Скачать или смотреть Complete Guide to Transformers: RNNs, Attention & BERT Explained

Embark on a comprehensive journey into the world of Natural Language Processing (NLP), culminating in a deep understanding of the revolutionary Transformer architecture. This full course meticulously builds your knowledge from the ground up.

We begin with an introduction to NLP and its significance, then dive into the fundamentals of Recurrent Neural Networks (RNNs). You'll learn how RNNs function, their structure, the concept of hidden states, and how weights and biases operate across time steps, complete with mathematical formulations and an exploration of backpropagation. We'll cover various RNN architectures (Many-to-One, Many-to-Many, One-to-Many, One-to-One) and guide you through building your first RNN model in PyTorch, including text pre-processing, vocabulary building, zero padding, and data preparation.

Next, we address RNN limitations and introduce crucial concepts like Word Embeddings, exploring popular methods such as Skip-gram, GloVe, and FastText, along with their practical implementation. The course then transitions to more advanced RNN structures, focusing on the powerful Encoder-Decoder architecture, discussing teacher forcing, and demonstrating its implementation.
The limitations of traditional sequence-to-sequence models pave the way for the game-changing Attention Mechanism. You'll gain a thorough understanding of how attention works and the specifics of Self-Attention calculations. This knowledge forms the bedrock for understanding the Transformer architecture itself.

Finally, we explore state-of-the-art Transformer models like BERT, delving into its pre-training and fine-tuning processes. We'll also touch upon the T5 model, compare BERT and GPT, and even demonstrate using a pre-trained model for a practical task like generating headlines. This course provides both the theoretical underpinnings and practical insights needed to master modern NLP techniques.

Chapters-
0:00 - Introduction
1:12 = Recurrent Neural Networks
2:19 - How RNN functions
3:14 - Structure of a simple representation of RNN
4:49 - Understanding Hidden State
5:36 - Output Layer
6:17 - Hidden States
7:32 - Weights and bias
8:37 Formulas for hidden state and output state
9:04 - Backward propagation in RNN
10:24 - 4 types of RNN architectures
13:59 - Problem Statement for Building our first RNN model
15:42 - Building our first RNN model
23:56 - Word Embeddings
25:40 - 3 popular methods of creating word embedding
32:59 - Implementation of word embedding
36:46 Limitations of RNN
36:52 - Multi-Layered RNN
38:32 - Bidirectional RNN
41:08 - Implementing Multi-Layered and Bi-directional RNN in Python
42:18 - GRU
47:44 - GRU Implementation
48:28 - LSTM
54:16- LSTM Implementation
55:52 - Problems with a Many-to-one RNN
56:28 - Many-to-many RNN models
1:00:20 - Encoder-decoder architecture
1:02:28 - Teacher forcing
1:05:23 - Implementation of encoder-decoder
1:24:17 - Limitations of the Encoder-Decoder model
1:25:49 - Attention Mechanism and Its Working
1:32:02 - Implementation of Attention Mechanism
1:37:17 - Transformers (In-Depth)
1:58:40 - BERT
2:04:17 - Fine-tuning a pre-training language model
2:07:28 - Finding Pre-Trained Transformer Models
2:10:31 - Using Pre-Trained model to classify sentiments
2:14:24 - Using pre-trained model to generate headlines
2:16:23 - Comparison between BERT and GPT
2:17:43 - Conclusion