copying a rare token

Logit Lens

The code to reproduce our results can be found here.

It seems like the neuron basically adds the embedding of “ an” to the residual stream, which increases the output probability for “ an” since the unembedding step consists of taking the dot product of the final residual with each token2.

The second purpose of skip connections is specific to transformers — preserving the original input sequence.

Skip connections serve two purposes. The first is that they help keep the gradient smooth, which is a big help for backpropagation. Attention is a filter, which means that when it’s working correctly it will block most of what tries to pass through it.

Once we have the result of our attention step, a vector that includes the most recent word and a small collection of the words that have preceded it, we need to translate that into features, each of which is a word pair. Attention masking gets us the raw material that we need, but it doesn’t build those word pair features. To do that, we can use a single layer fully connected neural network.

The central object in the transformer is the residual stream.

the Elhage et al.(2021) study showing an information-copying role for self-attention.

One of the main features of the high level architecture of a transformer is that each layer adds its results into what we call the “residual stream.”Constructing models with a residual stream traces back to early work by the Schmidhuber group, such as highway networks  and LSTMs, which have found significant modern success in the more recent residual network architecture . In transformers, the residual stream vectors are often called the “embedding.” We prefer the residual stream terminology, both because it emphasizes the residual nature (which we believe to be important) and also because we believe the residual stream often dedicates subspaces to tokens other than the present token, breaking the intuitions the embedding terminology suggests. The residual stream is simply the sum of the output of all the previous layers and the original embedding. We generally think of the residual stream as a communication channel, since it doesn't do any processing itself and all layers communicate through it.

A transformer starts with a token embedding, followed by a series of “residual blocks”, and finally a token unembedding. Each residual block consists of an attention layer, followed by an MLP layer. Both the attention and MLP layers each “read” their input from the residual stream (by performing a linear projection), and then “write” their result to the residual stream by adding a linear projection back in. Each attention layer consists of multiple heads, which operate in parallel.

You see the values of the self-attention weights are computed on the fly. They are data-dependent dynamic weights because they change dynamically in response to the data (fast weights).

This input embedding is the initial value of the residual stream, which all attention layers and MLPs read from and write to.

To see how this plays out, we can continue looking at matrix shapes. Tracing the matrix shape through the branches and weaves of the multihead attention blocks requires three more numbers. d_k: dimensions in the embedding space used for keys and queries. 64 in the paper. d_v: dimensions in the embedding space used for values. 64 in the paper. h: the number of heads. 8 in the paper.

Now, the progression of NLP, as discussed, tells a story. We begin with tokens and then build representations of these tokens. We use these representations to find similarities between tokens and embed them in a high-dimensional space. The same embeddings are also passed into sequential models that can process sequential data. Those models are used to build context and, through an ingenious way, attend to parts of the input sentence that are useful to the output sentence in translation.

The dominant idea is one of attention, by which a representation at a position is computed as a weighted combination of representations from other positions. A common self-supervision objective in a transformer model is to mask out occasional words in a text. The model works out what word used to be there. It does this by calculating from each word position (including mask positions) vectors that represent a query, key, and value at that position. The query at a position is compared with the value at every position to calculate how much attention to pay to each position; based on this, a weighted average of the values at all positions is calculated. This operation is repeated many times at each level of the transformer neural net, and the resulting value is further manipulated through a fully connected neural net layer and through use of normalization layers and residual connections to produce a new vector for each word. This whole process is repeated many times, giving extra layers of depth to the transformer neural net. At the end, the representation above a mask position should capture the word that was there in the original text: for instance, committee as illustrated in Figure 1.

This trick of using a one-hot vector to pull out a particular row of a matrix is at the core of how transformers work.

The source sequence will be pass to the TransformerEncoder, which will produce a new representation of it. This new representation will then be passed to the TransformerDecoder, together with the target sequence so far (target words 0 to N). The TransformerDecoder will then seek to predict the next words in the target sequence (N+1 and beyond).

The transformer model introduces the idea of instead of adding another complex mechanism (attention) to an already complex Seq2Seq model; we can simplify the solution by forgetting about everything else and just focusing on attention.

The selective-second-order-with-skips model is a useful way to think about what transformers do, at least in the decoder side. It captures, to a first approximation, what generative language models like OpenAI's GPT-3 are doing.

The Query word can be interpreted as the word for which we are calculating Attention. The Key and Value word is the word to which we are paying attention ie. how relevant is that word to the Query word.

Other work on interpreting transformer internals has focused mostly on what the attention is looking at. The logit lens focuses on what GPT "believes" after each step of processing, rather than how it updates that belief inside the step.

The attention layer (W in the diagram) computes three vectors based on the input, termed key, query, and value.

Attention is a means of selectively weighting different elements in input data, so that they will have an adjusted impact on the hidden states of downstream layers.

So for each word, we create a Query vector, a Key vector, and a Value vector. These vectors are created by multiplying the embedding by three matrices that we trained during the training process.

We show that BigBird is a universal approximator of sequence functions and is Turing complete,