Boi Faltings

Martin Josifoski, Maxime Peyrard, Frano Rajic et al.

A critical component of a successful language generation pipeline is the decoding algorithm. However, the general principles that should guide the choice of a decoding algorithm remain unclear. Previous works only compare decoding algorithms in narrow scenarios, and their findings do not generalize across tasks. We argue that the misalignment between the model's likelihood and the task-specific notion of utility is the key factor to understanding the effectiveness of decoding algorithms. To structure the discussion, we introduce a taxonomy of misalignment mitigation strategies (MMSs), providing a unifying view of decoding as a tool for alignment. The MMS taxonomy groups decoding algorithms based on their implicit assumptions about likelihood--utility misalignment, yielding general statements about their applicability across tasks. Specifically, by analyzing the correlation between the likelihood and the utility of predictions across a diverse set of tasks, we provide empirical evidence supporting the proposed taxonomy and a set of principles to structure reasoning when choosing a decoding algorithm. Crucially, our analysis is the first to relate likelihood-based decoding algorithms with algorithms that rely on external information, such as value-guided methods and prompting, and covers the most diverse set of tasks to date. Code, data, and models are available at https://github.com/epfl-dlab/understanding-decoding.

Debjit Paul, Mete Ismayilzada, Maxime Peyrard et al.

Language models (LMs) have recently shown remarkable performance on reasoning tasks by explicitly generating intermediate inferences, e.g., chain-of-thought prompting. However, these intermediate inference steps may be inappropriate deductions from the initial context and lead to incorrect final predictions. Here we introduce REFINER, a framework for finetuning LMs to explicitly generate intermediate reasoning steps while interacting with a critic model that provides automated feedback on the reasoning. Specifically, the critic provides structured feedback that the reasoning LM uses to iteratively improve its intermediate arguments. Empirical evaluations of REFINER on three diverse reasoning tasks show significant improvements over baseline LMs of comparable scale. Furthermore, when using GPT-3.5 or ChatGPT as the reasoner, the trained critic significantly improves reasoning without finetuning the reasoner. Finally, our critic model is trained without expensive human-in-the-loop data but can be substituted with humans at inference time.

Beatriz Borges, Negar Foroutan, Deniz Bayazit et al.

AI assistants are being increasingly used by students enrolled in higher education institutions. While these tools provide opportunities for improved teaching and education, they also pose significant challenges for assessment and learning outcomes. We conceptualize these challenges through the lens of vulnerability, the potential for university assessments and learning outcomes to be impacted by student use of generative AI. We investigate the potential scale of this vulnerability by measuring the degree to which AI assistants can complete assessment questions in standard university-level STEM courses. Specifically, we compile a novel dataset of textual assessment questions from 50 courses at EPFL and evaluate whether two AI assistants, GPT-3.5 and GPT-4 can adequately answer these questions. We use eight prompting strategies to produce responses and find that GPT-4 answers an average of 65.8% of questions correctly, and can even produce the correct answer across at least one prompting strategy for 85.1% of questions. When grouping courses in our dataset by degree program, these systems already pass non-project assessments of large numbers of core courses in various degree programs, posing risks to higher education accreditation that will be amplified as these models improve. Our results call for revising program-level assessment design in higher education in light of advances in generative AI.

Luca Mouchel, Debjit Paul, Shaobo Cui et al.

Despite the remarkable performance of Large Language Models (LLMs) in natural language processing tasks, they still struggle with generating logically sound arguments, resulting in potential risks such as spreading misinformation. To address this issue, we introduce FIPO, a fallacy-informed framework that leverages preference optimization methods to steer LLMs toward logically sound arguments. FIPO includes a classification loss, to capture the fine-grained information on fallacy types. Our results on argumentation datasets show that our method reduces the fallacy errors by up to 17.5%. Furthermore, our human evaluation results indicate that the quality of the generated arguments by our method significantly outperforms the fine-tuned baselines, as well as other preference optimization methods, such as DPO. These findings highlight the importance of ensuring models are aware of logical fallacies for effective argument generation. Our code is available at github.com/lucamouchel/Logical-Fallacies.

Diego Antognini, Shuyang Li, Boi Faltings et al.

There has recently been growing interest in the automatic generation of cooking recipes that satisfy some form of dietary restrictions, thanks in part to the availability of online recipe data. Prior studies have used pre-trained language models, or relied on small paired recipe data (e.g., a recipe paired with a similar one that satisfies a dietary constraint). However, pre-trained language models generate inconsistent or incoherent recipes, and paired datasets are not available at scale. We address these deficiencies with RecipeCrit, a hierarchical denoising auto-encoder that edits recipes given ingredient-level critiques. The model is trained for recipe completion to learn semantic relationships within recipes. Our work's main innovation is our unsupervised critiquing module that allows users to edit recipes by interacting with the predicted ingredients; the system iteratively rewrites recipes to satisfy users' feedback. Experiments on the Recipe1M recipe dataset show that our model can more effectively edit recipes compared to strong language-modeling baselines, creating recipes that satisfy user constraints and are more correct, serendipitous, coherent, and relevant as measured by human judges.

Ljubomir Rokvic, Panayiotis Danassis, Sai Praneeth Karimireddy et al.

In Federated Learning, it is crucial to handle low-quality, corrupted, or malicious data. However, traditional data valuation methods are not suitable due to privacy concerns. To address this, we propose a simple yet effective approach that utilizes a new influence approximation called "lazy influence" to filter and score data while preserving privacy. To do this, each participant uses their own data to estimate the influence of another participant's batch and sends a differentially private obfuscated score to the central coordinator. Our method has been shown to successfully filter out biased and corrupted data in various simulated and real-world settings, achieving a recall rate of over $>90\%$ (sometimes up to $100\%$) while maintaining strong differential privacy guarantees with $\varepsilon \leq 1$.

