Saved in:
Bibliographic Details
Main Authors: Liu, Tian Yu, Golatkar, Aditya, Soatto, Stefano
Format: Preprint
Published: 2023
Subjects:
Online Access:https://arxiv.org/abs/2307.08122
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917666005975040
author Liu, Tian Yu
Golatkar, Aditya
Soatto, Stefano
author_facet Liu, Tian Yu
Golatkar, Aditya
Soatto, Stefano
contents We introduce Tangent Attention Fine-Tuning (TAFT), a method for fine-tuning linearized transformers obtained by computing a First-order Taylor Expansion around a pre-trained initialization. We show that the Jacobian-Vector Product resulting from linearization can be computed efficiently in a single forward pass, reducing training and inference cost to the same order of magnitude as its original non-linear counterpart, while using the same number of parameters. Furthermore, we show that, when applied to various downstream visual classification tasks, the resulting Tangent Transformer fine-tuned with TAFT can perform comparably with fine-tuning the original non-linear network. Since Tangent Transformers are linear with respect to the new set of weights, and the resulting fine-tuning loss is convex, we show that TAFT enjoys several advantages compared to non-linear fine-tuning when it comes to model composition, parallel training, machine unlearning, and differential privacy. Our code is available at: https://github.com/tianyu139/tangent-model-composition
format Preprint
id arxiv_https___arxiv_org_abs_2307_08122
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Tangent Transformers for Composition, Privacy and Removal
Liu, Tian Yu
Golatkar, Aditya
Soatto, Stefano
Machine Learning
We introduce Tangent Attention Fine-Tuning (TAFT), a method for fine-tuning linearized transformers obtained by computing a First-order Taylor Expansion around a pre-trained initialization. We show that the Jacobian-Vector Product resulting from linearization can be computed efficiently in a single forward pass, reducing training and inference cost to the same order of magnitude as its original non-linear counterpart, while using the same number of parameters. Furthermore, we show that, when applied to various downstream visual classification tasks, the resulting Tangent Transformer fine-tuned with TAFT can perform comparably with fine-tuning the original non-linear network. Since Tangent Transformers are linear with respect to the new set of weights, and the resulting fine-tuning loss is convex, we show that TAFT enjoys several advantages compared to non-linear fine-tuning when it comes to model composition, parallel training, machine unlearning, and differential privacy. Our code is available at: https://github.com/tianyu139/tangent-model-composition
title Tangent Transformers for Composition, Privacy and Removal
topic Machine Learning
url https://arxiv.org/abs/2307.08122