The paper simplifies influence computations for large-scale machine learning models.
problem Improving training efficiency and accuracy in large-scale models.
method Study influence functions, define memorization, simplify computations.
result Influence functions can be practical for large-scale models, indicating memorization.
Proposes a method to explain deep neural networks by identifying representer points in the training set.
problem Explaining the predictions of deep neural networks.
method Identifying representer points in the training set to decompose neural network predictions.
result Provides a deeper understanding of neural network predictions through positive and negative representer values.
Causal influence measures for machine learnt classifiers shed light on the reasons behind classification, and aid in identifying influential input features and revealing their biases. However, such analyses involve evaluating the classifier using datapoints that may be atypical of its training distribution. Standard me…
New method estimates data influence efficiently by leveraging test samples.
problem Efficiently estimating influence of training data on model predictions.
method Mirrored Influence Hypothesis, forward pass for test samples.
result Significant improvement in efficiency over existing methods.
Captures data influence changes during training.
problem Traditional influence functions fail for modern training methods.
method Formalized trajectory-specific LOO influence, using data value embedding.
result Data influence varies by training stage, early and late stages have greater impact.
Influence functions accurately measure group effects in large datasets.
problem Measuring the effect of large groups of training points in models.
method First-order Taylor approximation of influence functions, tested on various real-world datasets.
result Influence functions correlate well with actual group effects, even with large errors.
Study examines influence diagnostics in high-dimensional M-estimation.
problem Understanding influence diagnostics in high-dimensional settings.
method Characterized the distribution of leave-one-out influences in high-dimensional Gaussian M-estimation.
result The distribution of influences converges to a limiting measure in high-dimensional settings.
In2Core selects a coreset for efficient LLM fine-tuning with reduced data.
problem Costly fine-tuning of large language models due to extensive parameters and data requirements.
method Analyzes model gradients to estimate training sample influence, optimizing for efficiency.
result Achieves similar performance with 50% of training data using In2Core.
How can we explain the predictions of a black-box model? In this paper, we use influence functions -- a classic technique from robust statistics -- to trace a model's prediction through the learning algorithm and back to its training data, thereby identifying training points most responsible for a given prediction. To …
Proposes second-order influence functions for identifying influential groups in test-time predictions.
problem Identifying influential groups in test-time predictions for black-box models.
method Second-order approximations of the effect of removing a group of training samples on model predictions.
result Improves the correlation between computed influence values and ground truth values for linear models.
AVA combines feature attribution methods for better model explanations.
problem Improving feature attribution methods for machine learning models.
method AVA: Aggregate Valuation of Antecedents, fusing antecedent event influence and value attribution.
result AVA provides better local and global model explanations.
Efficient method for choosing data points in machine learning models.
problem Choosing optimal data points for semi-supervised learning.
method Computing a 'regularity tangent' vector to measure model complexity and influence.
result The method efficiently calculates a measure of model complexity and data point influence.
The paper extends influence functions to sequence tagging tasks for better model interpretability.
problem Lack of interpretability methods for sequence tagging models.
method Define and compute influence of training instance segments on test segment predictions.
result The segment influence method tracks with true influence and identifies annotation errors.
Proposes DeGLIF to denoise graph data for label noise robustness.
problem Label noise in graph data makes node classification challenging.
method Uses leave-one-out influence function to denoise graph data.
result DeGLIF improves accuracy in node classification on noisy datasets.
Paper presents a new way to estimate model changes without full model evaluation.
problem Efficiently estimating changes in model parameters and outputs due to data point removal.
method Dual representation of influence functions for linearizable models, reducing computational complexity.
result The dual representation can be an efficient alternative to original influence functions, especially for large models.
The paper analyzes methods to identify influential data points in deep models.
problem Interpreting deep learning models and debugging datasets.
method Curated experiments to analyze influence of data points on classifiers.
result Training loss-based sample selection outperformed other methods in detecting mislabels.
A new method, VIF, calculates influence for non-decomposable losses efficiently.
problem Efficiently calculating influence for complex machine learning models with non-decomposable losses.
method Revisiting influence function from robust statistics, proposing Versatile Influence Function (VIF) for any non-decomposable loss.
result VIF method is up to 10^3 times faster than brute-force methods and closely matches influence results.
