Machine Learning (cs.LG)

Abstract: Multicalibration is a notion of fairness that aims to provide accurate predictions across a large set of groups. Multicalibration is known to be a different goal than loss minimization, even for simple predictors such as linear functions. In this note, we show that for (almost all) large neural network sizes, optimally minimizing squared error leads to multicalibration. Our results are about representational aspects of neural networks, and not about algorithmic or sample complexity considerations. Previous such results were known only for predictors that were nearly Bayes-optimal and were therefore representation independent. We emphasize that our results do not apply to specific algorithms for optimizing neural networks, such as SGD, and they should not be interpreted as "fairness comes for free from optimizing neural networks".

Abstract: Decoupling representation learning and classifier learning has been shown to be effective in classification with long-tailed data. There are two main ingredients in constructing a decoupled learning scheme; 1) how to train the feature extractor for representation learning so that it provides generalizable representations and 2) how to re-train the classifier that constructs proper decision boundaries by handling class imbalances in long-tailed data. In this work, we first apply Stochastic Weight Averaging (SWA), an optimization technique for improving the generalization of deep neural networks, to obtain better generalizing feature extractors for long-tailed classification. We then propose a novel classifier re-training algorithm based on stochastic representation obtained from the SWA-Gaussian, a Gaussian perturbed SWA, and a self-distillation strategy that can harness the diverse stochastic representations based on uncertainty estimates to build more robust classifiers. Extensive experiments on CIFAR10/100-LT, ImageNet-LT, and iNaturalist-2018 benchmarks show that our proposed method improves upon previous methods both in terms of prediction accuracy and uncertainty estimation.

Abstract: A Neural Process (NP) estimates a stochastic process implicitly defined with neural networks given a stream of data, rather than pre-specifying priors already known, such as Gaussian processes. An ideal NP would learn everything from data without any inductive biases, but in practice, we often restrict the class of stochastic processes for the ease of estimation. One such restriction is the use of a finite-dimensional latent variable accounting for the uncertainty in the functions drawn from NPs. Some recent works show that this can be improved with more "data-driven" source of uncertainty such as bootstrapping. In this work, we take a different approach based on the martingale posterior, a recently developed alternative to Bayesian inference. For the martingale posterior, instead of specifying prior-likelihood pairs, a predictive distribution for future data is specified. Under specific conditions on the predictive distribution, it can be shown that the uncertainty in the generated future data actually corresponds to the uncertainty of the implicitly defined Bayesian posteriors. Based on this result, instead of assuming any form of the latent variables, we equip a NP with a predictive distribution implicitly defined with neural networks and use the corresponding martingale posteriors as the source of uncertainty. The resulting model, which we name as Martingale Posterior Neural Process (MPNP), is demonstrated to outperform baselines on various tasks.

Abstract: Embodied control requires agents to leverage multi-modal pre-training to quickly learn how to act in new environments, where video demonstrations contain visual and motion details needed for low-level perception and control, and language instructions support generalization with abstract, symbolic structures. While recent approaches apply contrastive learning to force alignment between the two modalities, we hypothesize better modeling their complementary differences can lead to more holistic representations for downstream adaption. To this end, we propose Emergent Communication for Embodied Control (EC^2), a novel scheme to pre-train video-language representations for few-shot embodied control. The key idea is to learn an unsupervised "language" of videos via emergent communication, which bridges the semantics of video details and structures of natural language. We learn embodied representations of video trajectories, emergent language, and natural language using a language model, which is then used to finetune a lightweight policy network for downstream control. Through extensive experiments in Metaworld and Franka Kitchen embodied benchmarks, EC^2 is shown to consistently outperform previous contrastive learning methods for both videos and texts as task inputs. Further ablations confirm the importance of the emergent language, which is beneficial for both video and language learning, and significantly superior to using pre-trained video captions. We also present a quantitative and qualitative analysis of the emergent language and discuss future directions toward better understanding and leveraging emergent communication in embodied tasks.

Abstract: We discuss the discontinuities that arise when mapping unordered objects to neural network outputs of fixed permutation, referred to as the responsibility problem. Prior work has proved the existence of the issue by identifying a single discontinuity. Here, we show that discontinuities under such models are uncountably infinite, motivating further research into neural networks for unordered data.

