Machine Learning (cs.LG)

Abstract: Simulating and modeling the long-term dynamics of multi-object physical systems is an essential and challenging task. Current studies model the physical systems utilizing Graph Neural Networks (GNNs) with equivariant properties. Specifically, they model the dynamics as a sequence of discrete states with a fixed time interval and learn a direct mapping for all the two adjacent states. However, this direct mapping overlooks the continuous nature between the two states. Namely, we have verified that there are countless possible trajectories between two discrete dynamic states in current GNN-based direct mapping models. This issue greatly hinders the model generalization ability, leading to poor performance of the long-term simulation. In this paper, to better model the latent trajectory through discrete supervision signals, we propose a Physics-Inspired Neural Graph ODE (PINGO) algorithm. In PINGO, to ensure the uniqueness of the trajectory, we construct a Physics-Inspired Neural ODE framework to update the latent trajectory. Meanwhile, to effectively capture intricate interactions among objects, we use a GNN-based model to parameterize Neural ODE in a plug-and-play manner. Furthermore, we prove that the discrepancy between the learned trajectory of PIGNO and the true trajectory can be theoretically bounded. Extensive experiments verify our theoretical findings and demonstrate that our model yields an order-of-magnitude improvement over the state-of-the-art baselines, especially on long-term predictions and roll-out errors.

Abstract: In learning-to-rank (LTR), optimizing only the relevance (or the expected ranking utility) can cause representational harm to certain categories of items. Moreover, if there is implicit bias in the relevance scores, LTR models may fail to optimize for true relevance. Previous works have proposed efficient algorithms to train stochastic ranking models that achieve fairness of exposure to the groups ex-ante (or, in expectation), which may not guarantee representation fairness to the groups ex-post, that is, after realizing a ranking from the stochastic ranking model. Typically, ex-post fairness is achieved by post-processing, but previous work does not train stochastic ranking models that are aware of this post-processing. In this paper, we propose a novel objective that maximizes expected relevance only over those rankings that satisfy given representation constraints to ensure ex-post fairness. Building upon recent work on an efficient sampler for ex-post group-fair rankings, we propose a group-fair Plackett-Luce model and show that it can be efficiently optimized for our objective in the LTR framework. Experiments on three real-world datasets show that our group-fair algorithm guarantees fairness alongside usually having better relevance compared to the LTR baselines. In addition, our algorithm also achieves better relevance than post-processing baselines, which also ensures ex-post fairness. Further, when implicit bias is injected into the training data, our algorithm typically outperforms existing LTR baselines in relevance.

Abstract: Model-free RL-based recommender systems have recently received increasing research attention due to their capability to handle partial feedback and long-term rewards. However, most existing research has ignored a critical feature in recommender systems: one user's feedback on the same item at different times is random. The stochastic rewards property essentially differs from that in classic RL scenarios with deterministic rewards, which makes RL-based recommender systems much more challenging. In this paper, we first demonstrate in a simulator environment where using direct stochastic feedback results in a significant drop in performance. Then to handle the stochastic feedback more efficiently, we design two stochastic reward stabilization frameworks that replace the direct stochastic feedback with that learned by a supervised model. Both frameworks are model-agnostic, i.e., they can effectively utilize various supervised models. We demonstrate the superiority of the proposed frameworks over different RL-based recommendation baselines with extensive experiments on a recommendation simulator as well as an industrial-level recommender system.

Abstract: Permutation matrices play a key role in matching and assignment problems across the fields, especially in computer vision and robotics. However, memory for explicitly representing permutation matrices grows quadratically with the size of the problem, prohibiting large problem instances. In this work, we propose to tackle the curse of dimensionality of large permutation matrices by approximating them using low-rank matrix factorization, followed by a nonlinearity. To this end, we rely on the Kissing number theory to infer the minimal rank required for representing a permutation matrix of a given size, which is significantly smaller than the problem size. This leads to a drastic reduction in computation and memory costs, e.g., up to $3$ orders of magnitude less memory for a problem of size $n=20000$, represented using $8.4\times10^5$ elements in two small matrices instead of using a single huge matrix with $4\times 10^8$ elements. The proposed representation allows for accurate representations of large permutation matrices, which in turn enables handling large problems that would have been infeasible otherwise. We demonstrate the applicability and merits of the proposed approach through a series of experiments on a range of problems that involve predicting permutation matrices, from linear and quadratic assignment to shape matching problems.

