Week 12

$$\gdef \sam #1 {\mathrm{softargmax}(#1)}$$ $$\gdef \vect #1 {\boldsymbol{#1}} $$ $$\gdef \matr #1 {\boldsymbol{#1}} $$ $$\gdef \E {\mathbb{E}} $$ $$\gdef \V {\mathbb{V}} $$ $$\gdef \R {\mathbb{R}} $$ $$\gdef \N {\mathbb{N}} $$ $$\gdef \relu #1 {\texttt{ReLU}(#1)} $$ $$\gdef \D {\,\mathrm{d}} $$ $$\gdef \deriv #1 #2 {\frac{\D #1}{\D #2}}$$ $$\gdef \pd #1 #2 {\frac{\partial #1}{\partial #2}}$$ $$\gdef \set #1 {\left\lbrace #1 \right\rbrace} $$

Lecture part A

In this section we discuss the various architectures used in NLP applications, beginning with CNNs, RNNs, and eventually covering the state of-the art architecture, transformers. We then discuss the various modules that comprise transformers and how they make transformers advantageous for NLP tasks. Finally, we discuss tricks that allow transformers to be trained effectively.

Lecture part B

In this section we introduce beam search as a middle ground between greedy decoding and exhaustive search. We consider the case of wanting to sample from the generative distribution (i.e. when generating text) and introduce “top-k” sampling. Subsequently, we introduce sequence to sequence models (with a transformer variant) and backtranslation. We then introduce unsupervised learning approaches for learning embeddings and discuss word2vec, GPT, and BERT.

Practicum

We introduce attention, focusing on self-attention and its hidden layer representations of the inputs. Then, we introduce the key-value store paradigm and discuss how to represent queries, keys, and values as rotations of an input. Finally, we use attention to interpret the transformer architecture, taking a forward pass through a basic transformer, and comparing the encoder-decoder paradigm to sequential architectures.