Machine Learning (cs.LG)

Abstract: Variational autoencoders (VAEs) are used for transfer learning across various research domains such as music generation or medical image analysis. However, there is no principled way to assess before transfer which components to retrain or whether transfer learning is likely to help on a target task. We propose to explore this question through the lens of representational similarity. Specifically, using Centred Kernel Alignment (CKA) to evaluate the similarity of VAEs trained on different datasets, we show that encoders' representations are generic but decoders' specific. Based on these insights, we discuss the implications for selecting which components of a VAE to retrain and propose a method to visually assess whether transfer learning is likely to help on classification tasks.

Abstract: Exploration is a fundamental aspect of reinforcement learning (RL), and its effectiveness crucially decides the performance of RL algorithms, especially when facing sparse extrinsic rewards. Recent studies showed the effectiveness of encouraging exploration with intrinsic rewards estimated from novelty in observations. However, there is a gap between the novelty of an observation and an exploration in general, because the stochasticity in the environment as well as the behavior of an agent may affect the observation. To estimate exploratory behaviors accurately, we propose DEIR, a novel method where we theoretically derive an intrinsic reward from a conditional mutual information term that principally scales with the novelty contributed by agent explorations, and materialize the reward with a discriminative forward model. We conduct extensive experiments in both standard and hardened exploration games in MiniGrid to show that DEIR quickly learns a better policy than baselines. Our evaluations in ProcGen demonstrate both generalization capabilities and the general applicability of our intrinsic reward.

Abstract: Federated learning (FL) is vulnerable to poisoning attacks, where adversaries corrupt the global aggregation results and cause denial-of-service (DoS). Unlike recent model poisoning attacks that optimize the amplitude of malicious perturbations along certain prescribed directions to cause DoS, we propose a Flexible Model Poisoning Attack (FMPA) that can achieve versatile attack goals. We consider a practical threat scenario where no extra knowledge about the FL system (e.g., aggregation rules or updates on benign devices) is available to adversaries. FMPA exploits the global historical information to construct an estimator that predicts the next round of the global model as a benign reference. It then fine-tunes the reference model to obtain the desired poisoned model with low accuracy and small perturbations. Besides the goal of causing DoS, FMPA can be naturally extended to launch a fine-grained controllable attack, making it possible to precisely reduce the global accuracy. Armed with precise control, malicious FL service providers can gain advantages over their competitors without getting noticed, hence opening a new attack surface in FL other than DoS. Even for the purpose of DoS, experiments show that FMPA significantly decreases the global accuracy, outperforming six state-of-the-art attacks.

Abstract: Data collected from the real world tends to be biased, unbalanced, and at risk of exposing sensitive and private information. This reality has given rise to the idea of creating synthetic datasets to alleviate risk, bias, harm, and privacy concerns inherent in the real data. This concept relies on Generative AI models to produce unbiased, privacy-preserving synthetic data while being true to the real data. In this new paradigm, how can we tell if this approach delivers on its promises? We present an auditing framework that offers a holistic assessment of synthetic datasets and AI models trained on them, centered around bias and discrimination prevention, fidelity to the real data, utility, robustness, and privacy preservation. We showcase our framework by auditing multiple generative models on diverse use cases, including education, healthcare, banking, human resources, and across different modalities, from tabular, to time-series, to natural language. Our use cases demonstrate the importance of a holistic assessment in order to ensure compliance with socio-technical safeguards that regulators and policymakers are increasingly enforcing. For this purpose, we introduce the trust index that ranks multiple synthetic datasets based on their prescribed safeguards and their desired trade-offs. Moreover, we devise a trust-index-driven model selection and cross-validation procedure via auditing in the training loop that we showcase on a class of transformer models that we dub TrustFormers, across different modalities. This trust-driven model selection allows for controllable trust trade-offs in the resulting synthetic data. We instrument our auditing framework with workflows that connect different stakeholders from model development to audit and certification via a synthetic data auditing report.