Diego Antognini, Boi Faltings

Providing explanations for recommended items allows users to refine the recommendations by critiquing parts of the explanations. As a result of revisiting critiquing from the perspective of multimodal generative models, recent work has proposed M&Ms-VAE, which achieves state-of-the-art performance in terms of recommendation, explanation, and critiquing. M&Ms-VAE and similar models allow users to negatively critique (i.e., explicitly disagree). However, they share a significant drawback: users cannot positively critique (i.e., highlight a desired feature). We address this deficiency with M&Ms-VAE+, an extension of M&Ms-VAE that enables positive and negative critiquing. In addition to modeling users' interactions and keyphrase-usage preferences, we model their keyphrase-usage dislikes. Moreover, we design a novel critiquing module that is trained in a self-supervised fashion. Our experiments on two datasets show that M&Ms-VAE+ matches or exceeds M&Ms-VAE in recommendation and explanation performance. Furthermore, our results demonstrate that representing positive and negative critiques differently enables M&Ms-VAE+ to significantly outperform M&Ms-VAE and other models in positive and negative multi-step critiquing.

Shuangqi Li, Diego Antognini, Boi Faltings

Explanation is important for text classification tasks. One prevalent type of explanation is rationales, which are text snippets of input text that suffice to yield the prediction and are meaningful to humans. A lot of research on rationalization has been based on the selective rationalization framework, which has recently been shown to be problematic due to the interlocking dynamics. In this paper, we show that we address the interlocking problem in the multi-aspect setting, where we aim to generate multiple rationales for multiple outputs. More specifically, we propose a multi-stage training method incorporating an additional self-supervised contrastive loss that helps to generate more semantically diverse rationales. Empirical results on the beer review dataset show that our method improves significantly the rationalization performance.

Zhuan Shi, Jing Yan, Xiaoli Tang et al.

The increasing sophistication of text-to-image generative models has led to complex challenges in defining and enforcing copyright infringement criteria and protection. Existing methods, such as watermarking and dataset deduplication, fail to provide comprehensive solutions due to the lack of standardized metrics and the inherent complexity of addressing copyright infringement in diffusion models. To deal with these challenges, we propose a Reinforcement Learning-based Copyright Protection(RLCP) method for Text-to-Image Diffusion Model, which minimizes the generation of copyright-infringing content while maintaining the quality of the model-generated dataset. Our approach begins with the introduction of a novel copyright metric grounded in copyright law and court precedents on infringement. We then utilize the Denoising Diffusion Policy Optimization (DDPO) framework to guide the model through a multi-step decision-making process, optimizing it using a reward function that incorporates our proposed copyright metric. Additionally, we employ KL divergence as a regularization term to mitigate some failure modes and stabilize RL fine-tuning. Experiments conducted on 3 mixed datasets of copyright and non-copyright images demonstrate that our approach significantly reduces copyright infringement risk while maintaining image quality.

Shaobo Cui, Junyou Li, Luca Mouchel et al.

To address this gap, our study introduces the concept of causal epistemic consistency, which focuses on the self-consistency of Large Language Models (LLMs) in differentiating intermediates with nuanced differences in causal reasoning. We propose a suite of novel metrics -- intensity ranking concordance, cross-group position agreement, and intra-group clustering -- to evaluate LLMs on this front. Through extensive empirical studies on 21 high-profile LLMs, including GPT-4, Claude3, and LLaMA3-70B, we have favoring evidence that current models struggle to maintain epistemic consistency in identifying the polarity and intensity of intermediates in causal reasoning. Additionally, we explore the potential of using internal token probabilities as an auxiliary tool to maintain causal epistemic consistency. In summary, our study bridges a critical gap in AI research by investigating the self-consistency over fine-grained intermediates involved in causal reasoning.

Diego Antognini, Boi Faltings

Today, recommender systems are an inevitable part of everyone's daily digital routine and are present on most internet platforms. State-of-the-art deep learning-based models require a large number of data to achieve their best performance. Many datasets fulfilling this criterion have been proposed for multiple domains, such as Amazon products, restaurants, or beers. However, works and datasets in the hotel domain are limited: the largest hotel review dataset is below the million samples. Additionally, the hotel domain suffers from a higher data sparsity than traditional recommendation datasets and therefore, traditional collaborative-filtering approaches cannot be applied to such data. In this paper, we propose HotelRec, a very large-scale hotel recommendation dataset, based on TripAdvisor, containing 50 million reviews. To the best of our knowledge, HotelRec is the largest publicly available dataset in the hotel domain (50M versus 0.9M) and additionally, the largest recommendation dataset in a single domain and with textual reviews (50M versus 22M). We release HotelRec for further research: https://github.com/Diego999/HotelRec.

Diego Antognini, Boi Faltings

Today's research progress in the field of multi-document summarization is obstructed by the small number of available datasets. Since the acquisition of reference summaries is costly, existing datasets contain only hundreds of samples at most, resulting in heavy reliance on hand-crafted features or necessitating additional, manually annotated data. The lack of large corpora therefore hinders the development of sophisticated models. Additionally, most publicly available multi-document summarization corpora are in the news domain, and no analogous dataset exists in the video game domain. In this paper, we propose GameWikiSum, a new domain-specific dataset for multi-document summarization, which is one hundred times larger than commonly used datasets, and in another domain than news. Input documents consist of long professional video game reviews as well as references of their gameplay sections in Wikipedia pages. We analyze the proposed dataset and show that both abstractive and extractive models can be trained on it. We release GameWikiSum for further research: https://github.com/Diego999/GameWikiSum.

Debjit Paul, Robert West, Antoine Bosselut et al.

