Estimating Training Data Influence by Tracking Gradient Descent

By Garima Pruthi, Frederick Liu, Mukund Sundararajan, Satyen Kale. 2020

704 citations recorded by the source corpus.

We introduce a method called TrackIn that computes the influence of a training example on a prediction made by the model, by tracking 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 TrackIn via a combination of a few key ideas: (a) a first-order approximation to the exact computation, (b) using random projections to speed up the computation of the first-order approximation for large models, (c) using saved checkpoints of standard training procedures, and (d) cherry-picking layers of a deep neural network. An experimental evaluation shows that TrackIn is more effective in identifying mislabelled training examples than other related methods such as influence functions and representer points. We also discuss insights from applying the method on vision, regression and natural language tasks.

Subfields: Management Science and Operations Research, Organizational Behavior and Human Resource Management, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Management Science and Operations Research

In Organizational Behavior and Human Resource Management

In Optimization & Training

Citations in the corpus

Measuring Data (2022, background): “…and limited when it is done at all (Jo and Gebru, 2020), we are currently in the midst of creating, sharing, and training models based on datasets that we know very little about, contributing to behaviour from models that we can neither predict nor trace (Akyürek et al., 2022; Pruthi et al., 2020).”
What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation (2020, background): “Unlike some of the existing techniques [19, 30, 24] it is also completely model-agnostic and is itself easy to explain.”
LESS: Selecting Influential Data for Targeted Instruction Tuning (2024, methodology): “Our formulation of influence has been used in identifying mislabeled examples (Pruthi et al., 2020), analyzing memorization effects (Feldman & Zhang, 2020), and deriving various inter-pretability insights (Madsen et al., 2022), but it does not permit counterfactual reasoning.”
On Memorization in Probabilistic Deep Generative Models (2021, background): “Concurrent work in [43] focuses on influence functions for variational autoencoders based on the “TracIn” approximation of [44] computed over the training run.”
Post-hoc Interpretability for Neural NLP: A Survey (2021, methodology): “[88] repeat the label-correction experiment of inluence functions and Representer Point Selection, and ind that their method can better select mislabeled observations.”
Revisiting Methods for Finding Influential Examples (2021, methodology): “In addition, two simpler heuristics, Grad-Dot and Grad-Cos (Charpiat et al. 2019), are commonly used as baselines (Pruthi et al. 2020a).”
Tracing Knowledge in Language Models Back to the Training Data (2022, methodology): “Therefore, we turn to a more recent influence technique that has demonstrated both better tractability and stronger empirical results: TracIn (Pruthi et al., 2020), which seeks to estimate influence by asking a credit-assignment question rather than a counterfactual perturbation question.”
DataComp: In search of the next generation of multimodal datasets (2023, methodology): “[102] Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan.”
Explainability for Large Language Models: A Survey (2023, methodology): “Based on such an approach, TracIn derives the influence of training points using the calculus theorem with checkpoints during the training process [Pruthi et al. 2020].”
Influence Tuning: Demoting Spurious Correlations via Instance Attribution and Instance-Driven Updates (2021, methodology): “Apart from influence functions (Koh and Liang, 2017) and TracIn (Pruthi et al., 2020b) that are already introduced, other instance attribution methods include representer point selection (Yeh et al.”
Learning to Synthesize Programs as Interpretable and Generalizable Policies (2021, background): “[96] Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan.”
Influence Functions in Deep Learning Are Fragile (2020, methodology): “Influential training examples can also be obtained by tracking the change in loss for a test-prediction through model-checkpoints, which are stored during the training time [17].”
Debugging Tests for Model Explanations (2020, methodology): “We only focused on feature attributions: however, other methods such as approaches based on ‘concept’ activation [28], model representation dissection [9], and training point ranking [32, 48, 69] might be more suited to the debugging tasks studied here.”
Datamodels: Predicting Predictions from Training Data (2022, methodology): “…sets R(x) as follows: (a) setting R(x) to be the nearest k training examples to the target example x using the baseline methods described in Appendix E.2: influence function (Koh & Liang, 2017), TracIn (Pruthi et al., 2020), or distance in pre-trained representation space (Bengio et al., 2013).”
TRAK: Attributing Model Behavior at Scale (2023, neutral): “Number of training examples removed D ro p in co si ne si m ila ri ty CLIP similarity TracIn (Pruthi et al., 2020) TRAK (ours)”
Machine Unlearning for Random Forests (2020, background): “Instance-Based Interpretability A popular form of interpretability looks at how much each training instance contributes to a given prediction (Koh & Liang, 2017; Yeh et al., 2018; Sharchilev et al., 2018; Pruthi et al., 2020; Chen et al., 2021).”