Abstract:In this paper, we contend that a natural objective of representation learning is to compress and transform the distribution of the data, say sets of tokens, towards a low-dimensional Gaussian mixture supported on incoherent subspaces. The goodness of such a representation can be evaluated by a principled measure, called sparse rate reduction, that simultaneously maximizes the intrinsic information gain and extrinsic sparsity of the learned representation. From this perspective, popular deep network architectures, including transformers, can be viewed as realizing iterative schemes to optimize this measure. Particularly, we derive a transformer block from alternating optimization on parts of this objective: the multi-head self-attention operator compresses the representation by implementing an approximate gradient descent step on the coding rate of the features, and the subsequent multi-layer perceptron sparsifies the features. This leads to a family of white-box transformer-like deep network architectures, named CRATE, which are mathematically fully interpretable. We show, by way of a novel connection between denoising and compression, that the inverse to the aforementioned compressive encoding can be realized by the same class of CRATE architectures. Thus, the so-derived white-box architectures are universal to both encoders and decoders. Experiments show that these networks, despite their simplicity, indeed learn to compress and sparsify representations of large-scale real-world image and text datasets, and achieve performance very close to highly engineered transformer-based models: ViT, MAE, DINO, BERT, and GPT2. We believe the proposed computational framework demonstrates great potential in bridging the gap between theory and practice of deep learning, from a unified perspective of data compression. Code is available at: <a class="link-external link-https" href="https://ma-lab-berkeley.github.io/CRATE" rel="external noopener nofollow">this https URL</a> .

Compressed-Transformer

Joint Structured Pruning and Dense Knowledge Distillation for Efficient Transformer Model Compression

Multi-Dimension Compression of Feed-Forward Network in Vision Transformers

PET: Parameter-efficient Knowledge Distillation on Transformer

An Efficient Transformer Decoder with Compressed Sub-layers

Smartformer: An Intelligent Model Compression Framework for Transformer

A Survey on Transformer Compression

Modular Transformers: Compressing Transformers into Modularized Layers for Flexible Efficient Inference

Quantization-Aware and Tensor-Compressed Training of Transformers for Natural Language Understanding

Blockwise Compression of Transformer-based Models without Retraining

Smartformer: an Intelligent Transformer Compression Framework for Time-Series Modeling

Compressing Transformers: Features Are Low-Rank, but Weights Are Not!

Exploring Extreme Parameter Compression for Pre-trained Language Models

Train Large, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers

Bag of Tricks for Optimizing Transformer Efficiency

White-Box Transformers via Sparse Rate Reduction: Compression Is All There Is?

Transformer on a Diet

Compressing Transformer-based self-supervised models for speech processing

Efficient Encoder-Decoder Transformer Decoding for Decomposable Tasks

A Comparative Analysis of Task-Agnostic Distillation Methods for Compressing Transformer Language Models

Transformer Compression via Subspace Projection