Large language models (LLMs) have been shown to perform better when asked to reason step-by-step before answering a question. However, it is unclear to what degree the model's final answer is faithful to the stated reasoning steps. In this paper, we perform a causal mediation analysis on twelve LLMs to examine how intermediate reasoning steps generated by the LLM influence the final outcome and find that LLMs do not reliably use their intermediate reasoning steps when generating an answer. To address this issue, we introduce FRODO, a framework to tailor small-sized LMs to generate correct reasoning steps and robustly reason over these steps. FRODO consists of an inference module that learns to generate correct reasoning steps using an implicit causal reward function and a reasoning module that learns to faithfully reason over these intermediate inferences using a counterfactual and causal preference objective. Our experiments show that FRODO significantly outperforms four competitive baselines. Furthermore, FRODO improves the robustness and generalization ability of the reasoning LM, yielding higher performance on out-of-distribution test sets. Finally, we find that FRODO's rationales are more faithful to its final answer predictions than standard supervised fine-tuning.

Zeki Doruk Erden, Boi Faltings

Adaptive networks today rely on overparameterized fixed topologies that cannot break through the statistical conflicts they encounter in the data they are exposed to, and are prone to "catastrophic forgetting" as the network attempts to reuse the existing structures to learn new task. We propose a structural adaptation method, DIRAD, that can complexify as needed and in a directed manner without being limited by statistical conflicts within a dataset. We then extend this method and present the PREVAL framework, designed to prevent "catastrophic forgetting" in continual learning by detection of new data and assigning encountered data to suitable models adapted to process them, without needing task labels anywhere in the workflow. We show the reliability of the DIRAD in growing a network with high performance and orders-of-magnitude simpler than fixed topology networks; and demonstrate the proof-of-concept operation of PREVAL, in which continual adaptation to new tasks is observed while being able to detect and discern previously-encountered tasks.

Zeki Doruk Erden, Donia Gasmi, Boi Faltings

Continual learning for reinforcement learning agents remains a significant challenge, particularly in preserving and leveraging existing information without an external signal to indicate changes in tasks or environments. In this study, we explore the effectiveness of autoencoders in detecting new tasks and matching observed environments to previously encountered ones. Our approach integrates policy optimization with familiarity autoencoders within an end-to-end continual learning system. This system can recognize and learn new tasks or environments while preserving knowledge from earlier experiences and can selectively retrieve relevant knowledge when re-encountering a known environment. Initial results demonstrate successful continual learning without external signals to indicate task changes or reencounters, showing promise for this methodology.

Shaobo Cui, Lazar Milikic, Yiyang Feng et al.

Defeasibility in causal reasoning implies that the causal relationship between cause and effect can be strengthened or weakened. Namely, the causal strength between cause and effect should increase or decrease with the incorporation of strengthening arguments (supporters) or weakening arguments (defeaters), respectively. However, existing works ignore defeasibility in causal reasoning and fail to evaluate existing causal strength metrics in defeasible settings. In this work, we present $δ$-CAUSAL, the first benchmark dataset for studying defeasibility in causal reasoning. $δ$-CAUSAL includes around 11K events spanning ten domains, featuring defeasible causality pairs, i.e., cause-effect pairs accompanied by supporters and defeaters. We further show current causal strength metrics fail to reflect the change of causal strength with the incorporation of supporters or defeaters in $δ$-CAUSAL. To this end, we propose CESAR (Causal Embedding aSsociation with Attention Rating), a metric that measures causal strength based on token-level causal relationships. CESAR achieves a significant 69.7% relative improvement over existing metrics, increasing from 47.2% to 80.1% in capturing the causal strength change brought by supporters and defeaters. We further demonstrate even Large Language Models (LLMs) like GPT-3.5 still lag 4.5 and 10.7 points behind humans in generating supporters and defeaters, emphasizing the challenge posed by $δ$-CAUSAL.

Zeki Doruk Erden, Boi Faltings

Contemporary machine learning paradigm excels in statistical data analysis, solving problems that classical AI couldn't. However, it faces key limitations, such as a lack of integration with planning, incomprehensible internal structure, and inability to learn continually. We present the initial design for an AI system, Agential AI (AAI), in principle operating independently or on top of statistical methods, designed to overcome these issues. AAI's core is a learning method that models temporal dynamics with guarantees of completeness, minimality, and continual learning, using component-level variation and selection to learn the structure of the environment. It integrates this with a behavior algorithm that plans on a learned model and encapsulates high-level behavior patterns. Preliminary experiments on a simple environment show AAI's effectiveness and potential.

Shunchang Liu, Zhuan Shi, Lingjuan Lyu et al.

Assessing whether AI-generated images are substantially similar to source works is a crucial step in resolving copyright disputes. In this paper, we propose CopyJudge, a novel automated infringement identification framework that leverages large vision-language models (LVLMs) to simulate practical court processes for determining substantial similarity between copyrighted images and those generated by text-to-image diffusion models. Specifically, we employ an abstraction-filtration-comparison test framework based on the multi-LVLM debate to assess the likelihood of infringement and provide detailed judgment rationales. Based on these judgments, we further introduce a general LVLM-based mitigation strategy that automatically optimizes infringing prompts by avoiding sensitive expressions while preserving the non-infringing content. Furthermore, assuming the input noise is controllable, our approach can be enhanced by iteratively exploring non-infringing noise vectors within the diffusion latent space, even without modifying the original prompts. Experimental results show that our automated identification method achieves comparable state-of-the-art performance, while offering superior generalization and interpretability across various forms of infringement, and that our mitigation method more effectively mitigates memorization and IP infringement with a high degree of alignment to the original non-infringing expressions.