Abstract: We study Individual Fairness (IF) for Bayesian neural networks (BNNs). Specifically, we consider the $\epsilon$-$\delta$-individual fairness notion, which requires that, for any pair of input points that are $\epsilon$-similar according to a given similarity metrics, the output of the BNN is within a given tolerance $\delta>0.$ We leverage bounds on statistical sampling over the input space and the relationship between adversarial robustness and individual fairness to derive a framework for the systematic estimation of $\epsilon$-$\delta$-IF, designing Fair-FGSM and Fair-PGD as global,fairness-aware extensions to gradient-based attacks for BNNs. We empirically study IF of a variety of approximately inferred BNNs with different architectures on fairness benchmarks, and compare against deterministic models learnt using frequentist techniques. Interestingly, we find that BNNs trained by means of approximate Bayesian inference consistently tend to be markedly more individually fair than their deterministic counterparts.

Abstract: Classification trees continue to be widely adopted in machine learning applications due to their inherently interpretable nature and scalability. We propose a rolling subtree lookahead algorithm that combines the relative scalability of the myopic approaches with the foresight of the optimal approaches in constructing trees. The limited foresight embedded in our algorithm mitigates the learning pathology observed in optimal approaches. At the heart of our algorithm lies a novel two-depth optimal binary classification tree formulation flexible to handle any loss function. We show that the feasible region of this formulation is an integral polyhedron, yielding the LP relaxation solution optimal. Through extensive computational analyses, we demonstrate that our approach outperforms optimal and myopic approaches in 808 out of 1330 problem instances, improving the out-of-sample accuracy by up to 23.6% and 14.4%, respectively.

Abstract: In recent years, graph neural networks (GNNs) have achieved great success in the field of graph representation learning. Although prior work has shed light into the expressiveness of those models (\ie whether they can distinguish pairs of non-isomorphic graphs), it is still not clear what structural information is encoded into the node representations that are learned by those models. In this paper, we investigate which properties of graphs are captured purely by these models, when no node attributes are available. Specifically, we study four popular GNN models, and we show that two of them embed all nodes into the same feature vector, while the other two models generate representations that are related to the number of walks over the input graph. Strikingly, structurally dissimilar nodes can have similar representations at some layer $k>1$, if they have the same number of walks of length $k$. We empirically verify our theoretical findings on real datasets.

Abstract: Continual Learning is an important and challenging problem in machine learning, where models must adapt to a continuous stream of new data without forgetting previously acquired knowledge. While existing frameworks are built on PyTorch, the rising popularity of JAX might lead to divergent codebases, ultimately hindering reproducibility and progress. To address this problem, we introduce SequeL, a flexible and extensible library for Continual Learning that supports both PyTorch and JAX frameworks. SequeL provides a unified interface for a wide range of Continual Learning algorithms, including regularization-based approaches, replay-based approaches, and hybrid approaches. The library is designed towards modularity and simplicity, making the API suitable for both researchers and practitioners. We release SequeL\footnote{\url{https://github.com/nik-dim/sequel}} as an open-source library, enabling researchers and developers to easily experiment and extend the library for their own purposes.

Abstract: The quality of data driven learning algorithms scales significantly with the quality of data available. One of the most straight-forward ways to generate good data is to sample or explore the data source intelligently. Smart sampling can reduce the cost of gaining samples, reduce computation cost in learning, and enable the learning algorithm to adapt to unforeseen events. In this paper, we teach three Deep Q-Networks (DQN) with different exploration strategies to solve a problem of puncturing ongoing transmissions for URLLC messages. We demonstrate the efficiency of two adaptive exploration candidates, variance-based and Maximum Entropy-based exploration, compared to the standard, simple epsilon-greedy exploration approach.

Abstract: Transformer architectures have exhibited promising performance in various autonomous driving applications in recent years. On the other hand, its dedicated hardware acceleration on portable computational platforms has become the next critical step for practical deployment in real autonomous vehicles. This survey paper provides a comprehensive overview, benchmark, and analysis of Transformer-based models specifically tailored for autonomous driving tasks such as lane detection, segmentation, tracking, planning, and decision-making. We review different architectures for organizing Transformer inputs and outputs, such as encoder-decoder and encoder-only structures, and explore their respective advantages and disadvantages. Furthermore, we discuss Transformer-related operators and their hardware acceleration schemes in depth, taking into account key factors such as quantization and runtime. We specifically illustrate the operator level comparison between layers from convolutional neural network, Swin-Transformer, and Transformer with 4D encoder. The paper also highlights the challenges, trends, and current insights in Transformer-based models, addressing their hardware deployment and acceleration issues within the context of long-term autonomous driving applications.

