Training the Untrainable: Introducing Inductive Bias via Representational Alignment

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Subramaniam, Vighnesh, Mayo, David, Conwell, Colin, Poggio, Tomaso, Katz, Boris, Cheung, Brian, Barbu, Andrei
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911228582952960
author Subramaniam, Vighnesh
Mayo, David
Conwell, Colin
Poggio, Tomaso
Katz, Boris
Cheung, Brian
Barbu, Andrei
author_facet Subramaniam, Vighnesh
Mayo, David
Conwell, Colin
Poggio, Tomaso
Katz, Boris
Cheung, Brian
Barbu, Andrei
contents We demonstrate that architectures which traditionally are considered to be ill-suited for a task can be trained using inductive biases from another architecture. We call a network untrainable when it overfits, underfits, or converges to poor results even when tuning their hyperparameters. For example, fully connected networks overfit on object recognition while deep convolutional networks without residual connections underfit. The traditional answer is to change the architecture to impose some inductive bias, although the nature of that bias is unknown. We introduce guidance, where a guide network steers a target network using a neural distance function. The target minimizes its task loss plus a layerwise representational similarity against the frozen guide. If the guide is trained, this transfers over the architectural prior and knowledge of the guide to the target. If the guide is untrained, this transfers over only part of the architectural prior of the guide. We show that guidance prevents FCN overfitting on ImageNet, narrows the vanilla RNN-Transformer gap, boosts plain CNNs toward ResNet accuracy, and aids Transformers on RNN-favored tasks. We further identify that guidance-driven initialization alone can mitigate FCN overfitting. Our method provides a mathematical tool to investigate priors and architectures, and in the long term, could automate architecture design.
format Preprint
id arxiv_https___arxiv_org_abs_2410_20035
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Training the Untrainable: Introducing Inductive Bias via Representational Alignment
Subramaniam, Vighnesh
Mayo, David
Conwell, Colin
Poggio, Tomaso
Katz, Boris
Cheung, Brian
Barbu, Andrei
Machine Learning
Artificial Intelligence
Computation and Language
We demonstrate that architectures which traditionally are considered to be ill-suited for a task can be trained using inductive biases from another architecture. We call a network untrainable when it overfits, underfits, or converges to poor results even when tuning their hyperparameters. For example, fully connected networks overfit on object recognition while deep convolutional networks without residual connections underfit. The traditional answer is to change the architecture to impose some inductive bias, although the nature of that bias is unknown. We introduce guidance, where a guide network steers a target network using a neural distance function. The target minimizes its task loss plus a layerwise representational similarity against the frozen guide. If the guide is trained, this transfers over the architectural prior and knowledge of the guide to the target. If the guide is untrained, this transfers over only part of the architectural prior of the guide. We show that guidance prevents FCN overfitting on ImageNet, narrows the vanilla RNN-Transformer gap, boosts plain CNNs toward ResNet accuracy, and aids Transformers on RNN-favored tasks. We further identify that guidance-driven initialization alone can mitigate FCN overfitting. Our method provides a mathematical tool to investigate priors and architectures, and in the long term, could automate architecture design.
title Training the Untrainable: Introducing Inductive Bias via Representational Alignment
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2410.20035