HYDRA: Hypergradient Data Relevance Analysis for Interpreting Deep Neural Networks

02/04/2021
by   Yuanyuan Chen, et al.
0

The behaviors of deep neural networks (DNNs) are notoriously resistant to human interpretations. In this paper, we propose Hypergradient Data Relevance Analysis, or HYDRA, which interprets the predictions made by DNNs as effects of their training data. Existing approaches generally estimate data contributions around the final model parameters and ignore how the training data shape the optimization trajectory. By unrolling the hypergradient of test loss w.r.t. the weights of training data, HYDRA assesses the contribution of training data toward test data points throughout the training trajectory. In order to accelerate computation, we remove the Hessian from the calculation and prove that, under moderate conditions, the approximation error is bounded. Corroborating this theoretical claim, empirical results indicate the error is indeed small. In addition, we quantitatively demonstrate that HYDRA outperforms influence functions in accurately estimating data contribution and detecting noisy data labels. The source code is available at https://github.com/cyyever/aaai_hydra_8686.

READ FULL TEXT

page 5

page 10

research
11/21/2019

Neural Network Memorization Dissection

Deep neural networks (DNNs) can easily fit a random labeling of the trai...
research
02/01/2022

Datamodels: Predicting Predictions from Training Data

We present a conceptual framework, datamodeling, for analyzing the behav...
research
12/06/2021

Scaling Up Influence Functions

We address efficient calculation of influence functions for tracking pre...
research
05/05/2021

A Theoretical-Empirical Approach to Estimating Sample Complexity of DNNs

This paper focuses on understanding how the generalization error scales ...
research
12/31/2020

FastIF: Scalable Influence Functions for Efficient Model Interpretation and Debugging

Influence functions approximate the 'influences' of training data-points...
research
10/03/2022

Understanding Influence Functions and Datamodels via Harmonic Analysis

Influence functions estimate effect of individual data points on predict...
research
03/22/2021

Data Cleansing for Deep Neural Networks with Storage-efficient Approximation of Influence Functions

Identifying the influence of training data for data cleansing can improv...

Please sign up or login with your details

Forgot password? Click here to reset