Abstract: The use of machine learning (ML) inference for various applications is growing drastically. ML inference services engage with users directly, requiring fast and accurate responses. Moreover, these services face dynamic workloads of requests, imposing changes in their computing resources. Failing to right-size computing resources results in either latency service level objectives (SLOs) violations or wasted computing resources. Adapting to dynamic workloads considering all the pillars of accuracy, latency, and resource cost is challenging. In response to these challenges, we propose InfAdapter, which proactively selects a set of ML model variants with their resource allocations to meet latency SLO while maximizing an objective function composed of accuracy and cost. InfAdapter decreases SLO violation and costs up to 65% and 33%, respectively, compared to a popular industry autoscaler (Kubernetes Vertical Pod Autoscaler).

Abstract: Graph Neural Networks (GNNs) are well-suited for learning on homophilous graphs, i.e., graphs in which edges tend to connect nodes of the same type. Yet, achievement of consistent GNN performance on heterophilous graphs remains an open research problem. Recent works have proposed extensions to standard GNN architectures to improve performance on heterophilous graphs, trading off model simplicity for prediction accuracy. However, these models fail to capture basic graph properties, such as neighborhood label distribution, which are fundamental for learning. In this work, we propose GCN for Heterophily (GCNH), a simple yet effective GNN architecture applicable to both heterophilous and homophilous scenarios. GCNH learns and combines separate representations for a node and its neighbors, using one learned importance coefficient per layer to balance the contributions of center nodes and neighborhoods. We conduct extensive experiments on eight real-world graphs and a set of synthetic graphs with varying degrees of heterophily to demonstrate how the design choices for GCNH lead to a sizable improvement over a vanilla GCN. Moreover, GCNH outperforms state-of-the-art models of much higher complexity on four out of eight benchmarks, while producing comparable results on the remaining datasets. Finally, we discuss and analyze the lower complexity of GCNH, which results in fewer trainable parameters and faster training times than other methods, and show how GCNH mitigates the oversmoothing problem.

Abstract: Online experiments such as Randomised Controlled Trials (RCTs) or A/B-tests are the bread and butter of modern platforms on the web. They are conducted continuously to allow platforms to estimate the causal effect of replacing system variant "A" with variant "B", on some metric of interest. These variants can differ in many aspects. In this paper, we focus on the common use-case where they correspond to machine learning models. The online experiment then serves as the final arbiter to decide which model is superior, and should thus be shipped. The statistical literature on causal effect estimation from RCTs has a substantial history, which contributes deservedly to the level of trust researchers and practitioners have in this "gold standard" of evaluation practices. Nevertheless, in the particular case of machine learning experiments, we remark that certain critical issues remain. Specifically, the assumptions that are required to ascertain that A/B-tests yield unbiased estimates of the causal effect, are seldom met in practical applications. We argue that, because variants typically learn using pooled data, a lack of model interference cannot be guaranteed. This undermines the conclusions we can draw from online experiments with machine learning models. We discuss the implications this has for practitioners, and for the research literature.

Abstract: Medical coding is the task of assigning medical codes to clinical free-text documentation. Healthcare professionals manually assign such codes to track patient diagnoses and treatments. Automated medical coding can considerably alleviate this administrative burden. In this paper, we reproduce, compare, and analyze state-of-the-art automated medical coding machine learning models. We show that several models underperform due to weak configurations, poorly sampled train-test splits, and insufficient evaluation. In previous work, the macro F1 score has been calculated sub-optimally, and our correction doubles it. We contribute a revised model comparison using stratified sampling and identical experimental setups, including hyperparameters and decision boundary tuning. We analyze prediction errors to validate and falsify assumptions of previous works. The analysis confirms that all models struggle with rare codes, while long documents only have a negligible impact. Finally, we present the first comprehensive results on the newly released MIMIC-IV dataset using the reproduced models. We release our code, model parameters, and new MIMIC-III and MIMIC-IV training and evaluation pipelines to accommodate fair future comparisons.

Abstract: Behavioural cloning is an imitation learning technique that teaches an agent how to behave via expert demonstrations. Recent approaches use self-supervision of fully-observable unlabelled snapshots of the states to decode state pairs into actions. However, the iterative learning scheme employed by these techniques is prone to get trapped into bad local minima. Previous work uses goal-aware strategies to solve this issue. However, this requires manual intervention to verify whether an agent has reached its goal. We address this limitation by incorporating a discriminator into the original framework, offering two key advantages and directly solving a learning problem previous work had. First, it disposes of the manual intervention requirement. Second, it helps in learning by guiding function approximation based on the state transition of the expert's trajectories. Third, the discriminator solves a learning issue commonly present in the policy model, which is to sometimes perform a `no action' within the environment until the agent finally halts.