Abstract: Federated Learning (FL) allows several clients to construct a common global machine-learning model without having to share their data. FL, however, faces the challenge of statistical heterogeneity between the client's data, which degrades performance and slows down the convergence toward the global model. In this paper, we provide theoretical proof that minimizing heterogeneity between clients facilitates the convergence of a global model for every single client. This becomes particularly important under empirical concept shifts among clients, rather than merely considering imbalanced classes, which have been studied until now. Therefore, we propose a method for knowledge transfer between clients where the server trains client-specific generators. Each generator generates samples for the corresponding client to remove the conflict with other clients' models. Experiments conducted on synthetic and real data, along with a theoretical study, support the effectiveness of our method in constructing a well-generalizable global model by reducing the conflict between local models.

Abstract: As some recent information security legislation endowed users with unconditional rights to be forgotten by any trained machine learning model, personalized IoT service providers have to put unlearning functionality into their consideration. The most straightforward method to unlearn users' contribution is to retrain the model from the initial state, which is not realistic in high throughput applications with frequent unlearning requests. Though some machine unlearning frameworks have been proposed to speed up the retraining process, they fail to match decentralized learning scenarios. In this paper, we design a decentralized unlearning framework called HDUS, which uses distilled seed models to construct erasable ensembles for all clients. Moreover, the framework is compatible with heterogeneous on-device models, representing stronger scalability in real-world applications. Extensive experiments on three real-world datasets show that our HDUS achieves state-of-the-art performance.

Abstract: Quality-Diversity is a branch of stochastic optimization that is often applied to problems from the Reinforcement Learning and control domains in order to construct repertoires of well-performing policies/skills that exhibit diversity with respect to a behavior space. Such archives are usually composed of a finite number of reactive agents which are each associated to a unique behavior descriptor, and instantiating behavior descriptors outside of that coarsely discretized space is not straight-forward. While a few recent works suggest solutions to that issue, the trajectory that is generated is not easily customizable beyond the specification of a target behavior descriptor. We propose to jointly solve those problems in environments where semantic information about static scene elements is available by leveraging a Large Language Model to augment the repertoire with natural language descriptions of trajectories, and training a policy conditioned on those descriptions. Thus, our method allows a user to not only specify an arbitrary target behavior descriptor, but also provide the model with a high-level textual prompt to shape the generated trajectory. We also propose an LLM-based approach to evaluating the performance of such generative agents. Furthermore, we develop a benchmark based on simulated robot navigation in a 2d maze that we use for experimental validation.

Abstract: Hyperbolic space is becoming a popular choice for representing data due to the hierarchical structure - whether implicit or explicit - of many real-world datasets. Along with it comes a need for algorithms capable of solving fundamental tasks, such as classification, in hyperbolic space. Recently, multiple papers have investigated hyperbolic alternatives to hyperplane-based classifiers, such as logistic regression and SVMs. While effective, these approaches struggle with more complex hierarchical data. We, therefore, propose to generalize the well-known random forests to hyperbolic space. We do this by redefining the notion of a split using horospheres. Since finding the globally optimal split is computationally intractable, we find candidate horospheres through a large-margin classifier. To make hyperbolic random forests work on multi-class data and imbalanced experiments, we furthermore outline a new method for combining classes based on their lowest common ancestor and a class-balanced version of the large-margin loss. Experiments on standard and new benchmarks show that our approach outperforms both conventional random forest algorithms and recent hyperbolic classifiers.

Abstract: Assessment is a crucial part of education. Traditional marking is a source of inconsistencies and unconscious bias, placing a high cognitive load on the assessors. An approach to address these issues is comparative judgement (CJ). In CJ, the assessor is presented with a pair of items and is asked to select the better one. Following a series of comparisons, a rank is derived using a ranking model, for example, the BTM, based on the results. While CJ is considered a reliable method for marking, there are concerns around transparency, and the ideal number of pairwise comparisons to generate a reliable estimation of the rank order is not known. Additionally, there have been attempts to generate a method of selecting pairs that should be compared next in an informative manner, but some existing methods are known to have created their own bias within results inflating the reliability metric used. As a result, a random selection approach is usually deployed. We propose a novel Bayesian approach to CJ (BCJ) for determining the ranks of compared items alongside a new way to select the pairs to present to the marker(s) using active learning (AL), addressing the key shortcomings of traditional CJ. Furthermore, we demonstrate how the entire approach may provide transparency by providing the user insights into how it is making its decisions and, at the same time, being more efficient. Results from our experiments confirm that the proposed BCJ combined with entropy-driven AL pair-selection method is superior to other alternatives. We also find that the more comparisons done, the more accurate BCJ becomes, which solves the issue the current method has of the model deteriorating if too many comparisons are performed. As our approach can generate the complete predicted rank distribution for an item, we also show how this can be utilised in devising a predicted grade, guided by the assessor.

