Ideally, dialogue systems should generate responses that are faithful to the knowledge contained in relevant documents. However, many models generate hallucinated responses instead that contradict it or contain unverifiable information. To mitigate such undesirable behaviour, it has been proposed to fine-tune a `negative expert' on negative examples and subtract its parameters from those of a pre-trained model. However, intuitively, this does not take into account that some parameters are more responsible than others in causing hallucinations. Thus, we propose to weigh their individual importance via (an approximation of) the Fisher Information matrix, which measures the uncertainty of their estimate. We call this method Elastic Weight Removal (EWR). We evaluate our method -- using different variants of Flan-T5 as a backbone language model -- on multiple datasets for information-seeking dialogue generation and compare our method with state-of-the-art techniques for faithfulness, such as CTRL, Quark, DExperts, and Noisy Channel reranking. Extensive automatic and human evaluation shows that EWR systematically increases faithfulness at minor costs in terms of other metrics. However, we notice that only discouraging hallucinations may increase extractiveness, i.e. shallow copy-pasting of document spans, which can be undesirable. Hence, as a second main contribution, we show that our method can be extended to simultaneously discourage hallucinations and extractive responses. We publicly release the code for reproducing EWR and all baselines.

Continual learning (CL) aims to learn a sequence of tasks over time, with data distributions shifting from one task to another. When training on new task data, data representations from old tasks may drift. Some negative representation drift can result in catastrophic forgetting, by causing the locally learned class prototypes and data representations to correlate poorly across tasks. To mitigate such representation drift, we propose a method that finds global prototypes to guide the learning, and learns data representations with the regularization of the self-supervised information. Specifically, for NLP tasks, we formulate each task in a masked language modeling style, and learn the task via a neighbor attention mechanism over a pre-trained language model. Experimental results show that our proposed method can learn fairly consistent representations with less representation drift, and significantly reduce catastrophic forgetting in CL without resampling data from past tasks.

Offline reinforcement learning (RL) allows for the training of competent agents from offline datasets without any interaction with the environment. Online finetuning of such offline models can further improve performance. But how should we ideally finetune agents obtained from offline RL training? While offline RL algorithms can in principle be used for finetuning, in practice, their online performance improves slowly. In contrast, we show that it is possible to use standard online off-policy algorithms for faster improvement. However, we find this approach may suffer from policy collapse, where the policy undergoes severe performance deterioration during initial online learning. We investigate the issue of policy collapse and how it relates to data diversity, algorithm choices and online replay distribution. Based on these insights, we propose a conservative policy optimization procedure that can achieve stable and sample-efficient online learning from offline pretraining.

We investigated some difficulties that students often face when studying linear algebra at the undergraduate level, and identified some common mistakes and difficulties they often encountered when dealing with topics that require algorithmic thinking skills such as matrix factorization. In particular, we focused on (orthogonal) diagonalization and singular value decomposition (SVD). We also offered the possibility of exploring these topics using SageMath, a Python-based free open software computer algebra system (CAS) that has been identified to be useful for assisting many students in the computational process even though its output is static by nature. We then explored dynamic ChatGPT by inquiring the chatbot about the topic, either by asking to provide an example or to solve a problem, that is by constructing an (orthogonal) diagonalization or SVD from a particular matrix. By consolidating essential concepts in linear algebra and improving computational skills through effective practice, mastering these topics would become easier and mistakes could be minimized. Static SageMath, in particular, is a great aid for calculation confirmation and handling tedious computations. Although dynamic ChatGPT is relatively unreliable for solving problems in linear algebra, the mistakes it produces could become a valuable tool for improving critical thinking skills.

Language models (LMs) are increasingly being used in open-ended contexts, where the opinions reflected by LMs in response to subjective queries can have a profound impact, both on user satisfaction, as well as shaping the views of society at large. In this work, we put forth a quantitative framework to investigate the opinions reflected by LMs -- by leveraging high-quality public opinion polls and their associated human responses. Using this framework, we create OpinionsQA, a new dataset for evaluating the alignment of LM opinions with those of 60 US demographic groups over topics ranging from abortion to automation. Across topics, we find substantial misalignment between the views reflected by current LMs and those of US demographic groups: on par with the Democrat-Republican divide on climate change. Notably, this misalignment persists even after explicitly steering the LMs towards particular demographic groups. Our analysis not only confirms prior observations about the left-leaning tendencies of some human feedback-tuned LMs, but also surfaces groups whose opinions are poorly reflected by current LMs (e.g., 65+ and widowed individuals). Our code and data are available at https://github.com/tatsu-lab/opinions_qa.

Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel concepts from continuously shifting training data. Typical solutions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emergence of large-scale pre-trained vision transformer models has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well-established rehearsal-free continual learning setting. However, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sacrificing new task accuracy, and inability to benefit from expanded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input-conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the current SOTA method DualPrompt on established benchmarks by as much as 4.5% in average final accuracy. We also outperform the state of art by as much as 4.4% accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, corresponding to many practical settings. Our code is available at https://github.com/GT-RIPL/CODA-Prompt

This article presents a three-step framework for learning and solving partial differential equations (PDEs) using kernel methods. Given a training set consisting of pairs of noisy PDE solutions and source/boundary terms on a mesh, kernel smoothing is utilized to denoise the data and approximate derivatives of the solution. This information is then used in a kernel regression model to learn the algebraic form of the PDE. The learned PDE is then used within a kernel based solver to approximate the solution of the PDE with a new source/boundary term, thereby constituting an operator learning framework. Numerical experiments compare the method to state-of-the-art algorithms and demonstrate its competitive performance.

Nowadays, the ever-increasing urbanization and industrialization has led to a growing of water demand and a decrease in water supply and resources, thus creating a huge divergence between demand and supply. Therefore, water resources can play an important role in regional socio-economic and environmental development [Setegn, 2015]. The effective distribution of water resources in both civil and industry life indicates the levels of urban sustainability and social inclusiveness. Proper water distribution and forecasting can act as a baseline for achieving optimal resource allocation and mitigating the gap between supply and demand, thus improving operations, planning and management. In Greece, the recent years, the need for accurate water demand forecasting has become particularly important [Bithas and Chrysostomos, 2006]. The systematically extraction of non-renewable ground water, the insertion of chemicals for water purification, the drought caused by climate changes in the region of the Mediterranean and the sudden rise of water demand due to the increase of refugees and migrants has created many environmental issues on the quantity and quality of the water resources as well as previously unseen socio-economic and political problems. Therefore, an accurate forecasting of water consumption can be a decisive factor for proper planning, management and optimization. Water consumption data are seen as time series, since a measurement of water consumption levels is taken periodically (weekly, monthly, quarterly).

Time-dependent data-generating distributions have proven to be difficult for gradient-based training of neural networks, as the greedy updates result in catastrophic forgetting of previously learned knowledge. Despite the progress in the field of continual learning to overcome this forgetting, we show that a set of common state-of-the-art methods still suffers from substantial forgetting upon starting to learn new tasks, except that this forgetting is temporary and followed by a phase of performance recovery. We refer to this intriguing but potentially problematic phenomenon as the stability gap. The stability gap had likely remained under the radar due to standard practice in the field of evaluating continual learning models only after each task. Instead, we establish a framework for continual evaluation that uses per-iteration evaluation and we define a new set of metrics to quantify worst-case performance. Empirically we show that experience replay, constraintbased replay, knowledge-distillation, and parameter regularization methods are all prone to the stability gap; and that the stability gap can be observed in class-, task-, and domain-incremental learning benchmarks. Additionally, a controlled experiment shows that the stability gap increases when tasks are more dissimilar. Finally, by disentangling gradients into plasticity and stability components, we propose a conceptual explanation for the stability gap. The fast convergence in gradient-based optimization has resulted in many successes with highly overparameterized neural networks (Krizhevsky et al., 2012; Mnih et al., 2013; Devlin et al., 2018). In the standard training paradigm, these results are conditional on having a static data-generating distribution. However, when non-stationarity is introduced by a time-varying data-generating distribution, the gradient-based updates greedily overwrite the parameters of the previous solution. This results in catastrophic forgetting (French, 1999) and is one of the main hurdles in continual or lifelong learning.

In a recent article, Alon, Hanneke, Holzman, and Moran (FOCS '21) introduced a unifying framework to study the learnability of classes of partial concepts. One of the central questions studied in their work is whether the learnability of a partial concept class is always inherited from the learnability of some ``extension'' of it to a total concept class. They showed this is not the case for PAC learning but left the problem open for the stronger notion of online learnability. We resolve this problem by constructing a class of partial concepts that is online learnable, but no extension of it to a class of total concepts is online learnable (or even PAC learnable).