Estimating Training Data Influence by Tracing Gradient Descent

Abstract

We introduce a method called TracIn that computes the influence of a training example on a prediction made by the model. The idea is to trace how the loss on the test point changes during the training process whenever the training example of interest was utilized. We provide a scalable implementation of TracIn via: (a) a first-order gradient approximation to the exact computation, (b) saved checkpoints of standard training procedures, and (c) cherry-picking layers of a deep neural network. In contrast with previously proposed methods, TracIn is simple to implement; all it needs is the ability to work with gradients, checkpoints, and loss functions. The method is general. It applies to any machine learning model trained using stochastic gradient descent or a variant of it, agnostic of architecture, domain and task. We expect the method to be widely useful within processes that study and improve training data.

arXiv commentNeurIPS 2020

Cited by in corpus (23)

• Competition-Level Code Generation with AlphaCode
• Deep Learning on a Data Diet: Finding Important Examples Early in Training
• Influence Functions in Deep Learning Are Fragile
• Influence-guided Data Augmentation for Neural Tensor Completion
• Understanding Instance-based Interpretability of Variational Auto-Encoders
• Sparse Oblique Decision Trees: A Tool to Understand and Manipulate Neural Net Features
• Learning to Synthesize Programs as Interpretable and Generalizable Policies
• Rationalizing Predictions by Adversarial Information Calibration
• Debugging Tests for Model Explanations
• TracInAD: Measuring Influence for Anomaly Detection
• The Thousand Faces of Explainable AI Along the Machine Learning Life Cycle: Industrial Reality and Current State of Research
• A Unified Framework for Task-Driven Data Quality Management
• Towards Using Data-Influence Methods to Detect Noisy Samples in Source Code Corpora
• Why is "Problems" Predictive of Positive Sentiment? A Case Study of Explaining Unintuitive Features in Sentiment Classification
• On Sample Based Explanation Methods for NLP:Efficiency, Faithfulness, and Semantic Evaluation
• Fortifying Toxic Speech Detectors Against Veiled Toxicity
• Mixture-of-Experts Graph Transformers for Interpretable Particle Collision Detection
• Differential Privacy, Linguistic Fairness, and Training Data Influence: Impossibility and Possibility Theorems for Multilingual Language Models
• EDDA: Explanation-driven Data Augmentation to Improve Explanation Faithfulness
• Adversarial Attacks on Knowledge Graph Embeddings via Instance Attribution Methods
• Influence Tuning: Demoting Spurious Correlations via Instance Attribution and Instance-Driven Updates
• DIVA: Dataset Derivative of a Learning Task
• HERALD: An Annotation Efficient Method to Detect User Disengagement in Social Conversations