Abstract: Split learning of deep neural networks (SplitNN) has provided a promising solution to learning jointly for the mutual interest of a guest and a host, which may come from different backgrounds, holding features partitioned vertically. However, SplitNN creates a new attack surface for the adversarial participant, holding back its practical use in the real world. By investigating the adversarial effects of highly threatening attacks, including property inference, data reconstruction, and feature hijacking attacks, we identify the underlying vulnerability of SplitNN and propose a countermeasure. To prevent potential threats and ensure the learning guarantees of SplitNN, we design a privacy-preserving tunnel for information exchange between the guest and the host. The intuition is to perturb the propagation of knowledge in each direction with a controllable unified solution. To this end, we propose a new activation function named R3eLU, transferring private smashed data and partial loss into randomized responses in forward and backward propagations, respectively. We give the first attempt to secure split learning against three threatening attacks and present a fine-grained privacy budget allocation scheme. The analysis proves that our privacy-preserving SplitNN solution provides a tight privacy budget, while the experimental results show that our solution performs better than existing solutions in most cases and achieves a good tradeoff between defense and model usability.

Abstract: Supervised Deep Learning (DL) models are currently the leading approach for sensor-based Human Activity Recognition (HAR) on wearable and mobile devices. However, training them requires large amounts of labeled data whose collection is often time-consuming, expensive, and error-prone. At the same time, due to the intra- and inter-variability of activity execution, activity models should be personalized for each user. In this work, we propose SelfAct: a novel framework for HAR combining self-supervised and active learning to mitigate these problems. SelfAct leverages a large pool of unlabeled data collected from many users to pre-train through self-supervision a DL model, with the goal of learning a meaningful and efficient latent representation of sensor data. The resulting pre-trained model can be locally used by new users, which will fine-tune it thanks to a novel unsupervised active learning strategy. Our experiments on two publicly available HAR datasets demonstrate that SelfAct achieves results that are close to or even better than the ones of fully supervised approaches with a small number of active learning queries.

Abstract: Decadal temperature prediction provides crucial information for quantifying the expected effects of future climate changes and thus informs strategic planning and decision-making in various domains. However, such long-term predictions are extremely challenging, due to the chaotic nature of temperature variations. Moreover, the usefulness of existing simulation-based and machine learning-based methods for this task is limited because initial simulation or prediction errors increase exponentially over time. To address this challenging task, we devise a novel prediction method involving an information tracking mechanism that aims to track and adapt to changes in temperature dynamics during the prediction phase by providing probabilistic feedback on the prediction error of the next step based on the current prediction. We integrate this information tracking mechanism, which can be considered as a model calibrator, into the objective function of our method to obtain the corrections needed to avoid error accumulation. Our results show the ability of our method to accurately predict global land-surface temperatures over a decadal range. Furthermore, we demonstrate that our results are meaningful in a real-world context: the temperatures predicted using our method are consistent with and can be used to explain the well-known teleconnections within and between different continents.

Abstract: This paper proposes an exploration technique for multi-agent reinforcement learning (MARL) with graph-based communication among agents. We assume the individual rewards received by the agents are independent of the actions by the other agents, while their policies are coupled. In the proposed framework, neighbouring agents collaborate to estimate the uncertainty about the state-action space in order to execute more efficient explorative behaviour. Different from existing works, the proposed algorithm does not require counting mechanisms and can be applied to continuous-state environments without requiring complex conversion techniques. Moreover, the proposed scheme allows agents to communicate in a fully decentralized manner with minimal information exchange. And for continuous-state scenarios, each agent needs to exchange only a single parameter vector. The performance of the algorithm is verified with theoretical results for discrete-state scenarios and with experiments for continuous ones.

