Pretrained Transformers as Universal Computation Engines Lu et al. 2021

Pravish Sainath

Directions - Compare Representations

Comparison of VVS and CNN layer representations

How do pretrained transformers generalize to other modalities with minimal fine tuning?

Background

Transformer - Architecture Recap

​

​

Objective

An investigation on -��Pretrained language models (transformer models that predict next token)

1. These work well on other language tasks that they were not explicitly trained for
2. Could there be a more generic mechanism learned by such networks?
3. How can we analyze these patterns of computations performed in transformers by decoupling the effect of the learned representations?

Hypothesis

The self-attention layers of transformers pretrained on a data-rich modality are useful for arbitrary data sequences, enabling downstream transfer to different modalities.

​

Pretrained LMs generalize to other modalities

Are language models inherently capable of universal computation?

​

universal computation - �the ability to learn representations for predictive learning across diverse modalities

Method

Tasks

To test universal computation

• Numerical Computation
• Image Classification
• Homology

Tasks - Numerical Computation

Bit Memory

• The Model is shown 5 bitstrings each of length 1000�
• Finally, the model is shown a masked version of one of the bitstrings�
• It is expected to produce the original bitstring�
• The bitstrings are broken up into sequences of length 50, so that the models are fed 120 tokens of dimension 50

Tasks - Numerical Computation

Bit XOR

• The Model is shown 2 bitstrings each of length 5�
• It is expected to produce the bitwise XOR of the two bitstrings�
• The models are fed 10 tokens of dimension 1 (one bit at a time)

Tasks - Numerical Computation

ListOps

• The Model is shown a sequence of list operations (max, min, etc.)�
• It is expected to parse the expression and produce the output �
• The models are fed 512 tokens of dimension 15 (one token at a time)��Example�Input = [MAX 4 3 [ MIN 2 3 ] 1 0 ]�Output = 4

Tasks - Image Classification

CIFAR-10

• The Model is shown a sequence of 4x4 image patches�
• It is expected to produce the image category (0-9) as output �
• The models are fed 64 tokens of dimension 16 (one patch at a time)

CIFAR-10 LRA (Long Range Arena Benchmark)

• Similar to above but with grayscale and flattened vectors with 1-dim tokens�(longer sequence without much spatial inductive bias)��

Tasks - Homology

Remote Homology Detection

• The Model is shown a sequence of amino acids (protein)�
• It is expected to produce the fold prediction as output �
• The models are fed 1024 tokens of dimension 25 (one token at a time)��

Architecture

Pretrained GPT-2 Model�Base: 768-dim embeddings with 12 layers

• Output Layer
• Input Layer
• Layer Norm
• Positional Embeddings
• Self-Attention & Feed-Forward Layers

Operation inside a Transformer

​

Universality of Operation inside a Transformer

​

Frozen Pretrained Transformer (FPT)

​

Comparison

Performance in these tasks of different architectures

• FPT
• Full Transformer
• Full LSTM

Results

Performance comparison of FPT with others

​

1 Can PLMs transfer to different modalities?

• FPT achieves comparable performance to Full Transformer�
• This indicates that the internal computation of a Transformer is modality-agnostic with learned representations�
• Superior (100%) performance in the bit tasks indicates large memory of FPT�

​

2 Importance of the pretraining modality?

• The significance of pretraining vs the transformer architecture �
• Pretraining with language fairly significantly superior than most other methods�
• Pretrained ViT fares slightly better for vision tasks, it is worse for homology�
• Any pretraining is better than randomly intitalized Transformers

​

3 Transformer vs LSTM Architecture

• LSTM finetuned similar to FPT
• Transformers (FPT) outperform LSTMs across all tasks�
• Transformers are comparable to LSTMs with architectural improvements (like residual connections, etc.) in some tasks. �Positional Embeddings + Residual connections comparable to Transfo�
• Demonstrates the power of Self-Attention mechanism

​

4 Compute Efficiency over random initialization

• Efficiency measured as the number of gradient steps to converge for FPT vs random transformer models�
• FPT converges faster�
• Language pretrainining offers compute benefits for non-language tasks

​

5 Are Frozen attention layers modality-specific?

​

• FPT attends to semantically meaningful patterns in the data�
• Visualizing the first layer attention weights (softmax of query-key dot product) helps understand this relationship�
• FPT gives an interpretable attention pattern despite not training the self-attention layers themselves.�
• The bit tasks are more interpretable than the rest

​

5 Are Frozen attention layers modality-specific?

5 Are Frozen attention layers modality-specific?

5 Are Frozen attention layers modality-specific?

6 Freezing the Transf - overfitting / underfitting

• FPT underfit the data vs Full Transformers => They can improve further by expanding model capacity�
• FPT provides generalizable task representations�
• Other Transformers overfit poorly in low-data regimes

​

7 Scaling of performance with model size

• Most parameters are in the transformer layers (self-attention) �
• Full transformer exhibits more overfitting and divergence during training with larger models�
• FPT - increasing model size stably increases the capacity of the models. �
• Likely to scale as we move towards larger models and higher-data regimes.