Influence functions are inaccurate in deep learning models, especially for deeper networks.
problem Inaccuracies in influence functions in deep learning models.
method Empirical study of influence functions in neural network models trained on various datasets.
result Influence estimates are often erroneous for deeper networks and require regularization.
DataInf efficiently approximates data influence in large models, improving transparency and identifying mislabeled data.
problem Efficiently estimating data influence in large-scale models like LoRA-tuned LLMs and diffusion models.
method DataInf uses a closed-form expression to approximate influence scores efficiently.
result DataInf outperforms existing methods in computational and memory efficiency, accurately identifying influential data points.
Develops a faster model selection method using influence functions.
problem Time-consuming and computationally inefficient model selection.
method Model-specific data subsampling guided by influence functions.
result Empirically shows improved selection of high-quality models.
TracIn measures how training examples affect model predictions.
problem Measuring the influence of training examples on model predictions.
method Tracing changes in loss during training when a specific example is used.
result TracIn provides a scalable method to estimate example influence.
Identifies minimal training subset to flip a prediction.
problem Flipping predictions in machine learning models.
method Extended influence function for relabeling minimal subset.
result Relabeling fewer than 2% of training points can flip a prediction.
Study shows influence functions are poor for neural networks but useful for identifying influential examples.
problem Influence functions misalign with leave-one-out retraining in neural networks.
method Decomposed the discrepancy into five terms and studied their contributions across different architectures and datasets.
result Influence functions are a good approximation to the proximal Bregman response function (PBRF), useful for identifying influential examples.
Improves understanding of neural network predictions using influence functions.
problem Challenges in understanding neural network predictions.
method Utilized NTK theory to calculate influence functions for over-parameterized neural networks.
result Proved that the approximation error of IF can be arbitrarily small in the over-parameterized regime.
Develops a method to audit indirect feature influence in complex models.
problem Auditing indirect feature influence in complex, black-box models.
method Disentangled influence audits using disentangled representations.
result Can detect proxy features and show which ones affect model outcomes most.
Dynamic Influence Tracker measures changing sample importance during model training.
problem Static influence measurements during training overlook how sample importance varies over time.
method Dynamic Influence Tracker (DIT) captures time-varying sample influence across arbitrary time windows.
result DIT reveals distinct learning phases with shifting priorities and detects corrupted samples more efficiently.
The paper proposes a method to improve fairness in machine learning models without refitting.
problem Mitigating biases in machine learning models that disadvantage certain groups.
method Infinitesimal jackknife-based approach to drop selected training data points.
result The intervention improves fairness without significantly reducing predictive performance.
A scheme incentivizes agents to provide high-quality data early in federated learning.
problem Incentivizing high-quality data in federated learning.
method Characterization of influence on loss function, payment structure design, efficient influence approximation.
result The scheme ensures truthful data reporting as a game-theoretic equilibrium and is robust to mixtures of truthful and heuristic data reports.
New experiments show deep networks benefit from memorizing rare data points.
problem Understanding why neural networks perform well on training data.
method Estimating influence and memorization values of training examples.
result Deep networks benefit from memorizing rare data points for good performance.
Retraining stabilizes model influence on data.
problem Performativity in predictive models leads to feedback loops.
method Developed the stable signal principle to address retraining dynamics.
result Repeated risk minimization converges geometrically to stable signal direction.
RelatIF selects more intuitive training examples for explaining model predictions.
problem Influence functions identify outliers as explanatory examples, leading to poor explanations.
method RelatIF separates global and local influence, optimizing for local relative to global effects.
result Examples selected by RelatIF are more intuitive than those from influence functions.
Efficiently reduces data augmentation size with similar accuracy.
problem Explosive growth in dataset size due to data augmentation.
method Subsampling policies based on model influence and loss.
result Achieves a 90% reduction in augmentation set size while maintaining accuracy.
Improved scalability and interpretability in training data attribution.
problem Identifying which training data drives specific behaviors, especially unintended ones.
method Leveraging interpretable structures within the model to attribute model behavior to semantic directions, not individual test examples.
result Simple probe-based attribution methods are first-order approximations of Concept Influence that achieve comparable performance while being over an order-of-magnitude faster.
