
The encoder CNN generating unrolled feature maps
Implementing Bahdanau attention
The decoder RNN using attention
Generating captions using attention
Training the attention-based image captioning model
Visualizing the model's attention

Conclusion

Summary and next steps