English

DYAD: A Descriptive Yet Abjuring Density efficient approximation to linear neural network layers

Machine Learning 2023-12-13 v1 Computation and Language

Abstract

We devise, implement and performance-asses DYAD, a layer which can serve as a faster and more memory-efficient approximate replacement for linear layers, (nn.Linear() in Pytorch). These layers appear in common subcomponents, such as in the ff module of Transformers. DYAD is based on a bespoke near-sparse matrix structure which approximates the dense "weight" matrix W that matrix-multiplies the input in the typical realization of such a layer, a.k.a DENSE. Our alternative near-sparse matrix structure is decomposable to a sum of 2 matrices permutable to a block-sparse counterpart. These can be represented as 3D tensors, which in unison allow a faster execution of matrix multiplication with the mini-batched input matrix X compared to DENSE (O(rows(W ) x cols(W )) --> O( rows(W ) x cols(W ) # of blocks )). As the crux of our experiments, we pretrain both DYAD and DENSE variants of 2 sizes of the OPT arch and 1 size of the Pythia arch, including at different token scales of the babyLM benchmark. We find DYAD to be competitive (>= 90%) of DENSE performance on zero-shot (e.g. BLIMP), few-shot (OPENLM) and finetuning (GLUE) benchmarks, while being >=7-15% faster to train on-GPU even at 125m scale, besides surfacing larger speedups at increasing scale and model width.

Keywords

Cite

@article{arxiv.2312.06881,
  title  = {DYAD: A Descriptive Yet Abjuring Density efficient approximation to linear neural network layers},
  author = {Sarin Chandy and Varun Gangal and Yi Yang and Gabriel Maggiotti},
  journal= {arXiv preprint arXiv:2312.06881},
  year   = {2023}
}

Comments

Accepted at WANT workshop at NeurIPS 2023; code at https://github.com/asappresearch/dyad

R2 v1 2026-06-28T13:47:49.986Z