AdaBoost's success explained through noise influence measure.
problem Understanding why AdaBoost is a successful classifier.
method Introduced a measure of noise influence (ION) to explain AdaBoost's success.
result ION decreases with iteration number and base learner complexity.
Improves understanding of PWS by calculating influence of sources and data.
problem Understanding the influence of each component in PWS.
method Proposes source-aware Influence Function (IF) to decompose and calculate influence.
result Improves end model's generalization performance and identifies mislabeling.
Neuro-robotics model shows gestures improve counting ability.
problem Improving counting ability in robots and children.
method Introduces a neuro-robotics model trained with pointing data, tests various model and training conditions.
result Model's performance aligns with human children's learning to count.
There is an especially strong need in modern large-scale data analysis to prioritize samples for manual inspection. For example, the inspection could target important mislabeled samples or key vulnerabilities exploitable by an adversarial attack. In order to solve the "needle in the haystack" problem of which samples t…
New method reveals how training data influence diffusion model outputs.
problem Difficulty in assessing training data impact on diffusion model outputs.
method Use of ensembles trained on carefully engineered splits of training data to identify influential training examples.
result Demonstrated the viability of ensembles as generative models and validity of assessing influence.
Proposes a new framework to disentangle event influences in MTPP.
problem Underexplored how individual events influence overall dynamics over time.
method Decoupled MTPP framework using Neural Ordinary Differential Equations (Neural ODEs).
result Significantly improves performance on real-life datasets compared to state-of-the-art methods.
The paper explores how structured representations influence learning dynamics in neural networks.
problem Understanding the training dynamics of deep neural networks.
method Investigates a family of enriched transformation layers with constrained pathways and adaptive corrections.
result Improved robustness, smoother optimization, and scalable depth behavior are achieved through structured representations.
Influence functions help study large language model generalization, revealing surprising decay patterns.
problem Understanding and mitigating risks in large language models (LLMs).
method Eigenvalue-corrected Kronecker-Factored Approximation (EK-FAC) to scale influence functions to LLMs.
result Influences decay to near-zero when key phrases order is flipped, revealing a surprising limitation.
GGDA simplifies DA for large models, speeding up attribution by up to 50x.
problem Computational intensity of existing DA methods limits their applicability to large-scale models.
method Generalized Group Data Attribution (GGDA) framework attributing to groups of training points.
result GGDA achieves up to 50x speedups over standard DA methods while maintaining effectiveness.
New IF method improves accuracy in deep neural networks with noisy data.
problem Inaccurate influence estimates in deep neural networks, especially with noisy data.
method Established a connection between influence estimation error, validation set risk, and sharpness, introducing a novel estimation form for flat validation minima.
result Our novel Influence Function approach provides more accurate influence estimates, validated across various tasks.
Develops new methods to evaluate data influence in SAM for improved model training.
problem Challenges in mislabeled noisy data and privacy concerns in SAM.
method Two innovative data valuation methods based on influence functions (IF) for SAM.
result Demonstrates effectiveness in identifying mislabeled data and enhancing interpretability.
Advances in deep learning for spatio-temporal event modeling.
problem Limitations of traditional parametric models in capturing nonstationary dynamics.
method Integration of deep neural architectures to model conditional intensity function and influence kernels.
result Deep influence kernel approach enhances expressiveness and statistical explainability.
Gradient rollback explains neural models by identifying influential training examples.
problem Explain predictions of neural black-box models, especially in applications requiring user trust.
method Gradient rollback, a general approach for influence estimation applicable to neural models with limited parameter updates.
result Gradient rollback provides faithful explanations for neural models, including knowledge graph embedding methods.
Paper introduces a novel point process model for graph data using GNNs.
problem Modeling discrete event data over graphs with influence kernel.
method Combines Hawkes kernel and Graph Neural Networks (GNN) for event prediction.
result Achieves superior predictive performance compared to state-of-the-art.
Early SGD hyperparameters affect deep neural network training, showing a break-even point.
problem Understanding how early SGD hyperparameters influence deep neural network training.
method Analysis of stochastic gradient descent (SGD) hyperparameters and their effects on the optimization trajectory.
result A break-even point exists where SGD implicitly regularizes the loss surface and improves gradient conditioning.