Abstract: Machine learning techniques, in particular the so-called normalizing flows, are becoming increasingly popular in the context of Monte Carlo simulations as they can effectively approximate target probability distributions. In the case of lattice field theories (LFT) the target distribution is given by the exponential of the action. The common loss function's gradient estimator based on the "reparametrization trick" requires the calculation of the derivative of the action with respect to the fields. This can present a significant computational cost for complicated, non-local actions like e.g. fermionic action in QCD. In this contribution, we propose an estimator for normalizing flows based on the REINFORCE algorithm that avoids this issue. We apply it to two dimensional Schwinger model with Wilson fermions at criticality and show that it is up to ten times faster in terms of the wall-clock time as well as requiring up to $30\%$ less memory than the reparameterization trick estimator. It is also more numerically stable allowing for single precision calculations and the use of half-float tensor cores. We present an in-depth analysis of the origins of those improvements. We believe that these benefits will appear also outside the realm of the LFT, in each case where the target probability distribution is computationally intensive.

Abstract: In this paper, we investigate federated contextual linear bandit learning within a wireless system that comprises a server and multiple devices. Each device interacts with the environment, selects an action based on the received reward, and sends model updates to the server. The primary objective is to minimize cumulative regret across all devices within a finite time horizon. To reduce the communication overhead, devices communicate with the server via over-the-air computation (AirComp) over noisy fading channels, where the channel noise may distort the signals. In this context, we propose a customized federated linear bandits scheme, where each device transmits an analog signal, and the server receives a superposition of these signals distorted by channel noise. A rigorous mathematical analysis is conducted to determine the regret bound of the proposed scheme. Both theoretical analysis and numerical experiments demonstrate the competitive performance of our proposed scheme in terms of regret bounds in various settings.

Abstract: Compact neural network offers many benefits for real-world applications. However, it is usually challenging to train the compact neural networks with small parameter sizes and low computational costs to achieve the same or better model performance compared to more complex and powerful architecture. This is particularly true for multitask learning, with different tasks competing for resources. We present a simple, efficient and effective multitask learning overparameterisation neural network design by overparameterising the model architecture in training and sharing the overparameterised model parameters more effectively across tasks, for better optimisation and generalisation. Experiments on two challenging multitask datasets (NYUv2 and COCO) demonstrate the effectiveness of the proposed method across various convolutional networks and parameter sizes.

Abstract: Pre-trained foundation models, owing primarily to their enormous capacity and exposure to vast amount of training data scraped from the internet, enjoy the advantage of storing knowledge about plenty of real-world concepts. Such models are typically fine-tuned on downstream datasets to produce remarkable state-of-the-art performances. While various fine-tuning methods have been devised and are shown to be highly effective, we observe that a fine-tuned model's ability to recognize concepts on tasks $\textit{different}$ from the downstream one is reduced significantly compared to its pre-trained counterpart. This is clearly undesirable as a huge amount of time and money went into learning those very concepts in the first place. We call this undesirable phenomenon "concept forgetting" and via experiments show that most end-to-end fine-tuning approaches suffer heavily from this side effect. To this end, we also propose a rather simple fix to this problem by designing a method called LDIFS (short for $\ell_2$ distance in feature space) that simply preserves the features of the original foundation model during fine-tuning. We show that LDIFS significantly reduces concept forgetting without having noticeable impact on the downstream task performance.

Abstract: Anomaly detection (AD) tasks have been solved using machine learning algorithms in various domains and applications. The great majority of these algorithms use normal data to train a residual-based model, and assign anomaly scores to unseen samples based on their dissimilarity with the learned normal regime. The underlying assumption of these approaches is that anomaly-free data is available for training. This is, however, often not the case in real-world operational settings, where the training data may be contaminated with a certain fraction of abnormal samples. Training with contaminated data, in turn, inevitably leads to a deteriorated AD performance of the residual-based algorithms. In this paper we introduce a framework for a fully unsupervised refinement of contaminated training data for AD tasks. The framework is generic and can be applied to any residual-based machine learning model. We demonstrate the application of the framework to two public datasets of multivariate time series machine data from different application fields. We show its clear superiority over the naive approach of training with contaminated data without refinement. Moreover, we compare it to the ideal, unrealistic reference in which anomaly-free data would be available for training. Since the approach exploits information from the anomalies, and not only from the normal regime, it is comparable and often outperforms the ideal baseline as well.