​

8 Is performance better because of better init?

• Layer-wise mean and standard deviation from the pretrained model is used to initialize a random transformer to see if better stats of initialization improves the performance�
• Using pretrained stats improves performance but not across all tasks�
• PLMs have superior performance across tasks

​

8 Is performance better because of better init?

• Most parameters are in the transformer layers (self-attention) �
• Full transformer exhibits more overfitting and divergence during training with larger models�
• FPT - increasing model size stably increases the capacity of the models. �
• Likely to scale as we move towards larger models and higher-data regimes.

​

9 Finetuning Transformer Output layer only?

• Fix a randomly initialized input layer and freeze all parts of the model except for the output. (similar to resevoir computing/echo state network)�
• Speedups significant and performance differences observed�
• Performance significantly degrades and the models also exhibit overfitting �(likely without regularization/dropout)

​

10 Finetuning Transformer Output layer only?

• Tokens need to mix and form interesting representations useful for downstream tasks�
• Importance of the depth of the transformer for generating representations which “mix” tokens.�
• Less layers and parameters are random => Less likely for tokens to be mixed well Increasing #layers increases chances of mixing�

​

10 Finetuning Transformer Output layer only?

With finetuning layer norm

​

10 Finetuning Transformer Output layer only?

Without finetuning layer norm

​

11 Performance improves as more params trained?

• In practical applications, it could be better to choose a more specialized finetuning scheme or add more trainable parameters.�
• Investigate additionally finetuning the self-attention and feedforward layers, which were previously frozen. �
• Add them to the list of parameters finetuned, without changing the optimization or learning rate scheme.�
• Finetuning the feedforward layers can improve performance, but finetuning the attention layers can lead to divergence.�

​

11 Performance improves as more params trained?

12 Which model params to finetune?

• Orthogonal initialization is important when input parameters are not trained�
• The layer norm parameters most important for finetuning.

Discussion

Transformers in multimodal settings

1. A single model could learn a variety of multimodal tasks with an attention architecture - they use distinct transformers to embed different modalities�Eg - Transformers for multimodal predictive tasks, such as images and text in ViLBERT and CLIP �
2. OpenAI found that some neurons learned by CLIP are activated by a particular semantic concept, regardless of language or picture input. �
3. Using FPT similar to DALL-E which uses a single transformer to embed both the image and text modalities generating a “universal latent space” that projects any type of input into a single latent space.�
4. Such a latent space would be useful for a model that could learn from many sources of supervision.

Transformers in transfer settings

• Many works for in-modality transfers such as ViT, T5, CL�
• CLIP showed that training on text in addition to images could allow for zero-shot classification via providing downstream labels as text. �
• Hernandez et al. (2021) do a thorough investigation of transfer with language pretraining, notably showing transfer from English to Python�
• Pretraining and fine tuning of transformer models - using adapter networks, etc.�

​

Self Attention Layers as Optimization Steps

• A single transformer self-attention block can be trained to perform an optimization step towards finding a stationary point, representing the solution to the task�
• The self-attention layer is a gradient step in a Hopfield network with a learning rate of 1, thus transformers are capable of storing and retrieving a large amount of patterns with an implicit energy function.�
• Similar to function overloading in programming where the key-value pair is the function signature�

​

Self Attention Layers as Optimization Steps

Types of fixed points in Hopfield Net

determined by how the pattern x_i is separated from others patterns:

1. a global fixed point n: �no separation of a pattern from the others����
2. a fixed point close to a single pattern: �pattern is separated from other patterns

��

​

• metastable state: �some patterns are similar to each other and �well separated from all other vectors.

​

Global Workspace Theory

• Finetuning the input and output layers to perform �generalizable computation is similar to the �Global Workspace Theory (Baars,1993)�
• GWT - there is a “blackboard” that different parts �of the brain send data to; �frozen language model as being a blackboard in this�setting.�
• Language might also be a natural choice of model �for this blackboard

​

​

Reservoir Computing

• ESN - A random RNN is frozen and only the output readout layer is trained.�Advantage = very fast to train as it is unnecessary to backpropagate over time.�
• ESNs are recurrent allowing the outputs of the random frozen network to modulate future inputs. �
• In FPT the input and positional embeddings are fine tuned, which allow the inputs to the frozen network to adapt to a particular modality/for a query to the frozen network to be learned.

​

​

Points of Discussion

• General-reasoning abilities of Transformer Pretrained Language Models
• Unified representations in Transformer Pretrained Language Models
• Something computationally special about language?�Is this emergence of universal computations because of the usage of language model in the pretraining? Or would any sufficiently diverse sequence dataset yield similar results?
• Language as a medium for cognition?�The neurons in the brain encode generic patterns and can specialize

​