Abstract: We introduce a novel self-attention mechanism, which we call CSA (Chromatic Self-Attention), which extends the notion of attention scores to attention _filters_, independently modulating the feature channels. We showcase CSA in a fully-attentional graph Transformer CGT (Chromatic Graph Transformer) which integrates both graph structural information and edge features, completely bypassing the need for local message-passing components. Our method flexibly encodes graph structure through node-node interactions, by enriching the original edge features with a relative positional encoding scheme. We propose a new scheme based on random walks that encodes both structural and positional information, and show how to incorporate higher-order topological information, such as rings in molecular graphs. Our approach achieves state-of-the-art results on the ZINC benchmark dataset, while providing a flexible framework for encoding graph structure and incorporating higher-order topology.

Abstract: This work provides a comprehensive derivation of the parameter gradients for GATv2 [4], a widely used implementation of Graph Attention Networks (GATs). GATs have proven to be powerful frameworks for processing graph-structured data and, hence, have been used in a range of applications. However, the achieved performance by these attempts has been found to be inconsistent across different datasets and the reasons for this remains an open research question. As the gradient flow provides valuable insights into the training dynamics of statistically learning models, this work obtains the gradients for the trainable model parameters of GATv2. The gradient derivations supplement the efforts of [2], where potential pitfalls of GATv2 are investigated.

Abstract: We consider the problem of control in the setting of reinforcement learning (RL), where model information is not available. Policy gradient algorithms are a popular solution approach for this problem and are usually shown to converge to a stationary point of the value function. In this paper, we propose two policy Newton algorithms that incorporate cubic regularization. Both algorithms employ the likelihood ratio method to form estimates of the gradient and Hessian of the value function using sample trajectories. The first algorithm requires an exact solution of the cubic regularized problem in each iteration, while the second algorithm employs an efficient gradient descent-based approximation to the cubic regularized problem. We establish convergence of our proposed algorithms to a second-order stationary point (SOSP) of the value function, which results in the avoidance of traps in the form of saddle points. In particular, the sample complexity of our algorithms to find an $\epsilon$-SOSP is $O(\epsilon^{-3.5})$, which is an improvement over the state-of-the-art sample complexity of $O(\epsilon^{-4.5})$.

Abstract: In intelligent transport systems, it is common and inevitable with missing data. While complete and valid traffic speed data is of great importance to intelligent transportation systems. A latent factorization-of-tensors (LFT) model is one of the most attractive approaches to solve missing traffic data recovery due to its well-scalability. A LFT model achieves optimization usually via a stochastic gradient descent (SGD) solver, however, the SGD-based LFT suffers from slow convergence. To deal with this issue, this work integrates the unique advantages of the proportional-integral-derivative (PID) controller into a Tucker decomposition based LFT model. It adopts two-fold ideas: a) adopting tucker decomposition to build a LFT model for achieving a better recovery accuracy. b) taking the adjusted instance error based on the PID control theory into the SGD solver to effectively improve convergence rate. Our experimental studies on two major city traffic road speed datasets show that the proposed model achieves significant efficiency gain and highly competitive prediction accuracy.

Abstract: We investigate the potential of GPT-4~\cite{gpt4} to perform Neural Architecture Search (NAS) -- the task of designing effective neural architectures. Our proposed approach, \textbf{G}PT-4 \textbf{I}nformed \textbf{N}eural \textbf{A}rchitecture \textbf{S}earch (GINAS),leverages the generative capabilities of GPT-4 as a black-box optimiser to quickly navigate the architecture search space, pinpoint promising candidates, and iteratively refine these candidates to improve performance.We assess GINAS across several benchmarks, comparing it with existing state-of-the-art NAS techniques to illustrate its effectiveness. Rather than targeting state-of-the-art performance, our objective is to highlight GPT-4's potential to assist research on a challenging technical problem through a simple prompting scheme that requires relatively limited domain expertise. More broadly, we believe our preliminary results point to future research that harnesses general purpose language models for diverse optimisation tasks. We also highlight important limitations to our study, and note implications for AI safety.