Zeki Doruk Erden, Boi Faltings

Inherent limitations of contemporary machine learning systems in crucial areas -- importantly in continual learning, information reuse, comprehensibility, and integration with deliberate behavior -- are receiving increasing attention. To address these challenges, we introduce a system design, fueled by a novel learning approach conceptually grounded in principles of evolutionary developmental biology, that overcomes key limitations of current methods. Our design comprises three core components: The Modeller, a gradient-free learning mechanism inherently capable of continual learning and structural adaptation; a planner for goal-directed action over learned models; and a behavior encapsulation mechanism that can decompose complex behaviors into a hierarchical structure. We demonstrate proof-of-principle operation in a simple test environment. Additionally, we extend our modeling framework to higher-dimensional network-structured spaces, using MNIST for a shape detection task. Our framework shows promise in overcoming multiple major limitations of contemporary machine learning systems simultaneously and in an organic manner.

Zeki Doruk Erden, Boi Faltings

This article critically examines the foundational principles of contemporary AI methods, exploring the limitations that hinder its potential. We draw parallels between the modern AI landscape and the 20th-century Modern Synthesis in evolutionary biology, and highlight how advancements in evolutionary theory that augmented the Modern Synthesis, particularly those of Evolutionary Developmental Biology, offer insights that can inform a new design paradigm for AI. By synthesizing findings across AI and evolutionary theory, we propose a pathway to overcome existing limitations, enabling AI to achieve its aspirational goals.

Zeki Doruk Erden, Boi Faltings

Artificial intelligence (AI), propelled by advancements in machine learning, has made significant strides in solving complex tasks. However, the current neural network-based paradigm, while effective, is heavily constrained by inherent limitations, primarily a lack of structural organization and a progression of learning that displays undesirable properties. As AI research progresses without a unifying framework, it either tries to patch weaknesses heuristically or draws loosely from biological mechanisms without strong theoretical foundations. Meanwhile, the recent paradigm shift in evolutionary understanding -- driven primarily by evolutionary developmental biology (EDB) -- has been largely overlooked in AI literature, despite a striking analogy between the Modern Synthesis and contemporary machine learning, evident in their shared assumptions, approaches, and limitations upon careful analysis. Consequently, the principles of adaptation from EDB that reshaped our understanding of the evolutionary process can also form the foundation of a unifying conceptual framework for the next design philosophy in AI, going beyond mere inspiration and grounded firmly in biology's first principles. This article provides a detailed overview of the analogy between the Modern Synthesis and modern machine learning, and outlines the core principles of a new AI design paradigm based on insights from EDB. To exemplify our analysis, we also present two learning system designs grounded in specific developmental principles -- regulatory connections, somatic variation and selection, and weak linkage -- that resolve multiple major limitations of contemporary machine learning in an organic manner, while also providing deeper insights into the role of these mechanisms in biological evolution.

Yiyang Feng, Yichen Wang, Shaobo Cui et al.

Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning, positioning them as promising tools for supporting human problem-solving. However, what happens when their performance is affected by misinformation, i.e., incorrect inputs introduced by users due to oversights or gaps in knowledge? Such misinformation is prevalent in real-world interactions with LLMs, yet how it propagates within LLMs' reasoning process remains underexplored. Focusing on mathematical reasoning, we present a comprehensive analysis of how misinformation affects intermediate reasoning steps and final answers. We also examine how effectively LLMs can correct misinformation when explicitly instructed to do so. Even with explicit instructions, LLMs succeed less than half the time in rectifying misinformation, despite possessing correct internal knowledge, leading to significant accuracy drops (10.02% - 72.20%), and the degradation holds with thinking models (4.30% - 19.97%). Further analysis shows that applying factual corrections early in the reasoning process most effectively reduces misinformation propagation, and fine-tuning on synthesized data with early-stage corrections significantly improves reasoning factuality. Our work offers a practical approach to mitigating misinformation propagation.

Ljubomir Rokvic, Panayiotis Danassis, Boi Faltings

In Federated Learning, heterogeneity in client data distributions often means that a single global model does not have the best performance for individual clients. Consider for example training a next-word prediction model for keyboards: user-specific language patterns due to demographics (dialect, age, etc.), language proficiency, and writing style result in a highly non-IID dataset across clients. Other examples are medical images taken with different machines, or driving data from different vehicle types. To address this, we propose a simple yet effective personalized federated learning framework (pFedLIA) that utilizes a computationally efficient influence approximation, called `Lazy Influence', to cluster clients in a distributed manner before model aggregation. Within each cluster, data owners collaborate to jointly train a model that captures the specific data patterns of the clients. Our method has been shown to successfully recover the global model's performance drop due to the non-IID-ness in various synthetic and real-world settings, specifically a next-word prediction task on the Nordic languages as well as several benchmark tasks. It matches the performance of a hypothetical Oracle clustering, and significantly improves on existing baselines, e.g., an improvement of 17% on CIFAR100.

Zhuan Shi, Yifei Song, Xiaoli Tang et al.

Generative art using Diffusion models has achieved remarkable performance in image generation and text-to-image tasks. However, the increasing demand for training data in generative art raises significant concerns about copyright infringement, as models can produce images highly similar to copyrighted works. Existing solutions attempt to mitigate this by perturbing Diffusion models to reduce the likelihood of generating such images, but this often compromises model performance. Another approach focuses on economically compensating data holders for their contributions, yet it fails to address copyright loss adequately. Our approach begin with the introduction of a novel copyright metric grounded in copyright law and court precedents on infringement. We then employ the TRAK method to estimate the contribution of data holders. To accommodate the continuous data collection process, we divide the training into multiple rounds. Finally, We designed a hierarchical budget allocation method based on reinforcement learning to determine the budget for each round and the remuneration of the data holder based on the data holder's contribution and copyright loss in each round. Extensive experiments across three datasets show that our method outperforms all eight benchmarks, demonstrating its effectiveness in optimizing budget distribution in a copyright-aware manner. To the best of our knowledge, this is the first technical work that introduces to incentive contributors and protect their copyrights by compensating them.

