In recent years, predictive machine learning models have gained prominence across various scientific domains. However, their black-box nature necessitates establishing trust in them before accepting their predictions as accurate. One promising strategy involves employing explanation techniques that elucidate the rationale behind a model’s predictions in a way that humans can understand. However, assessing the degree of human interpretability of these explanations is a nontrivial challenge. In this work, we introduce interpretation entropy as a universal solution for evaluating the human interpretability of any linear model. Using this concept and drawing inspiration from classical thermodynamics, we present Thermodynamics-inspired Explainable Representations of AI and other black-box Paradigms, a method for generating optimally human-interpretable explanations in a model-agnostic manner. We demonstrate the wide-ranging applicability of this method by explaining predictions from various black-box model architectures across diverse domains, including molecular simulations, text, and image classification.
LEXPLAIN: Improving Model Explanations via Lexicon Supervision
Model explanations that shed light on the
model’s predictions are becoming a desired additional output of NLP models, alongside their
predictions. Challenges in creating these explanations include making them trustworthy
and faithful to the model’s predictions. In this
work, we propose a novel framework for guiding model explanations by supervising them explicitly. To this end, our method, LEXPLAIN,
uses task-related lexicons to directly supervise
model explanations. This approach consistently
improves the plausibility of model’s explanations without sacrificing performance on the
task, as we demonstrate on sentiment analysis
and toxicity detection. Our analyses show that
our method also demotes spurious correlations
(i.e., with respect to African American English
dialect) on toxicity detection, improving fairness.
more »
« less
- PAR ID:
- 10467897
- Publisher / Repository:
- Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023)
- Date Published:
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
more » « less
Abstract -
null (Ed.)We present an experimental assessment of the impact of feature attribution-style explanations on human performance in predicting the consensus toxicity of social media posts with advice from an unreliable machine learning model. By doing so we add to a small but growing body of literature inspecting the utility of interpretable machine learning in terms of human outcomes. We also evaluate interpretable machine learning for the first time in the important domain of online toxicity, where fully-automated methods have faced criticism as being inadequate as a measure of toxic behavior. We find that, contrary to expectations, explanations have no significant impact on accuracy or agreement with model predictions, through they do change the distribution of subject error somewhat while reducing the cognitive burden of the task for subjects. Our results contribute to the recognition of an intriguing expectation gap in the field of interpretable machine learning between the general excitement the field has engendered and the ambiguous results of recent experimental work, including this study.more » « less
-
Large Language Models (LLMs) can achieve strong performance on many tasks by producing step-by-step reasoning before giving a final output, often referred to as chain-of-thought reasoning (CoT). It is tempting to interpret these CoT explanations as the LLM’s process for solving a task. This level of transparency into LLMs’ predictions would yield significant safety benefits. However, we find that CoT explanations can systematically misrepresent the true reason for a model’s prediction. We demonstrate that CoT explanations can be heavily influenced by adding biasing features to model inputs—e.g., by reordering the multiple-choice options in a few-shot prompt to make the answer always “(A)”—which models systematically fail to mention in their explanations. When we bias models toward incorrect answers, they frequently generate CoT explanations rationalizing those answers. This causes accuracy to drop by as much as 36% on a suite of 13 tasks from BIG-Bench Hard, when testing with GPT-3.5 from OpenAI and Claude 1.0 from Anthropic. On a social-bias task, model explanations justify giving answers in line with stereotypes without mentioning the influence of these social biases. Our findings indicate that CoT explanations can be plausible yet misleading, which risks increasing our trust in LLMs without guaranteeing their safety. Building more transparent and explainable systems will require either improving CoT faithfulness through targeted efforts or abandoning CoT in favor of alternative methods.more » « less
-
Solar flares are transient space weather events that pose a significant threat to space and ground-based technological systems, making their precise and reliable prediction crucial for mitigating potential impacts. This paper contributes to the growing body of research on deep learning methods for solar flare prediction, primarily focusing on highly overlooked near-limb flares and utilizing the attribution methods to provide a post hoc qualitative explanation of the model’s predictions. We present a solar flare prediction model, which is trained using hourly full-disk line-of-sight magnetogram images and employs a binary prediction mode to forecast ≥M-class flares that may occur within the following 24-hour period. To address the class imbalance, we employ a fusion of data augmentation and class weighting techniques; and evaluate the overall performance of our model using the true skill statistic (TSS) and Heidke skill score (HSS). Moreover, we applied three attribution methods, namely Guided Gradient-weighted Class Activation Mapping, Integrated Gradients, and Deep Shapley Additive Explanations, to interpret and cross-validate our model’s predictions with the explanations. Our analysis revealed that full-disk prediction of solar flares aligns with characteristics related to active regions (ARs). In particular, the key findings of this study are: (1) our deep learning models achieved an average TSS∼0.51 and HSS∼0.35, and the results further demonstrate a competent capability to predict near-limb solar flares and (2) the qualitative analysis of the model’s explanation indicates that our model identifies and uses features associated with ARs in central and near-limb locations from full-disk magnetograms to make corresponding predictions. In other words, our models learn the shape and texture-based characteristics of flaring ARs even when they are at near-limb areas, which is a novel and critical capability that has significant implications for operational forecasting.more » « less
-
null (Ed.)Natural language inference (NLI) is the task of detecting the existence of entailment or contradiction in a given sentence pair. Although NLI techniques could help numerous information retrieval tasks, most solutions for NLI are neural approaches whose lack of interpretability prohibits both straightforward integration and diagnosis for further improvement. We target the task of generating token-level explanations for NLI from a neural model. Many existing approaches for token-level explanation are either computationally costly or require additional annotations for training. In this article, we first introduce a novel method for training an explanation generator that does not require additional human labels. Instead, the explanation generator is trained with the objective of predicting how the model’s classification output will change when parts of the inputs are modified. Second, we propose to build an explanation generator in a multi-task learning setting along with the original NLI task so the explanation generator can utilize the model’s internal behavior. The experiment results suggest that the proposed explanation generator outperforms numerous strong baselines. In addition, our method does not require excessive additional computation at prediction time, which renders it an order of magnitude faster than the best-performing baseline.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.18653/v1/2023.starsem-1.19