Abstract: Gaussian processes are cemented as the model of choice in Bayesian optimization and active learning. Yet, they are severely dependent on cleverly chosen hyperparameters to reach their full potential, and little effort is devoted to finding the right hyperparameters in the literature. We demonstrate the impact of selecting good hyperparameters for GPs and present two acquisition functions that explicitly prioritize this goal. Statistical distance-based Active Learning (SAL) considers the average disagreement among samples from the posterior, as measured by a statistical distance. It is shown to outperform the state-of-the-art in Bayesian active learning on a number of test functions. We then introduce Self-Correcting Bayesian Optimization (SCoreBO), which extends SAL to perform Bayesian optimization and active hyperparameter learning simultaneously. SCoreBO learns the model hyperparameters at improved rates compared to vanilla BO, while outperforming the latest Bayesian optimization methods on traditional benchmarks. Moreover, the importance of self-correction is demonstrated on an array of exotic Bayesian optimization tasks

Abstract: We present a new general-purpose algorithm for learning classes of $[0,1]$-valued functions in a generalization of the prediction model, and prove a general upper bound on the expected absolute error of this algorithm in terms of a scale-sensitive generalization of the Vapnik dimension proposed by Alon, Ben-David, Cesa-Bianchi and Haussler. We give lower bounds implying that our upper bounds cannot be improved by more than a constant factor in general. We apply this result, together with techniques due to Haussler and to Benedek and Itai, to obtain new upper bounds on packing numbers in terms of this scale-sensitive notion of dimension. Using a different technique, we obtain new bounds on packing numbers in terms of Kearns and Schapire's fat-shattering function. We show how to apply both packing bounds to obtain improved general bounds on the sample complexity of agnostic learning. For each $\epsilon > 0$, we establish weaker sufficient and stronger necessary conditions for a class of $[0,1]$-valued functions to be agnostically learnable to within $\epsilon$, and to be an $\epsilon$-uniform Glivenko-Cantelli class. This is a manuscript that was accepted by JCSS, together with a correction.

Abstract: Data-driven machine learning is playing a crucial role in the advancements of Industry 4.0, specifically in enhancing predictive maintenance and quality inspection. Federated learning (FL) enables multiple participants to develop a machine learning model without compromising the privacy and confidentiality of their data. In this paper, we evaluate the performance of different FL aggregation methods and compare them to central and local training approaches. Our study is based on four datasets with varying data distributions. The results indicate that the performance of FL is highly dependent on the data and its distribution among clients. In some scenarios, FL can be an effective alternative to traditional central or local training methods. Additionally, we introduce a new federated learning dataset from a real-world quality inspection setting.

Abstract: Recent advances in many domains require more and more complicated experiment design. Such complicated experiments often have many parameters, which necessitate parameter tuning. Tree-structured Parzen estimator (TPE), a Bayesian optimization method, is widely used in recent parameter tuning frameworks. Despite its popularity, the roles of each control parameter and the algorithm intuition have not been discussed so far. In this tutorial, we will identify the roles of each control parameter and their impacts on hyperparameter optimization using a diverse set of benchmarks. We compare our recommended setting drawn from the ablation study with baseline methods and demonstrate that our recommended setting improves the performance of TPE. Our TPE implementation is available at https://github.com/nabenabe0928/tpe/tree/single-opt.

Abstract: We propose an evolution strategies-based algorithm for estimating gradients in unrolled computation graphs, called ES-Single. Similarly to the recently-proposed Persistent Evolution Strategies (PES), ES-Single is unbiased, and overcomes chaos arising from recursive function applications by smoothing the meta-loss landscape. ES-Single samples a single perturbation per particle, that is kept fixed over the course of an inner problem (e.g., perturbations are not re-sampled for each partial unroll). Compared to PES, ES-Single is simpler to implement and has lower variance: the variance of ES-Single is constant with respect to the number of truncated unrolls, removing a key barrier in applying ES to long inner problems using short truncations. We show that ES-Single is unbiased for quadratic inner problems, and demonstrate empirically that its variance can be substantially lower than that of PES. ES-Single consistently outperforms PES on a variety of tasks, including a synthetic benchmark task, hyperparameter optimization, training recurrent neural networks, and training learned optimizers.