Sankarshan Damle, Boi Faltings

Large Language Models (LLMs) are increasingly expected to handle complex decision-making tasks, yet their ability to perform structured resource allocation remains underexplored. Evaluating their reasoning is also difficult due to data contamination and the static nature of existing benchmarks. We present a dual-purpose framework leveraging Participatory Budgeting (PB) both as (i) a practical setting for LLM-based resource allocation and (ii) an adaptive benchmark for evaluating their reasoning capabilities. We task LLMs with selecting project subsets under feasibility (e.g., budget) constraints via three prompting strategies: greedy selection, direct optimization, and a hill-climbing-inspired refinement. We benchmark LLMs' allocations against a utility-maximizing oracle. Interestingly, we also test whether LLMs can infer structured preferences from natural-language voter input or metadata, without explicit votes. By comparing allocations based on inferred preferences to those from ground-truth votes, we evaluate LLMs' ability to extract preferences from open-ended input. Our results underscore the role of prompt design and show that LLMs hold promise for mechanism design with unstructured inputs.

Zeki Doruk Erden, Boi Faltings

We analyze the ability of computational units to retain past responses after parameter updates, a key property for system-wide continual learning. Neural networks trained with gradient descent lack this capability, prompting us to propose Modelleyen, an alternative approach with inherent response preservation. We demonstrate through experiments on modeling the dynamics of a simple environment and on MNIST that, despite increased computational complexity and some representational limitations at its current stage, Modelleyen achieves continual learning without relying on sample replay or predefined task boundaries.

Shaobo Cui, Zhijing Jin, Bernhard Schölkopf et al.

Understanding commonsense causality is a unique mark of intelligence for humans. It helps people understand the principles of the real world better and benefits the decision-making process related to causation. For instance, commonsense causality is crucial in judging whether a defendant's action causes the plaintiff's loss in determining legal liability. Despite its significance, a systematic exploration of this topic is notably lacking. Our comprehensive survey bridges this gap by focusing on taxonomies, benchmarks, acquisition methods, qualitative reasoning, and quantitative measurements in commonsense causality, synthesizing insights from over 200 representative articles. Our work aims to provide a systematic overview, update scholars on recent advancements, provide a pragmatic guide for beginners, and highlight promising future research directions in this vital field.

Fei Mi, Tao Lin, Boi Faltings

The ability to quickly learn new knowledge (e.g. new classes or data distributions) is a big step towards human-level intelligence. In this paper, we consider scenarios that require learning new classes or data distributions quickly and incrementally over time, as it often occurs in real-world dynamic environments. We propose "Memory-based Hebbian Parameter Adaptation" (Hebb) to tackle the two major challenges (i.e., catastrophic forgetting and sample efficiency) towards this goal in a unified framework. To mitigate catastrophic forgetting, Hebb augments a regular neural classifier with a continuously updated memory module to store representations of previous data. To improve sample efficiency, we propose a parameter adaptation method based on the well-known Hebbian theory, which directly "wires" the output network's parameters with similar representations retrieved from the memory. We empirically verify the superior performance of Hebb through extensive experiments on a wide range of learning tasks (image classification, language model) and learning scenarios (continual, incremental, online). We demonstrate that Hebb effectively mitigates catastrophic forgetting, and it indeed learns new knowledge better and faster than the current state-of-the-art.

Fei Mi, Wanhao Zhou, Fengyu Cai et al.

As the labeling cost for different modules in task-oriented dialog (ToD) systems is expensive, a major challenge is to train different modules with the least amount of labeled data. Recently, large-scale pre-trained language models, have shown promising results for few-shot learning in ToD. In this paper, we devise a self-training approach to utilize the abundant unlabeled dialog data to further improve state-of-the-art pre-trained models in few-shot learning scenarios for ToD systems. Specifically, we propose a self-training approach that iteratively labels the most confident unlabeled data to train a stronger Student model. Moreover, a new text augmentation technique (GradAug) is proposed to better train the Student by replacing non-crucial tokens using a masked language model. We conduct extensive experiments and present analyses on four downstream tasks in ToD, including intent classification, dialog state tracking, dialog act prediction, and response selection. Empirical results demonstrate that the proposed self-training approach consistently improves state-of-the-art pre-trained models (BERT, ToD-BERT) when only a small number of labeled data are available.

Fengyu Cai, Wanhao Zhou, Fei Mi et al.

Utterance-level intent detection and token-level slot filling are two key tasks for natural language understanding (NLU) in task-oriented systems. Most existing approaches assume that only a single intent exists in an utterance. However, there are often multiple intents within an utterance in real-life scenarios. In this paper, we propose a multi-intent NLU framework, called SLIM, to jointly learn multi-intent detection and slot filling based on BERT. To fully exploit the existing annotation data and capture the interactions between slots and intents, SLIM introduces an explicit slot-intent classifier to learn the many-to-one mapping between slots and intents. Empirical results on three public multi-intent datasets demonstrate (1) the superior performance of SLIM compared to the current state-of-the-art for NLU with multiple intents and (2) the benefits obtained from the slot-intent classifier.

Diana Petrescu, Diego Antognini, Boi Faltings

Recommendations with personalized explanations have been shown to increase user trust and perceived quality and help users make better decisions. Moreover, such explanations allow users to provide feedback by critiquing them. Several algorithms for recommender systems with multi-step critiquing have therefore been developed. However, providing a user-friendly interface based on personalized explanations and critiquing has not been addressed in the last decade. In this paper, we introduce four different web interfaces (available under https://lia.epfl.ch/critiquing/) helping users making decisions and finding their ideal item. We have chosen the hotel recommendation domain as a use case even though our approach is trivially adaptable for other domains. Moreover, our system is model-agnostic (for both recommender systems and critiquing models) allowing a great flexibility and further extensions. Our interfaces are above all a useful tool to help research in recommendation with critiquing. They allow to test such systems on a real use case and also to highlight some limitations of these approaches to find solutions to overcome them.

