Black box models only provide results for deep learning tasks, and lack informative details about how these results were obtained. Knowing how input variables are related to outputs, in addition to why they are related, can be critical to translating predictions into laboratory experiments, or defending a model prediction under scrutiny. In this paper, we propose a general theory that defines a variance tolerance factor (VTF) inspired by influence function, to interpret features in the context of black box neural networks by ranking the importance of features, and construct a novel architecture consisting of a base model and feature model to explore the feature importance in a Rashomon set that contains all well-performing neural networks. Two feature importance ranking methods in the Rashomon set and a feature selection method based on the VTF are created and explored. A thorough evaluation on synthetic and benchmark datasets is provided, and the method is applied to two real world examples predicting the formation of noncrystalline gold nanoparticles and the chemical toxicity 1793 aromatic compounds exposed to a protozoan ciliate for 40 hours.

The continuous development of Question Answering (QA) datasets has drawn the research community's attention toward multi-domain models. A popular approach is to use multi-dataset models, which are models trained on multiple datasets to learn their regularities and prevent overfitting to a single dataset. However, with the proliferation of QA models in online repositories such as GitHub or Hugging Face, an alternative is becoming viable. Recent works have demonstrated that combining expert agents can yield large performance gains over multi-dataset models. To ease research in multi-agent models, we extend UKP-SQuARE, an online platform for QA research, to support three families of multi-agent systems: i) agent selection, ii) early-fusion of agents, and iii) late-fusion of agents. We conduct experiments to evaluate their inference speed and discuss the performance vs. speed trade-off compared to multi-dataset models. UKP-SQuARE is open-source and publicly available at http://square.ukp-lab.de.

The prevalence of large-scale graphs poses great challenges in time and storage for training and deploying graph neural networks (GNNs). Several recent works have explored solutions for pruning the large original graph into a small and highly-informative one, such that training and inference on the pruned and large graphs have comparable performance. Although empirically effective, current researches focus on static or non-temporal graphs, which are not directly applicable to dynamic scenarios. In addition, they require labels as ground truth to learn the informative structure, limiting their applicability to new problem domains where labels are hard to obtain. To solve the dilemma, we propose and study the problem of unsupervised graph pruning on dynamic graphs. We approach the problem by our proposed STEP, a self-supervised temporal pruning framework that learns to remove potentially redundant edges from input dynamic graphs. From a technical and industrial viewpoint, our method overcomes the trade-offs between the performance and the time & memory overheads. Our results on three real-world datasets demonstrate the advantages on improving the efficacy, robustness, and efficiency of GNNs on dynamic node classification tasks. Most notably, STEP is able to prune more than 50% of edges on a million-scale industrial graph Alipay (7M nodes, 21M edges) while approximating up to 98% of the original performance. Code is available at https://github.com/EdisonLeeeee/STEP.

Interest in reinforcement learning (RL) has recently surged due to the application of deep learning techniques, but these connectionist approaches are opaque compared with symbolic systems. Learning Classifier Systems (LCSs) are evolutionary machine learning systems that can be categorised as eXplainable AI (XAI) due to their rule-based nature. Michigan LCSs are commonly used in RL domains as the alternative Pittsburgh systems (e.g. SAMUEL) suffer from complex algorithmic design and high computational requirements; however they can produce more compact/interpretable solutions than Michigan systems. We aim to develop two novel Pittsburgh LCSs to address RL domains: PPL-DL and PPL-ST. The former acts as a "zeroth-level" system, and the latter revisits SAMUEL's core Monte Carlo learning mechanism for estimating rule strength. We compare our two Pittsburgh systems to the Michigan system XCS across deterministic and stochastic FrozenLake environments. Results show that PPL-ST performs on-par or better than PPL-DL and outperforms XCS in the presence of high levels of environmental uncertainty. Rulesets evolved by PPL-ST can achieve higher performance than those evolved by XCS, but in a more parsimonious and therefore more interpretable fashion, albeit with higher computational cost. This indicates that PPL-ST is an LCS well-suited to producing explainable policies in RL domains.

Monitoring the fuel moisture content (FMC) of vegetation is crucial for managing and mitigating the impact of wildland fires. The combination of in situ FMC observations with numerical weather prediction (NWP) models and satellite retrievals has enabled the development of machine learning (ML) models to estimate dead FMC retrievals over the contiguous US (CONUS). In this study, ML models were trained using variables from the National Water Model and the High-Resolution Rapid Refresh (HRRR) NWP models, and static variables characterizing the surface properties, as well as surface reflectances and land surface temperature (LST) retrievals from the VIIRS instrument on board the Suomi-NPP satellite system. Extensive hyper-parameter optimization yielded skillful FMC models compared to a daily climatography RMSE (+44\%) and to an hourly climatography RMSE (+24\%). Furthermore, VIIRS retrievals were important predictors for estimating FMC, contributing significantly as a group due to their high band-correlation. In contrast, individual predictors in the HRRR group had relatively high importance according to the explainability techniques used. When both HRRR and VIIRS retrievals were not used as model inputs, the performance dropped significantly. If VIIRS retrievals were not used, the RMSE performance was worse. This highlights the importance of VIIRS retrievals in modeling FMC, which yielded better models compared to MODIS. Overall, the importance of the VIIRS group of predictors corroborates the dynamic relationship between the 10-h fuel and the atmosphere and soil moisture. These findings emphasize the significance of selecting appropriate data sources for predicting FMC with ML models, with VIIRS retrievals and selected HRRR variables being critical components in producing skillful FMC estimates.

