TLDR: A new framework called f-INE uses hypothesis testing to reliably estimate the influence of individual data points on machine learning models, even when training randomness causes instability in other methods. It quantifies influence by statistically distinguishing model behavior with and without a data point, and an efficient algorithm computes this in a single training run. Experiments show f-INE effectively identifies mislabeled data and poisoned LLM instructions with greater consistency than previous approaches.
Machine learning models are becoming increasingly complex, and understanding how individual training data points affect their final behavior is crucial for debugging, improving, and ensuring fairness. This field, known as influence estimation, helps us identify which samples are most impactful. However, a significant challenge arises from “training randomness” – factors like random initialization, data shuffling, and batching can cause the same data point to appear influential in one training run and irrelevant in another. This inconsistency makes it difficult to trust and apply these methods in real-world scenarios like data curation or identifying problematic samples.
A new research paper, titled “f-INE: A Hypothesis Testing Framework for Estimating Influence under Training Randomness,” addresses this critical issue. Authored by Subhodip Panda, Dhruv Tarsadiya, Shashwat Sourav, Prathosh A.P., and Sai Praneeth Karimireddy, the paper introduces a novel approach to influence estimation that explicitly accounts for the inherent randomness in machine learning training processes.
The Problem with Existing Methods
Current influence estimation techniques, such as Influence Functions, TRAK, and TraceIn, often show significant variability. Imagine trying to decide whether to delete a data point because it seems “low-influence” in one run, only to find it was critical in another. This instability undermines their utility, as decisions based on inconsistent scores can lead to unexpected performance drops or a failure to identify truly harmful data.
Introducing f-influence and f-INE
The core innovation of this paper is the concept of “f-influence,” a new definition of influence grounded in hypothesis testing. Instead of trying to assign a single, potentially unstable score, f-influence re-frames the problem: can we statistically differentiate between a model trained with a specific data point and one trained without it, considering the natural randomness of the training process? If the difference is statistically significant, then the data point is deemed influential.
This framework leads to the “f-INfluence Estimation (f-INE)” algorithm. The influence of a data point is quantified by how easily one can distinguish between two distributions: one representing the model’s behavior when the data point is included in training, and another when it’s excluded. By estimating influence at a distributional level, f-INE inherently captures and accounts for training randomness.
Key Properties and Efficiency
The researchers establish desirable properties for f-influence, including compositionality and asymptotic normality. These properties are crucial because they allow the complex trade-off curves (which describe the difficulty of distinguishing the two distributions) to be characterized by a single, interpretable scalar parameter, µ, known as Gµ-influence. This single parameter provides a total ordering of influence, making it practical for decision-making.
A major practical advantage of f-INE is its efficiency. Unlike naive methods that would require retraining models hundreds of times, f-INE computes this Gµ-influence in a single training run. It achieves this by estimating “single-step influence” using gradient similarity and employing techniques to reduce dependencies among samples collected during training.
Empirical Validation
The paper demonstrates f-INE’s effectiveness across various tasks:
- Identifying Mislabeled Samples: In experiments with the MNIST dataset, f-INE performed comparably to or better than existing methods like TraceIn, TRAK, and Influence Functions in detecting mislabeled examples. Its recall curve was smoother and more predictable, indicating reduced variance in influence scores.
- Attributing LLM Model Behavior: f-INE was scaled to estimate the influence of instruction tuning data on Llama-3.1-8B, a large language model. It reliably detected “poisoned” samples designed to steer the model’s opinions (e.g., negative sentiment towards specific entities). f-INE identified a higher fraction of poisoned instructions in its top-ranked influential points compared to the LESS method.
- Lower Variability: Crucially, f-INE showed significantly lower variability in influence scores across multiple training runs (different random seeds) compared to LESS, confirming its robustness to training randomness.
- Qualitative Analysis: A qualitative case study highlighted f-INE’s ability to detect subtle data poisonings that methods like LESS (which only compare means) might miss. By comparing full distributions, f-INE could identify instances where the distributions were dissimilar even if their means were close.
Also Read:
- LeFCert: Securing Language Models Against Data Poisoning Attacks
- Unmasking Hidden Training Data in LLMs After Reinforcement Learning
Conclusion
By reframing influence estimation as a binary hypothesis test, f-INE provides a statistically rigorous and practical method for understanding data influence in machine learning, even under the unavoidable presence of training randomness. Its ability to provide a reliable, ordered notion of influence, coupled with its computational efficiency, makes it a valuable tool for data cleanup, model debugging, and attributing model behavior in high-stakes applications. This work also establishes a strong connection between influence estimation and privacy auditing, opening new avenues for future research. For more details, you can read the full research paper here.