Panayiotis Danassis, Aris Filos-Ratsikas, Haipeng Chen et al.

Traditional competitive markets do not account for negative externalities; indirect costs that some participants impose on others, such as the cost of over-appropriating a common-pool resource (which diminishes future stock, and thus harvest, for everyone). Quantifying appropriate interventions to market prices has proven to be quite challenging. We propose a practical approach to computing market prices and allocations via a deep reinforcement learning policymaker agent, operating in an environment of other learning agents. Our policymaker allows us to tune the prices with regard to diverse objectives such as sustainability and resource wastefulness, fairness, buyers' and sellers' welfare, etc. As a highlight of our findings, our policymaker is significantly more successful in maintaining resource sustainability, compared to the market equilibrium outcome, in scarce resource environments.

Diego Antognini, Boi Faltings

Automated predictions require explanations to be interpretable by humans. One type of explanation is a rationale, i.e., a selection of input features such as relevant text snippets from which the model computes the outcome. However, a single overall selection does not provide a complete explanation, e.g., weighing several aspects for decisions. To this end, we present a novel self-interpretable model called ConRAT. Inspired by how human explanations for high-level decisions are often based on key concepts, ConRAT extracts a set of text snippets as concepts and infers which ones are described in the document. Then, it explains the outcome with a linear aggregation of concepts. Two regularizers drive ConRAT to build interpretable concepts. In addition, we propose two techniques to boost the rationale and predictive performance further. Experiments on both single- and multi-aspect sentiment classification tasks show that ConRAT is the first to generate concepts that align with human rationalization while using only the overall label. Further, it outperforms state-of-the-art methods trained on each aspect label independently.

Panayiotis Danassis, Florian Wiedemair, Boi Faltings

We present a multi-agent learning algorithm, ALMA-Learning, for efficient and fair allocations in large-scale systems. We circumvent the traditional pitfalls of multi-agent learning (e.g., the moving target problem, the curse of dimensionality, or the need for mutually consistent actions) by relying on the ALMA heuristic as a coordination mechanism for each stage game. ALMA-Learning is decentralized, observes only own action/reward pairs, requires no inter-agent communication, and achieves near-optimal (<5% loss) and fair coordination in a variety of synthetic scenarios and a real-world meeting scheduling problem. The lightweight nature and fast learning constitute ALMA-Learning ideal for on-device deployment.

Diego Antognini, Boi Faltings

Recent studies have shown that providing personalized explanations alongside recommendations increases trust and perceived quality. Furthermore, it gives users an opportunity to refine the recommendations by critiquing parts of the explanations. On one hand, current recommender systems model the recommendation, explanation, and critiquing objectives jointly, but this creates an inherent trade-off between their respective performance. On the other hand, although recent latent linear critiquing approaches are built upon an existing recommender system, they suffer from computational inefficiency at inference due to the objective optimized at each conversation's turn. We address these deficiencies with M&Ms-VAE, a novel variational autoencoder for recommendation and explanation that is based on multimodal modeling assumptions. We train the model under a weak supervision scheme to simulate both fully and partially observed variables. Then, we leverage the generalization ability of a trained M&Ms-VAE model to embed the user preference and the critique separately. Our work's most important innovation is our critiquing module, which is built upon and trained in a self-supervised manner with a simple ranking objective. Experiments on four real-world datasets demonstrate that among state-of-the-art models, our system is the first to dominate or match the performance in terms of recommendation, explanation, and multi-step critiquing. Moreover, M&Ms-VAE processes the critiques up to 25.6x faster than the best baselines. Finally, we show that our model infers coherent joint and cross generation, even under weak supervision, thanks to our multimodal-based modeling and training scheme.

Panayiotis Danassis, Zeki Doruk Erden, Boi Faltings

Can artificial agents benefit from human conventions? Human societies manage to successfully self-organize and resolve the tragedy of the commons in common-pool resources, in spite of the bleak prediction of non-cooperative game theory. On top of that, real-world problems are inherently large-scale and of low observability. One key concept that facilitates human coordination in such settings is the use of conventions. Inspired by human behavior, we investigate the learning dynamics and emergence of temporal conventions, focusing on common-pool resources. Extra emphasis was given in designing a realistic evaluation setting: (a) environment dynamics are modeled on real-world fisheries, (b) we assume decentralized learning, where agents can observe only their own history, and (c) we run large-scale simulations (up to 64 agents). Uncoupled policies and low observability make cooperation hard to achieve; as the number of agents grow, the probability of taking a correct gradient direction decreases exponentially. By introducing an arbitrary common signal (e.g., date, time, or any periodic set of numbers) as a means to couple the learning process, we show that temporal conventions can emerge and agents reach sustainable harvesting strategies. The introduction of the signal consistently improves the social welfare (by 258% on average, up to 3306%), the range of environmental parameters where sustainability can be achieved (by 46% on average, up to 300%), and the convergence speed in low abundance settings (by 13% on average, up to 53%).

Panayiotis Danassis, Aleksei Triastcyn, Boi Faltings

We introduce a practical and scalable algorithm (PALMA) for solving one of the fundamental problems of multi-agent systems -- finding matches and allocations -- in unboundedly large settings (e.g., resource allocation in urban environments, mobility-on-demand systems, etc.), while providing strong worst-case privacy guarantees. PALMA is decentralized, runs on-device, requires no inter-agent communication, and converges in constant time under reasonable assumptions. We evaluate PALMA in a mobility-on-demand and a paper assignment scenario, using real data in both, and demonstrate that it provides a strong level of privacy ($\varepsilon \leq 1$ and median as low as $\varepsilon = 0.5$ across agents) and high-quality matchings (up to $86\%$ of the non-private optimal, outperforming even the privacy-preserving centralized maximum-weight matching baseline).