Abstract: Long-term time series forecasting is a vital task and has a wide range of real applications. Recent methods focus on capturing the underlying patterns from one single domain (e.g. the time domain or the frequency domain), and have not taken a holistic view to process long-term time series from the time-frequency domains. In this paper, we propose a Time-Frequency Enhanced Decomposed Network (TFDNet) to capture both the long-term underlying patterns and temporal periodicity from the time-frequency domain. In TFDNet, we devise a multi-scale time-frequency enhanced encoder backbone and develop two separate trend and seasonal time-frequency blocks to capture the distinct patterns within the decomposed trend and seasonal components in multi-resolutions. Diverse kernel learning strategies of the kernel operations in time-frequency blocks have been explored, by investigating and incorporating the potential different channel-wise correlation patterns of multivariate time series. Experimental evaluation of eight datasets from five benchmark domains demonstrated that TFDNet is superior to state-of-the-art approaches in both effectiveness and efficiency.

Abstract: Risky drivers account for 70% of fatal accidents in the United States. With recent advances in sensors and intelligent vehicular systems, there has been significant research on assessing driver behavior to improve driving experiences and road safety. This paper examines the various techniques used to analyze driver behavior using visual and vehicular data, providing an overview of the latest research in this field. The paper also discusses the challenges and open problems in the field and offers potential recommendations for future research. The survey concludes that integrating vision and vehicular information can significantly enhance the accuracy and effectiveness of driver behavior analysis, leading to improved safety measures and reduced traffic accidents.

Abstract: Scientific knowledge is predominantly stored in books and scientific journals, often in the form of PDFs. However, the PDF format leads to a loss of semantic information, particularly for mathematical expressions. We propose Nougat (Neural Optical Understanding for Academic Documents), a Visual Transformer model that performs an Optical Character Recognition (OCR) task for processing scientific documents into a markup language, and demonstrate the effectiveness of our model on a new dataset of scientific documents. The proposed approach offers a promising solution to enhance the accessibility of scientific knowledge in the digital age, by bridging the gap between human-readable documents and machine-readable text. We release the models and code to accelerate future work on scientific text recognition.

Abstract: While traditional deep learning models often lack interpretability, concept bottleneck models (CBMs) provide inherent explanations via their concept representations. Specifically, they allow users to perform interventional interactions on these concepts by updating the concept values and thus correcting the predictive output of the model. Traditionally, however, these interventions are applied to the model only once and discarded afterward. To rectify this, we present concept bottleneck memory models (CB2M), an extension to CBMs. Specifically, a CB2M learns to generalize interventions to appropriate novel situations via a two-fold memory with which it can learn to detect mistakes and to reapply previous interventions. In this way, a CB2M learns to automatically improve model performance from a few initially obtained interventions. If no prior human interventions are available, a CB2M can detect potential mistakes of the CBM bottleneck and request targeted interventions. In our experimental evaluations on challenging scenarios like handling distribution shifts and confounded training data, we illustrate that CB2M are able to successfully generalize interventions to unseen data and can indeed identify wrongly inferred concepts. Overall, our results show that CB2M is a great tool for users to provide interactive feedback on CBMs, e.g., by guiding a user's interaction and requiring fewer interventions.

Abstract: Despite the recent success of Graph Neural Networks (GNNs), it remains challenging to train GNNs on large-scale graphs due to neighbor explosions. As a remedy, distributed computing becomes a promising solution by leveraging abundant computing resources (e.g., GPU). However, the node dependency of graph data increases the difficulty of achieving high concurrency in distributed GNN training, which suffers from the massive communication overhead. To address it, Historical value approximation is deemed a promising class of distributed training techniques. It utilizes an offline memory to cache historical information (e.g., node embedding) as an affordable approximation of the exact value and achieves high concurrency. However, such benefits come at the cost of involving dated training information, leading to staleness, imprecision, and convergence issues. To overcome these challenges, this paper proposes SAT (Staleness-Alleviated Training), a novel and scalable distributed GNN training framework that reduces the embedding staleness adaptively. The key idea of SAT is to model the GNN's embedding evolution as a temporal graph and build a model upon it to predict future embedding, which effectively alleviates the staleness of the cached historical embedding. We propose an online algorithm to train the embedding predictor and the distributed GNN alternatively and further provide a convergence analysis. Empirically, we demonstrate that SAT can effectively reduce embedding staleness and thus achieve better performance and convergence speed on multiple large-scale graph datasets.

