Augmenting interpretable models with large language models during training (2022-09-23T00:00:00.000000Z)

TL;DR

Two instantiations of Aug-imodels in natural-language processing are explored: Aug-Linear, which augments a linear model with decoupled embeddings from an LLM and Aug-Tree, which augmented a decision tree with LLM feature expansions, which outperform their non-augmented, interpretable counterparts.

Abstract

Recent large language models (LLMs), such as ChatGPT, have demonstrated remarkable prediction performance for a growing array of tasks. However, their proliferation into high-stakes domains and compute-limited settings has created a burgeoning need for interpretability and efficiency. We address this need by proposing Aug-imodels, a framework for leveraging the knowledge learned by LLMs to build extremely efficient and interpretable prediction models. Aug-imodels use LLMs during fitting but not during inference, allowing complete transparency and often a speed/memory improvement of greater than 1000x for inference compared to LLMs. We explore two instantiations of Aug-imodels in natural-language processing: Aug-Linear, which augments a linear model with decoupled embeddings from an LLM and Aug-Tree, which augments a decision tree with LLM feature expansions. Across a variety of text-classification datasets, both outperform their non-augmented, interpretable counterparts. Aug-Linear can even outperform much larger models, e.g. a 6-billion parameter GPT-J model, despite having 10,000x fewer parameters and being fully transparent. We further explore Aug-imodels in a natural-language fMRI study, where they generate interesting interpretations from scientific data.

Authors

R. Caruana

8 papers

Jianfeng Gao

21 papers

Chandan Singh

7 papers

TL;DR

Abstract

Authors

References88 items

Ecosystem Graphs: The Social Footprint of Foundation Models

Sparks of Artificial General Intelligence: Early experiments with GPT-4

CHiLL: Zero-shot Custom Interpretable Feature Extraction from Clinical Notes with Large Language Models

One Embedder, Any Task: Instruction-Finetuned Text Embeddings

Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification

Predictive Coding or Just Feature Discovery? An Alternative Account of Why Language Models Fit Brain Data

A natural language fMRI dataset for voxelwise encoding models

Post-hoc Concept Bottleneck Models

Predictability and stability testing to assess clinical decision instrument performance for children after blunt torso trauma

Brains and algorithms partially converge in natural language processing

PromptSource: An Integrated Development Environment and Repository for Natural Language Prompts

Hierarchical Shrinkage: improving the accuracy and interpretability of tree-based methods

Yseop at FinSim-3 Shared Task 2021: Specializing Financial Domain Learning with Phrase Representations

Adaptive wavelet distillation from neural networks through interpretations

InferBERT: A Transformer-Based Causal Inference Framework for Enhancing Pharmacovigilance

Imodels: a Python Package for Fitting Interpretable Models

Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges

A statistical interpretation of term specificity and its application in retrieval

Learning from the Best: Rationalizing Prediction by Adversarial Information Calibration

Concept Bottleneck Models

The neural architecture of language: Integrative modeling converges on predictive processing

Generalized and Scalable Optimal Sparse Decision Trees

Language Models are Few-Shot Learners

Neural Additive Models: Interpretable Machine Learning with Neural Nets

TextAttack: A Framework for Adversarial Attacks, Data Augmentation, and Adversarial Training in NLP

Transformation Importance with Applications to Cosmology

Explaining Explanations: Axiomatic Feature Interactions for Deep Networks

Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer

HuggingFace's Transformers: State-of-the-art Natural Language Processing

DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

RoBERTa: A Robustly Optimized BERT Pretraining Approach

Disentangled Attribution Curves for Interpreting Random Forests and Boosted Trees

Explainable AI for Trees: From Local Explanations to Global Understanding

One neuron versus deep learning in aftershock prediction

Optimal Sparse Decision Trees

Definitions, methods, and applications in interpretable machine learning

Please Stop Explaining Black Box Models for High Stakes Decisions

Learning Global Additive Explanations for Neural Nets Using Model Distillation

This looks like that: deep learning for interpretable image recognition

Hierarchical interpretations for neural network predictions

The Building Blocks of Interpretability

Deep Contextualized Word Representations

Distilling a Neural Network Into a Soft Decision Tree

Deep Learning for Case-based Reasoning through Prototypes: A Neural Network that Explains its Predictions

Learning Certifiably Optimal Rule Lists

Optimal classification trees

An unexpected unity among methods for interpreting model predictions

Bag of Tricks for Efficient Text Classification

Deep learning for computational biology

European Union Regulations on Algorithmic Decision-Making and a "Right to Explanation"

XGBoost: A Scalable Tree Boosting System

Natural speech reveals the semantic maps that tile human cerebral cortex

“Why Should I Trust You?”: Explaining the Predictions of Any Classifier

Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-day Readmission

Distilling the Knowledge in a Neural Network

GloVe: Global Vectors for Word Representation

Distributed Representations of Words and Phrases and their Compositionality

Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank

The Emergence of Machine Learning Techniques in Criminology

Good debt or bad debt: Detecting semantic orientations in economic texts

Efficient Estimation of Word Representations in Vector Space

Classification and regression trees

Fairness through awareness

Scikit-learn: Machine Learning in Python

Modeling Annotators: A Generative Approach to Learning from Annotator Rationales

PREDICTIVE LEARNING VIA RULE ENSEMBLES

BART: Bayesian Additive Regression Trees

Generalized Additive Models

Seeing Stars: Exploiting Class Relationships for Sentiment Categorization with Respect to Rating Scales

Learning Question Classifiers

NLTK: The Natural Language Toolkit

Greedy function approximation: A gradient boosting machine.

Experiments with a New Boosting Algorithm

Generalized linear models. 2nd ed.

C4.5: Programs for Machine Learning

Route, Interpret, Repeat: Blurring the Line Between Post hoc Explainability and Interpretable Models