Fei Mi, Liangwei Chen, Mengjie Zhao et al.

Natural language generation (NLG) is an essential component of task-oriented dialog systems. Despite the recent success of neural approaches for NLG, they are typically developed in an offline manner for particular domains. To better fit real-life applications where new data come in a stream, we study NLG in a "continual learning" setting to expand its knowledge to new domains or functionalities incrementally. The major challenge towards this goal is catastrophic forgetting, meaning that a continually trained model tends to forget the knowledge it has learned before. To this end, we propose a method called ARPER (Adaptively Regularized Prioritized Exemplar Replay) by replaying prioritized historical exemplars, together with an adaptive regularization technique based on ElasticWeight Consolidation. Extensive experiments to continually learn new domains and intents are conducted on MultiWoZ-2.0 to benchmark ARPER with a wide range of techniques. Empirical results demonstrate that ARPER significantly outperforms other methods by effectively mitigating the detrimental catastrophic forgetting issue.

Milena Filipovic, Blagoj Mitrevski, Diego Antognini et al.

Recommender systems research tends to evaluate model performance offline and on randomly sampled targets, yet the same systems are later used to predict user behavior sequentially from a fixed point in time. Simulating online recommender system performance is notoriously difficult and the discrepancy between online and offline behaviors is typically not accounted for in offline evaluations. This disparity permits weaknesses to go unnoticed until the model is deployed in a production setting. In this paper, we first demonstrate how omitting temporal context when evaluating recommender system performance leads to false confidence. To overcome this, we postulate that offline evaluation protocols can only model real-life use-cases if they account for temporal context. Next, we propose a training procedure to further embed the temporal context in existing models. We use a multi-objective approach to introduce temporal context into traditionally time-unaware recommender systems and confirm its advantage via the proposed evaluation protocol. Finally, we validate that the Pareto Fronts obtained with the added objective dominate those produced by state-of-the-art models that are only optimized for accuracy on three real-world publicly available datasets. The results show that including our temporal objective can improve recall@20 by up to 20%.

Blagoj Mitrevski, Milena Filipovic, Diego Antognini et al.

Multi-objective gradient methods are becoming the standard for solving multi-objective problems. Among others, they show promising results in developing multi-objective recommender systems with both correlated and conflicting objectives. Classic multi-gradient~descent usually relies on the combination of the gradients, not including the computation of first and second moments of the gradients. This leads to a brittle behavior and misses important areas in the solution space. In this work, we create a multi-objective model-agnostic Adamize method that leverages the benefits of the Adam optimizer in single-objective problems. This corrects and stabilizes~the~gradients of every objective before calculating a common gradient descent vector that optimizes all the objectives simultaneously. We evaluate the benefits of Multi-objective Adamize on two multi-objective recommender systems and for three different objective combinations, both correlated or conflicting. We report significant improvements, measured with three different Pareto front metrics: hypervolume, coverage, and spacing. Finally, we show that the \textit{Adamized} Pareto front strictly dominates the previous one on multiple objective pairs.

Kirtan Padh, Diego Antognini, Emma Lejal Glaude et al.

The goal of fairness in classification is to learn a classifier that does not discriminate against groups of individuals based on sensitive attributes, such as race and gender. One approach to designing fair algorithms is to use relaxations of fairness notions as regularization terms or in a constrained optimization problem. We observe that the hyperbolic tangent function can approximate the indicator function. We leverage this property to define a differentiable relaxation that approximates fairness notions provably better than existing relaxations. In addition, we propose a model-agnostic multi-objective architecture that can simultaneously optimize for multiple fairness notions and multiple sensitive attributes and supports all statistical parity-based notions of fairness. We use our relaxation with the multi-objective architecture to learn fair classifiers. Experiments on public datasets show that our method suffers a significantly lower loss of accuracy than current debiasing algorithms relative to the unconstrained model.

Fei Mi, Xiaoyu Lin, Boi Faltings

Session-based recommendation has received growing attention recently due to the increasing privacy concern. Despite the recent success of neural session-based recommenders, they are typically developed in an offline manner using a static dataset. However, recommendation requires continual adaptation to take into account new and obsolete items and users, and requires "continual learning" in real-life applications. In this case, the recommender is updated continually and periodically with new data that arrives in each update cycle, and the updated model needs to provide recommendations for user activities before the next model update. A major challenge for continual learning with neural models is catastrophic forgetting, in which a continually trained model forgets user preference patterns it has learned before. To deal with this challenge, we propose a method called Adaptively Distilled Exemplar Replay (ADER) by periodically replaying previous training samples (i.e., exemplars) to the current model with an adaptive distillation loss. Experiments are conducted based on the state-of-the-art SASRec model using two widely used datasets to benchmark ADER with several well-known continual learning techniques. We empirically demonstrate that ADER consistently outperforms other baselines, and it even outperforms the method using all historical data at every update cycle. This result reveals that ADER is a promising solution to mitigate the catastrophic forgetting issue towards building more realistic and scalable session-based recommenders.