Abstract: Precise hardware performance models play a crucial role in code optimizations. They can assist compilers in making heuristic decisions or aid autotuners in identifying the optimal configuration for a given program. For example, the autotuner for XLA, a machine learning compiler, discovered 10-20% speedup on state-of-the-art models serving substantial production traffic at Google. Although there exist a few datasets for program performance prediction, they target small sub-programs such as basic blocks or kernels. This paper introduces TpuGraphs, a performance prediction dataset on full tensor programs, represented as computational graphs, running on Tensor Processing Units (TPUs). Each graph in the dataset represents the main computation of a machine learning workload, e.g., a training epoch or an inference step. Each data sample contains a computational graph, a compilation configuration, and the execution time of the graph when compiled with the configuration. The graphs in the dataset are collected from open-source machine learning programs, featuring popular model architectures, e.g., ResNet, EfficientNet, Mask R-CNN, and Transformer. TpuGraphs provides 25x more graphs than the largest graph property prediction dataset (with comparable graph sizes), and 770x larger graphs on average compared to existing performance prediction datasets on machine learning programs. This graph-level prediction task on large graphs introduces new challenges in learning, ranging from scalability, training efficiency, to model quality.

Abstract: This work introduces a novel approach for epistemic uncertainty estimation for ensemble models using pairwise-distance estimators (PaiDEs). These estimators utilize the pairwise-distance between model components to establish bounds on entropy and uses said bounds as estimates for information-based criterion. Unlike recent deep learning methods for epistemic uncertainty estimation, which rely on sample-based Monte Carlo estimators, PaiDEs are able to estimate epistemic uncertainty up to 100$\times$ faster, over a larger space (up to 100$\times$) and perform more accurately in higher dimensions. To validate our approach, we conducted a series of experiments commonly used to evaluate epistemic uncertainty estimation: 1D sinusoidal data, Pendulum-v0, Hopper-v2, Ant-v2 and Humanoid-v2. For each experimental setting, an Active Learning framework was applied to demonstrate the advantages of PaiDEs for epistemic uncertainty estimation.

Abstract: We present accumulator-aware quantization (A2Q), a novel weight quantization method designed to train quantized neural networks (QNNs) to avoid overflow when using low-precision accumulators during inference. A2Q introduces a unique formulation inspired by weight normalization that constrains the L1-norm of model weights according to accumulator bit width bounds that we derive. Thus, in training QNNs for low-precision accumulation, A2Q also inherently promotes unstructured weight sparsity to guarantee overflow avoidance. We apply our method to deep learning-based computer vision tasks to show that A2Q can train QNNs for low-precision accumulators while maintaining model accuracy competitive with a floating-point baseline. In our evaluations, we consider the impact of A2Q on both general-purpose platforms and programmable hardware. However, we primarily target model deployment on FPGAs because they can be programmed to fully exploit custom accumulator bit widths. Our experimentation shows accumulator bit width significantly impacts the resource efficiency of FPGA-based accelerators. On average across our benchmarks, A2Q offers up to a 2.3x reduction in resource utilization over 32-bit accumulator counterparts with 99.2% of the floating-point model accuracy.

Abstract: Graph Neural Networks (GNNs) are powerful tools for learning representations on graphs, such as social networks. However, their vulnerability to privacy inference attacks restricts their practicality, especially in high-stake domains. To address this issue, privacy-preserving GNNs have been proposed, focusing on preserving node and/or link privacy. This work takes a step back and investigates how GNNs contribute to privacy leakage. Through theoretical analysis and simulations, we identify message passing under structural bias as the core component that allows GNNs to \textit{propagate} and \textit{amplify} privacy leakage. Building upon these findings, we propose a principled privacy-preserving GNN framework that effectively safeguards both node and link privacy, referred to as dual-privacy preservation. The framework comprises three major modules: a Sensitive Information Obfuscation Module that removes sensitive information from node embeddings, a Dynamic Structure Debiasing Module that dynamically corrects the structural bias, and an Adversarial Learning Module that optimizes the privacy-utility trade-off. Experimental results on four benchmark datasets validate the effectiveness of the proposed model in protecting both node and link privacy while preserving high utility for downstream tasks, such as node classification.