Feature attribution methods (FAs) are popular approaches for providing insights into the model reasoning process of making predictions. The more faithful a FA is, the more accurately it reflects which parts of the input are more important for the prediction. Widely used faithfulness metrics, such as sufficiency and comprehensiveness use a hard erasure criterion, i.e. entirely removing or retaining the top most important tokens ranked by a given FA and observing the changes in predictive likelihood. However, this hard criterion ignores the importance of each individual token, treating them all equally for computing sufficiency and comprehensiveness. In this paper, we propose a simple yet effective soft erasure criterion. Instead of entirely removing or retaining tokens from the input, we randomly mask parts of the token vector representations proportionately to their FA importance. Extensive experiments across various natural language processing tasks and different FAs show that our soft-sufficiency and soft-comprehensiveness metrics consistently prefer more faithful explanations compared to hard sufficiency and comprehensiveness. Our code: https://github.com/casszhao/SoftFaith

Enhancing word usage is a desired feature for writing assistance. To further advance research in this area, this paper introduces "Smart Word Suggestions" (SWS) task and benchmark. Unlike other works, SWS emphasizes end-to-end evaluation and presents a more realistic writing assistance scenario. This task involves identifying words or phrases that require improvement and providing substitution suggestions. The benchmark includes human-labeled data for testing, a large distantly supervised dataset for training, and the framework for evaluation. The test data includes 1,000 sentences written by English learners, accompanied by over 16,000 substitution suggestions annotated by 10 native speakers. The training dataset comprises over 3.7 million sentences and 12.7 million suggestions generated through rules. Our experiments with seven baselines demonstrate that SWS is a challenging task. Based on experimental analysis, we suggest potential directions for future research on SWS. The dataset and related codes is available at https://github.com/microsoft/SmartWordSuggestions.

Large language models, e.g. ChatGPT are currently contributing enormously to make artificial intelligence even more popular, especially among the general population. However, such chatbot models were developed as tools to support natural language communication between humans. Problematically, it is very much a ``statistical correlation machine" (correlation instead of causality) and there are indeed ethical concerns associated with the use of AI language models such as ChatGPT, such as Bias, Privacy, and Abuse. This paper highlights specific ethical concerns on ChatGPT and articulates key challenges when ChatGPT is used in various applications. Practical commandments for different stakeholders of ChatGPT are also proposed that can serve as checklist guidelines for those applying ChatGPT in their applications. These commandment examples are expected to motivate the ethical use of ChatGPT.

Neural sequence models, especially transformers, exhibit a remarkable capacity for in-context learning. They can construct new predictors from sequences of labeled examples $(x, f(x))$ presented in the input without further parameter updates. We investigate the hypothesis that transformer-based in-context learners implement standard learning algorithms implicitly, by encoding smaller models in their activations, and updating these implicit models as new examples appear in the context. Using linear regression as a prototypical problem, we offer three sources of evidence for this hypothesis. First, we prove by construction that transformers can implement learning algorithms for linear models based on gradient descent and closed-form ridge regression. Second, we show that trained in-context learners closely match the predictors computed by gradient descent, ridge regression, and exact least-squares regression, transitioning between different predictors as transformer depth and dataset noise vary, and converging to Bayesian estimators for large widths and depths. Third, we present preliminary evidence that in-context learners share algorithmic features with these predictors: learners' late layers non-linearly encode weight vectors and moment matrices. These results suggest that in-context learning is understandable in algorithmic terms, and that (at least in the linear case) learners may rediscover standard estimation algorithms. Code and reference implementations are released at https://github.com/ekinakyurek/google-research/blob/master/incontext.

A recent trend in the domain of open-domain conversational agents is enabling them to converse empathetically to emotional prompts. Current approaches either follow an end-to-end approach or condition the responses on similar emotion labels to generate empathetic responses. But empathy is a broad concept that refers to the cognitive and emotional reactions of an individual to the observed experiences of another and it is more complex than mere mimicry of emotion. Hence, it requires identifying complex human conversational strategies and dynamics in addition to generic emotions to control and interpret empathetic responding capabilities of chatbots. In this work, we make use of a taxonomy of eight empathetic response intents in addition to generic emotion categories in building a dialogue response generation model capable of generating empathetic responses in a controllable and interpretable manner. It consists of two modules: 1) a response emotion/intent prediction module; and 2) a response generation module. We propose several rule-based and neural approaches to predict the next response's emotion/intent and generate responses conditioned on these predicted emotions/intents. Automatic and human evaluation results emphasize the importance of the use of the taxonomy of empathetic response intents in producing more diverse and empathetically more appropriate responses than end-to-end models.