Diego Antognini, Claudiu Musat, Boi Faltings

Using personalized explanations to support recommendations has been shown to increase trust and perceived quality. However, to actually obtain better recommendations, there needs to be a means for users to modify the recommendation criteria by interacting with the explanation. We present a novel technique using aspect markers that learns to generate personalized explanations of recommendations from review texts, and we show that human users significantly prefer these explanations over those produced by state-of-the-art techniques. Our work's most important innovation is that it allows users to react to a recommendation by critiquing the textual explanation: removing (symmetrically adding) certain aspects they dislike or that are no longer relevant (symmetrically that are of interest). The system updates its user model and the resulting recommendations according to the critique. This is based on a novel unsupervised critiquing method for single- and multi-step critiquing with textual explanations. Experiments on two real-world datasets show that our system is the first to achieve good performance in adapting to the preferences expressed in multi-step critiquing.

Fei Mi, Boi Faltings

Increasing concerns with privacy have stimulated interests in Session-based Recommendation (SR) using no personal data other than what is observed in the current browser session. Existing methods are evaluated in static settings which rarely occur in real-world applications. To better address the dynamic nature of SR tasks, we study an incremental SR scenario, where new items and preferences appear continuously. We show that existing neural recommenders can be used in incremental SR scenarios with small incremental updates to alleviate computation overhead and catastrophic forgetting. More importantly, we propose a general framework called Memory Augmented Neural model (MAN). MAN augments a base neural recommender with a continuously queried and updated nonparametric memory, and the predictions from the neural and the memory components are combined through another lightweight gating network. We empirically show that MAN is well-suited for the incremental SR task, and it consistently outperforms state-of-the-art neural and nonparametric methods. We analyze the results and demonstrate that it is particularly good at incrementally learning preferences on new and infrequent items.

Aleksei Triastcyn, Boi Faltings

This paper considers the problem of enhancing user privacy in common machine learning development tasks, such as data annotation and inspection, by substituting the real data with samples form a generative adversarial network. We propose employing Bayesian differential privacy as the means to achieve a rigorous theoretical guarantee while providing a better privacy-utility trade-off. We demonstrate experimentally that our approach produces higher-fidelity samples, compared to prior work, allowing to (1) detect more subtle data errors and biases, and (2) reduce the need for real data labelling by achieving high accuracy when training directly on artificial samples.

Panayiotis Danassis, Marija Sakota, Aris Filos-Ratsikas et al.

We study the optimization of large-scale, real-time ridesharing systems and propose a modular design methodology, Component Algorithms for Ridesharing (CAR). We evaluate a diverse set of CARs (14 in total), focusing on the key algorithmic components of ridesharing. We take a multi-objective approach, evaluating 12 metrics related to global efficiency, complexity, passenger, driver, and platform incentives, in settings designed to closely resemble reality in every aspect, focusing on vehicles of capacity two. To the best of our knowledge, this is the largest and most comprehensive evaluation to date. We (i) identify CARs that perform well on global, passenger, driver or platform metrics, (ii) demonstrate that lightweight relocation schemes can significantly improve the Quality of Service by up to $50\%$, and (iii) highlight a practical, scalable, on-device CAR that works well across all metrics.

Nikola Milojkovic, Diego Antognini, Giancarlo Bergamin et al.

Recommender systems need to mirror the complexity of the environment they are applied in. The more we know about what might benefit the user, the more objectives the recommender system has. In addition there may be multiple stakeholders - sellers, buyers, shareholders - in addition to legal and ethical constraints. Simultaneously optimizing for a multitude of objectives, correlated and not correlated, having the same scale or not, has proven difficult so far. We introduce a stochastic multi-gradient descent approach to recommender systems (MGDRec) to solve this problem. We show that this exceeds state-of-the-art methods in traditional objective mixtures, like revenue and recall. Not only that, but through gradient normalization we can combine fundamentally different objectives, having diverse scales, into a single coherent framework. We show that uncorrelated objectives, like the proportion of quality products, can be improved alongside accuracy. Through the use of stochasticity, we avoid the pitfalls of calculating full gradients and provide a clear setting for its applicability.

Aleksei Triastcyn, Boi Faltings

We consider the problem of reinforcing federated learning with formal privacy guarantees. We propose to employ Bayesian differential privacy, a relaxation of differential privacy for similarly distributed data, to provide sharper privacy loss bounds. We adapt the Bayesian privacy accounting method to the federated setting and suggest multiple improvements for more efficient privacy budgeting at different levels. Our experiments show significant advantage over the state-of-the-art differential privacy bounds for federated learning on image classification tasks, including a medical application, bringing the privacy budget below 1 at the client level, and below 0.1 at the instance level. Lower amounts of noise also benefit the model accuracy and reduce the number of communication rounds.

Aleksei Triastcyn, Boi Faltings

In this paper, we propose FedGP, a framework for privacy-preserving data release in the federated learning setting. We use generative adversarial networks, generator components of which are trained by FedAvg algorithm, to draw privacy-preserving artificial data samples and empirically assess the risk of information disclosure. Our experiments show that FedGP is able to generate labelled data of high quality to successfully train and validate supervised models. Finally, we demonstrate that our approach significantly reduces vulnerability of such models to model inversion attacks.

Diego Antognini, Claudiu Musat, Boi Faltings

Automated predictions require explanations to be interpretable by humans. Past work used attention and rationale mechanisms to find words that predict the target variable of a document. Often though, they result in a tradeoff between noisy explanations or a drop in accuracy. Furthermore, rationale methods cannot capture the multi-faceted nature of justifications for multiple targets, because of the non-probabilistic nature of the mask. In this paper, we propose the Multi-Target Masker (MTM) to address these shortcomings. The novelty lies in the soft multi-dimensional mask that models a relevance probability distribution over the set of target variables to handle ambiguities. Additionally, two regularizers guide MTM to induce long, meaningful explanations. We evaluate MTM on two datasets and show, using standard metrics and human annotations, that the resulting masks are more accurate and coherent than those generated by the state-of-the-art methods. Moreover, MTM is the first to also achieve the highest F1 scores for all the target variables simultaneously.