Timezone: »
Attention based Transformer architecture has enabled significant advances in the field of natural language processing. In addition to new pre-training techniques, recent improvements crucially rely on working with a relatively larger embedding dimension for tokens. Unfortunately, this leads to models that are prohibitively large to be employed in the downstream tasks. In this paper we identify one of the important factors contributing to the large embedding size requirement. In particular, our analysis highlights that the scaling between the number of heads and the size of each head in the current architecture gives rise to a low-rank bottleneck in attention heads, causing this limitation. We further validate this in our experiments. As a solution we propose to set the head size of an attention unit to input sequence length, and independent of the number of heads, resulting in multi-head attention layers with provably more expressive power. We empirically show that this allows us to train models with a relatively smaller embedding dimension and with better performance scaling.
Author Information
Srinadh Bhojanapalli (Google AI)
Chulhee (Charlie) Yun (MIT)
Ankit Singh Rawat (Google)
Sashank Jakkam Reddi (Google)
Sanjiv Kumar (Google Research, NY)
More from the Same Authors
-
2020 Poster: Does label smoothing mitigate label noise? »
Michal Lukasik · Srinadh Bhojanapalli · Aditya Menon · Sanjiv Kumar -
2020 Poster: Accelerating Large-Scale Inference with Anisotropic Vector Quantization »
Ruiqi Guo · Philip Sun · Erik Lindgren · Quan Geng · David Simcha · Felix Chern · Sanjiv Kumar -
2020 Poster: SCAFFOLD: Stochastic Controlled Averaging for Federated Learning »
Sai Praneeth Reddy Karimireddy · Satyen Kale · Mehryar Mohri · Sashank Jakkam Reddi · Sebastian Stich · Ananda Theertha Suresh -
2020 Poster: Federated Learning with Only Positive Labels »
Felix Xinnan Yu · Ankit Singh Rawat · Aditya Menon · Sanjiv Kumar -
2019 Poster: Escaping Saddle Points with Adaptive Gradient Methods »
Matthew Staib · Sashank Jakkam Reddi · Satyen Kale · Sanjiv Kumar · Suvrit Sra -
2019 Oral: Escaping Saddle Points with Adaptive Gradient Methods »
Matthew Staib · Sashank Jakkam Reddi · Satyen Kale · Sanjiv Kumar · Suvrit Sra -
2019 Poster: Learning a Compressed Sensing Measurement Matrix via Gradient Unrolling »
Shanshan Wu · Alexandros Dimakis · Sujay Sanghavi · Felix Xinnan Yu · Daniel Holtmann-Rice · Dmitry Storcheus · Afshin Rostamizadeh · Sanjiv Kumar -
2019 Oral: Learning a Compressed Sensing Measurement Matrix via Gradient Unrolling »
Shanshan Wu · Alexandros Dimakis · Sujay Sanghavi · Felix Xinnan Yu · Daniel Holtmann-Rice · Dmitry Storcheus · Afshin Rostamizadeh · Sanjiv Kumar -
2018 Poster: Loss Decomposition for Fast Learning in Large Output Spaces »
En-Hsu Yen · Satyen Kale · Felix Xinnan Yu · Daniel Holtmann-Rice · Sanjiv Kumar · Pradeep Ravikumar -
2018 Oral: Loss Decomposition for Fast Learning in Large Output Spaces »
En-Hsu Yen · Satyen Kale · Felix Xinnan Yu · Daniel Holtmann-Rice · Sanjiv Kumar · Pradeep Ravikumar -
2017 Poster: Stochastic Generative Hashing »
Bo Dai · Ruiqi Guo · Sanjiv Kumar · Niao He · Le Song -
2017 Talk: Stochastic Generative Hashing »
Bo Dai · Ruiqi Guo · Sanjiv Kumar · Niao He · Le Song -
2017 Poster: Distributed Mean Estimation with Limited Communication »
Ananda Theertha Suresh · Felix Xinnan Yu · Sanjiv Kumar · Brendan McMahan -
2017 Talk: Distributed Mean Estimation with Limited Communication »
Ananda Theertha Suresh · Felix Xinnan Yu · Sanjiv Kumar · Brendan McMahan