Deep neural networks demonstrate strong performance under aligned training-test distributions. However, real-world test data often exhibit domain shifts. Test-Time Adaptation (TTA) addresses this challenge by adapting the model to test data during inference. While most TTA studies assume that the training and test data share the same class set (closed-set TTA), real-world scenarios often involve open-set data (open-set TTA), which can degrade closed-set accuracy. A recent study showed that identifying open-set data during adaptation and maximizing its entropy is an effective solution. However, the previous method relies on the source model for filtering, resulting in suboptimal filtering accuracy on domain-shifted test data. In contrast, we found that the adapting model, which learns domain knowledge from noisy test streams, tends to be unstable and leads to error accumulation when used for filtering. To address this problem, we propose Primary-Auxiliary Filtering (PAF), which employs an auxiliary filter to validate data filtered by the primary filter. Furthermore, we propose Knowledge-Integrated Prediction (KIP), which calibrates the outputs of the adapting model, EMA model, and source model to integrate their complementary knowledge for OSTTA. We validate our approach across diverse closed-set and open-set datasets. Our method enhances both closed-set accuracy and open-set discrimination over existing methods. The code is available at https://github.com/powerpowe/PAF-KIP-OSTTA .

Street Scene Semantic Understanding (denoted as S3U) is a crucial but complex task for autonomous driving (AD) vehicles. Their inference models typically face poor generalization due to domain-shift. Federated Learning (FL) has emerged as a promising paradigm for enhancing the generalization of AD models through privacy-preserving distributed learning. However, these FL AD models face significant temporal catastrophic forgetting when deployed in dynamically evolving environments, where continuous adaptation causes abrupt erosion of historical knowledge. This paper proposes Federated Exponential Moving Average (FedEMA), a novel framework that addresses this challenge through two integral innovations: (I) Server-side model's historical fitting capability preservation via fusing current FL round's aggregation model and a proposed previous FL round's exponential moving average (EMA) model; (II) Vehicle-side negative entropy regularization to prevent FL models' possible overfitting to EMA-introduced temporal patterns. Above two strategies empower FedEMA a dual-objective optimization that balances model generalization and adaptability. In addition, we conduct theoretical convergence analysis for the proposed FedEMA. Extensive experiments both on Cityscapes dataset and Camvid dataset demonstrate FedEMA's superiority over existing approaches, showing 7.12% higher mean Intersection-over-Union (mIoU).

Weight averaging of Stochastic Gradient Descent (SGD) iterates is a popular method for training deep learning models. While it is often used as part of complex training pipelines to improve generalization or serve as a `teacher' model, weight averaging lacks proper evaluation on its own. In this work, we present a systematic study of the Exponential Moving Average (EMA) of weights. We first explore the training dynamics of EMA, give guidelines for hyperparameter tuning, and highlight its good early performance, partly explaining its success as a teacher. We also observe that EMA requires less learning rate decay compared to SGD since averaging naturally reduces noise, introducing a form of implicit regularization. Through extensive experiments, we show that EMA solutions differ from last-iterate solutions. EMA models not only generalize better but also exhibit improved i) robustness to noisy labels, ii) prediction consistency, iii) calibration and iv) transfer learning. Therefore, we suggest that an EMA of weights is a simple yet effective plug-in to improve the performance of deep learning models.

Predicting the behavior of a wireless link in terms of, e.g., the frame delivery ratio, is a critical task for optimizing the performance of wireless industrial communication systems. This is because industrial applications are typically characterized by stringent dependability and end-to-end latency requirements, which are adversely affected by channel quality degradation. In this work, we studied two neural network models for Wi-Fi link quality prediction in dense indoor environments. Experimental results show that their accuracy outperforms conventional methods based on exponential moving averages, due to their ability to capture complex patterns about communications, including the effects of shadowing and multipath propagation, which are particularly pronounced in industrial scenarios. This highlights the potential of neural networks for predicting spectrum behavior in challenging operating conditions, and suggests that they can be exploited to improve determinism and dependability of wireless communications, fostering their adoption in the industry.

Traditional language model alignment methods, such as Direct Preference Optimization (DPO), are limited by their dependence on static, pre-collected paired preference data, which hampers their adaptability and practical applicability. To overcome this limitation, we introduce Self-Augmented Preference Optimization (SAPO), an effective and scalable training paradigm that does not require existing paired data. Building on the self-play concept, which autonomously generates negative responses, we further incorporate an off-policy learning pipeline to enhance data exploration and exploitation. Specifically, we employ an Exponential Moving Average (EMA) model in conjunction with a replay buffer to enable dynamic updates of response segments, effectively integrating real-time feedback with insights from historical data. Our comprehensive evaluations of the LLaMA3-8B and Mistral-7B models across benchmarks, including the Open LLM Leaderboard, IFEval, AlpacaEval 2.0, and MT-Bench, demonstrate that SAPO matches or surpasses established offline contrastive baselines, such as DPO and Odds Ratio Preference Optimization, and outperforms offline self-play methods like SPIN. Our code is available at https://github.com/yinyueqin/SAPO