Abstract: The quality of air is closely linked with the life quality of humans, plantations, and wildlife. It needs to be monitored and preserved continuously. Transportations, industries, construction sites, generators, fireworks, and waste burning have a major percentage in degrading the air quality. These sources are required to be used in a safe and controlled manner. Using traditional laboratory analysis or installing bulk and expensive models every few miles is no longer efficient. Smart devices are needed for collecting and analyzing air data. The quality of air depends on various factors, including location, traffic, and time. Recent researches are using machine learning algorithms, big data technologies, and the Internet of Things to propose a stable and efficient model for the stated purpose. This review paper focuses on studying and compiling recent research in this field and emphasizes the Data sources, Monitoring, and Forecasting models. The main objective of this paper is to provide the astuteness of the researches happening to improve the various aspects of air polluting models. Further, it casts light on the various research issues and challenges also.

Abstract: Fine-tuning pre-trained models has recently yielded remarkable performance gains in graph neural networks (GNNs). In addition to pre-training techniques, inspired by the latest work in the natural language fields, more recent work has shifted towards applying effective fine-tuning approaches, such as parameter-efficient tuning (delta tuning). However, given the substantial differences between GNNs and transformer-based models, applying such approaches directly to GNNs proved to be less effective. In this paper, we present a comprehensive comparison of delta tuning techniques for GNNs and propose a novel delta tuning method specifically designed for GNNs, called AdapterGNN. AdapterGNN preserves the knowledge of the large pre-trained model and leverages highly expressive adapters for GNNs, which can adapt to downstream tasks effectively with only a few parameters, while also improving the model's generalization ability on the downstream tasks. Extensive experiments show that AdapterGNN achieves higher evaluation performance (outperforming full fine-tuning by 1.4% and 5.5% in the chemistry and biology domains respectively, with only 5% of its parameters tuned) and lower generalization gaps compared to full fine-tuning. Moreover, we empirically show that a larger GNN model can have a worse generalization ability, which differs from the trend observed in large language models. We have also provided a theoretical justification for delta tuning can improve the generalization ability of GNNs by applying generalization bounds.

Abstract: This paper investigates the energy savings that near-subthreshold processors can obtain in edge AI applications and proposes strategies to improve them while maintaining the accuracy of the application. The selected processors deploy adaptive voltage scaling techniques in which the frequency and voltage levels of the processor core are determined at the run-time. In these systems, embedded RAM and flash memory size is typically limited to less than 1 megabyte to save power. This limited memory imposes restrictions on the complexity of the neural networks model that can be mapped to these devices and the required trade-offs between accuracy and battery life. To address these issues, we propose and evaluate alternative 'big-little' neural network strategies to improve battery life while maintaining prediction accuracy. The strategies are applied to a human activity recognition application selected as a demonstrator that shows that compared to the original network, the best configurations obtain an energy reduction measured at 80% while maintaining the original level of inference accuracy.

Abstract: Group fairness is achieved by equalising prediction distributions between protected sub-populations; individual fairness requires treating similar individuals alike. These two objectives, however, are incompatible when a scoring model is calibrated through discontinuous probability functions, where individuals can be randomly assigned an outcome determined by a fixed probability. This procedure may provide two similar individuals from the same protected group with classification odds that are disparately different -- a clear violation of individual fairness. Assigning unique odds to each protected sub-population may also prevent members of one sub-population from ever receiving equal chances of a positive outcome to another, which we argue is another type of unfairness called individual odds. We reconcile all this by constructing continuous probability functions between group thresholds that are constrained by their Lipschitz constant. Our solution preserves the model's predictive power, individual fairness and robustness while ensuring group fairness.

Abstract: Mining data streams is one of the main studies in machine learning area due to its application in many knowledge areas. One of the major challenges on mining data streams is concept drift, which requires the learner to discard the current concept and adapt to a new one. Ensemble-based drift detection algorithms have been used successfully to the classification task but usually maintain a fixed size ensemble of learners running the risk of needlessly spending processing time and memory. In this paper we present improvements to the Scale-free Network Regressor (SFNR), a dynamic ensemble-based method for regression that employs social networks theory. In order to detect concept drifts SFNR uses the Adaptive Window (ADWIN) algorithm. Results show improvements in accuracy, especially in concept drift situations and better performance compared to other state-of-the-art algorithms in both real and synthetic data.

