11B: Influence Functions

Materials:

Date: Friday, 25-Oct-2024

Pre-work:

1. [tools] DEEL-PUNCC - a python toolbox for Conformal Prediction from DEEL.ai a project for Dependable, Certifiable, Explainable AI for Critical Systems. Checkout the sister projects from DEEl on Bias DEEL INFLUENCIAE, oodeel for OOD, xplique for XAI,
2. White Paper - Machine Learning in Certified Systems

In-Class

Topic: Gradients > IFs > {Debugging, OOD, Robustness, XAI}

Discussion:

1. Cook’s Distance
   • leverage
   • outlier
   • influence
2. Leveraing Influence Functions for Dataset Exploration and Cleaning
3. Interpreting Robust Optimization via Adversarial Influence Functions
4. Data Debugging: TraceIn - Estimating Training Data Influence by Tracing Gradient Descent
5. GradOrth A Simple yet Efficient Out-of-Distribution Detection with Orthogonal Projection of Gradients
6. Axiomatic Attribution for Deep Networks - integrated gradients for xai
7. Understanding Black-box Predictions via Influence Functions - Perhaps, the first applications of Influence Functions to explain Black Box Deep Learning models
8. Influence functions in Machine Learning tasks

Lab

1. IF from pyDVL library for data valuation and influence function computation.
2. OOD: Notebook
3. XAI in LLMs: Notebook Use IFs for XIA on LLMs, Blog
4. XAI with gradients xplique

Post-class

Influence Functions (IFs)

• Theory:
  • Hampel’s IF Seminal Paper - The Influence Curve and its Role in Robust Estimation, from Frank R. Hampel, that started this work
• Computation:
  • Scaling Up Influence Functions
  • Revisiting inverse Hessian vector products for calculating influence functions
  • M-FAC: Efficient Matrix-Free Approximations of Second-Order Information
  • FastIF FastIF: Scalable Influence Functions for Efficient Model Interpretation and Debugging, with some applications in NLP.
• (Early) Application in ML:
  • Understanding Black-box Predictions via Influence Functions - Perhaps, the first applications of Influence Functions to explain Black Box Deep Learning models
  • Influence Functions in Deep Learning Are Fragile
  • Influence Functions for Machine Learning: Nonparametric Estimators for Entropies, Divergences and Mutual Informations
  • Infuence functions in Machine Learning tasks
  • If Influence Functions are the Answer, Then What is the Question?
• Tools:
  • Influenciate from the DEEL Influenciae project
  • pyDVL library for data valuation and influence function computation.

LLMs

1. Studying Large Language Model Generalization with Influence Functions
2. Do Influence Functions Work on Large Language Models?
3. TextGrad Automatic ‘’Differentiation’’ via Text, paper

XAI using Gradients and IFs

1. Axiomatic Attribution for Deep Networks - integrated gradients for xai
2. For others DEEL.ai’ Xplique code paper

Data Debugging

1. Estimating Training Data Influence by Tracing Gradient Descent
2. Leveraing Influence Functions for Dataset Exploration and Cleaning
3. Understanding Black-box Predictions via Influence Functions
4. RelatIF: Identifying Explanatory Training Examples via Relative Influence
5. On the Accuracy of Influence Functions for Measuring Group Effects
6. Representer Point Selection for Explaining Deep Neural Networks
7. Representer Point Selection via Local Jacobian Expansion for Post-hoc Classifier Explanation of Deep Neural Networks and Ensemble Models
8. Understanding Influence Functions and Datamodels via Harmonic Analysis
9. An Automatic Finite-Sample Robustness Metric: When Can Dropping a Little Data Make a Big Difference?
10. LAVA: Data Valuation without Pre-Specified Learning Algorithms

OOD

1. Out-of-Distribution Generalization Analysis via Influence Function
2. How Useful are Gradients for OOD Detection Really? - a negative results on gradients
3. Gradient-Regularized Out-of-Distribution Detection
4. GradOrth A Simple yet Efficient Out-of-Distribution Detection with Orthogonal Projection of Gradients
5. oodeel - a toolbox for OOD detection

Robustness

1. Interpreting Robust Optimization via Adversarial Influence Functions
2. Robust inference: The approach based on influence functions
3. Robust inference by influence functions
4. Generalized Influence Functions and Robustness Analysis

Notes

Gradients is all you need

Modality	Aspect	Topics	Resources
Tabular	OOD	x	x
Tabular	XAI	x	x
Tabular	Bias	x	x
Tabular	CP	x	x
Tabular	Robustness	x	x
Text	OOD	x	x
Text	XAI	x	x
Text	Bias	x	x
Text	CP	x	x
Text	Robustness	x	x
Speech	OOD	x	x
Speech	XAI	x	x
Speech	Bias	x	x
Speech	CP	x	x
Speech	Robustness	x	x
Image	OOD	x	x
Image	XAI	x	x
Image	Bias	x	x
Image	CP	x	x
Image	Robustness	x	x
Video	OOD	x	x
Video	XAI	x	x
Video	Bias	x	x
Video	CP	x	x
Video	Robustness	x	x