A common phenomena confining the representation quality in Self-Supervised Learning (SSL) is dimensional collapse (also known as rank degeneration), where the learned representations are mapped to a low dimensional subspace of the representation space. The State-of-the-Art SSL methods have shown to suffer from dimensional collapse and fall behind maintaining full rank. Recent approaches to prevent this problem have proposed using contrastive losses, regularization techniques, or architectural tricks. We propose WERank, a new regularizer on the weight parameters of the network to prevent rank degeneration at different layers of the network. We provide empirical evidence and mathematical justification to demonstrate the effectiveness of the proposed regularization method in preventing dimensional collapse. We verify the impact of WERank on graph SSL where dimensional collapse is more pronounced due to the lack of proper data augmentation. We empirically demonstrate that WERank is effective in helping BYOL to achieve higher rank during SSL pre-training and consequently downstream accuracy during evaluation probing. Ablation studies and experimental analysis shed lights on the underlying factors behind the performance gains of the proposed approach.

The ability to predict the behavior of a wireless channel in terms of the frame delivery ratio is quite valuable, and permits, e.g., to optimize the operating parameters of a wireless network at runtime, or to proactively react to the degradation of the channel quality, in order to meet the stringent requirements about dependability and end-to-end latency that typically characterize industrial applications. In this work, prediction models based on the exponential moving average (EMA) are investigated in depth, which are proven to outperform other simple statistical methods and whose performance is nearly as good as artificial neural networks, but with dramatically lower computational requirements. Regarding the innovation and motivation of this work, a new model that we called EMA linear combination (ELC), is introduced, explained, and evaluated experimentally. Its prediction accuracy, tested on some databases acquired from a real setup based on Wi-Fi devices, showed that ELC brings tangible improvements over EMA in any experimental conditions, the only drawback being a slight increase in computational complexity.

Neural networks are very effective when trained on large datasets for a large number of iterations. However, when they are trained on non-stationary streams of data and in an online fashion, their performance is reduced (1) by the online setup, which limits the availability of data, (2) due to catastrophic forgetting because of the non-stationary nature of the data. Furthermore, several recent works (Caccia et al., 2022; Lange et al., 2023) showed that replay methods used in continual learning suffer from the stability gap, encountered when evaluating the model continually (rather than only on task boundaries). In this article, we study the effect of model ensembling as a way to improve performance and stability in online continual learning. We notice that naively ensembling models coming from a variety of training tasks increases the performance in online continual learning considerably. Starting from this observation, and drawing inspirations from semi-supervised learning ensembling methods, we use a lightweight temporal ensemble that computes the exponential moving average of the weights (EMA) at test time, and show that it can drastically increase the performance and stability when used in combination with several methods from the literature. Learning neural networks with backpropagation has been proven capable of good generalization properties even when using overparametrized networks (Krizhevsky et al., 2017). However, these good learning properties mainly occur when the data is provided in an independant and identically distributed manner. When learning on a stream which distribution varies over time, neural networks are known to suffer from catastrophic forgetting (McCloskey & Cohen, 1989; Goodfellow et al., 2014; Kirkpatrick et al., 2017), and tend to forget knowledge acquired in previous learning tasks. The field of continual learning aims to address this problem. Generally, incremental learning separates the learning into distinct tasks (identified by a task-ID) that are encountered sequentially by the agent. A variety of settings have been introduced in continual learning in order to evaluate several aspects of the continual learning agent; taskincremental learning (De Lange et al., 2021; van de Ven & Tolias, 2018), and class-incremental learning (Masana et al., 2022; Belouadah et al., 2021) are among the most popular. In this paper, we focus on the more challenging class-incremental setting, where the learner does not have access to the task-ID at inference time.

The ability to reliably predict the future quality of a wireless channel, as seen by the media access control layer, is a key enabler to improve performance of future industrial networks that do not rely on wires. Knowing in advance how much channel behavior may change can speed up procedures for adaptively selecting the best channel, making the network more deterministic, reliable, and less energy-hungry, possibly improving device roaming capabilities at the same time. To this aim, popular approaches based on moving averages and regression were compared, using multiple key performance indicators, on data captured from a real Wi-Fi setup. Moreover, a simple technique based on a linear combination of outcomes from different techniques was presented and analyzed, to further reduce the prediction error, and some considerations about lower bounds on achievable errors have been reported. We found that the best model is the exponential moving average, which managed to predict the frame delivery ratio with a 2.10\% average error and, at the same time, has lower computational complexity and memory consumption than the other models we analyzed.