Abstract: Model-based neural networks provide unparalleled performance for various tasks, such as sparse coding and compressed sensing problems. Due to the strong connection with the sensing model, these networks are interpretable and inherit prior structure of the problem. In practice, model-based neural networks exhibit higher generalization capability compared to ReLU neural networks. However, this phenomenon was not addressed theoretically. Here, we leverage complexity measures including the global and local Rademacher complexities, in order to provide upper bounds on the generalization and estimation errors of model-based networks. We show that the generalization abilities of model-based networks for sparse recovery outperform those of regular ReLU networks, and derive practical design rules that allow to construct model-based networks with guaranteed high generalization. We demonstrate through a series of experiments that our theoretical insights shed light on a few behaviours experienced in practice, including the fact that ISTA and ADMM networks exhibit higher generalization abilities (especially for small number of training samples), compared to ReLU networks.

Abstract: Wind turbine power curve models translate ambient conditions into turbine power output. They are essential for energy yield prediction and turbine performance monitoring. In recent years, data-driven machine learning methods have outperformed parametric, physics-informed approaches. However, they are often criticised for being opaque "black boxes" which raises concerns regarding their robustness in non-stationary environments, such as faced by wind turbines. We, therefore, introduce an explainable artificial intelligence (XAI) framework to investigate and validate strategies learned by data-driven power curve models from operational SCADA data. It combines domain-specific considerations with Shapley Values and the latest findings from XAI for regression. Our results suggest, that learned strategies can be better indicators for model robustness than validation or test set errors. Moreover, we observe that highly complex, state-of-the-art ML models are prone to learn physically implausible strategies. Consequently, we compare several measures to ensure physically reasonable model behaviour. Lastly, we propose the utilization of XAI in the context of wind turbine performance monitoring, by disentangling environmental and technical effects that cause deviations from an expected turbine output. We hope, our work can guide domain experts towards training and selecting more transparent and robust data-driven wind turbine power curve models.

Abstract: Multivariate probabilistic time series forecasts are commonly evaluated via proper scoring rules, i.e., functions that are minimal in expectation for the ground-truth distribution. However, this property is not sufficient to guarantee good discrimination in the non-asymptotic regime. In this paper, we provide the first systematic finite-sample study of proper scoring rules for time-series forecasting evaluation. Through a power analysis, we identify the "region of reliability" of a scoring rule, i.e., the set of practical conditions where it can be relied on to identify forecasting errors. We carry out our analysis on a comprehensive synthetic benchmark, specifically designed to test several key discrepancies between ground-truth and forecast distributions, and we gauge the generalizability of our findings to real-world tasks with an application to an electricity production problem. Our results reveal critical shortcomings in the evaluation of multivariate probabilistic forecasts as commonly performed in the literature.

Abstract: We consider functions from the real numbers to the real numbers, output by a neural network with 1 hidden activation layer, arbitrary width, and ReLU activation function. We assume that the parameters of the neural network are chosen uniformly at random with respect to various probability distributions, and compute the expected distribution of the points of non-linearity. We use these results to explain why the network may be biased towards outputting functions with simpler geometry, and why certain functions with low information-theoretic complexity are nonetheless hard for a neural network to approximate.

Abstract: Deep reinforcement learning (RL) works impressively in some environments and fails catastrophically in others. Ideally, RL theory should be able to provide an understanding of why this is, i.e. bounds predictive of practical performance. Unfortunately, current theory does not quite have this ability. We compare standard deep RL algorithms to prior sample complexity prior bounds by introducing a new dataset, BRIDGE. It consists of 155 MDPs from common deep RL benchmarks, along with their corresponding tabular representations, which enables us to exactly compute instance-dependent bounds. We find that prior bounds do not correlate well with when deep RL succeeds vs. fails, but discover a surprising property that does. When actions with the highest Q-values under the random policy also have the highest Q-values under the optimal policy, deep RL tends to succeed; when they don't, deep RL tends to fail. We generalize this property into a new complexity measure of an MDP that we call the effective horizon, which roughly corresponds to how many steps of lookahead search are needed in order to identify the next optimal action when leaf nodes are evaluated with random rollouts. Using BRIDGE, we show that the effective horizon-based bounds are more closely reflective of the empirical performance of PPO and DQN than prior sample complexity bounds across four metrics. We also show that, unlike existing bounds, the effective horizon can predict the effects of using reward shaping or a pre-trained exploration policy.