Deep neural networks (DNNs) have shown remarkable performance in a wide range of tasks over the past decade (Bengio et al. 2021). A mystery is that they generalize surprisingly well on unseen data, though having far more trainable parameters than the number of training examples (Belkin et al. 2019, Li et al. 2023). This phenomenon of benign overfitting inevitably casts shadows on the classical theory of statistical learning, which posits that models with high complexity tend to overfit the training data, whereas models with low complexity tend to underfit the training data. To reconcile the conflicts, some researchers argue that this is due to the regularization incurred during training, either implicitly imposed via use of stochastic gradient descent (SGD) (Advani et al. 2020, Barrett & Dherin 2021, Smith et al. 2021, Sclocchi & Wyart 2024) or explicitly via batch normalization (Ioffe & Szegedy 2015), weight decay (Krogh & Hertz 1992), dropout (Srivastava et al. 2014), etc. However, Zhang et al. (2017) questioned this widely received wisdom because they found that DNNs are still able to achieve zero training error with randomly labeled examples, which apparently cannot generalize. Prior to our work, there has been extensive study trying to explain the generalization behavior of DNNs and they roughly can be categorized into the following classes. The first class is the so-called norm-based bounds (Neyshabur et al. 2015, Bartlett et al. 2017, Neyshabur et al. 2018, Golowich et al. 2018) that are composed of the operator norm of layerwise weight matrices. However, recent studies suggest that these norm-based bounds might be problematic as they abnormally increase with the number of training examples (Nagarajan & Kolter 2019). Moreover, norm-based bounds are numerically vacuous as they are even several orders of magnitude larger than the number of network parameters.

Partial-Label Learning (PLL) is a typical problem of weakly supervised learning, where each training instance is annotated with a set of candidate labels. Self-training PLL models achieve state-of-the-art performance but suffer from error accumulation problem caused by mistakenly disambiguated instances. Although co-training can alleviate this issue by training two networks simultaneously and allowing them to interact with each other, most existing co-training methods train two structurally identical networks with the same task, i.e., are symmetric, rendering it insufficient for them to correct each other due to their similar limitations. Therefore, in this paper, we propose an asymmetric dual-task co-training PLL model called AsyCo, which forces its two networks, i.e., a disambiguation network and an auxiliary network, to learn from different views explicitly by optimizing distinct tasks. Specifically, the disambiguation network is trained with self-training PLL task to learn label confidence, while the auxiliary network is trained in a supervised learning paradigm to learn from the noisy pairwise similarity labels that are constructed according to the learned label confidence. Finally, the error accumulation problem is mitigated via information distillation and confidence refinement. Extensive experiments on both uniform and instance-dependent partially labeled datasets demonstrate the effectiveness of AsyCo. The code is available at https://github.com/libeibeics/AsyCo.

Self-supervised learning (SSL) models usually require weeks of pre-training with dozens of high-end GPUs. These models typically have a multi-headed self-attention (MHSA) context encoder. However, MHSA takes quadratic time and space in the input length, contributing to the high pre-training cost. Linear-complexity alternatives to MHSA have been proposed. For instance, in supervised training, the SummaryMixing model is the first to outperform MHSA across multiple speech processing tasks. However, these cheaper alternatives have not been explored for SSL yet. This paper studies a linear-complexity context encoder for SSL for the first time. With better or equivalent performance for the downstream tasks of the MP3S benchmark, SummaryMixing reduces the pre-training time and peak VRAM of wav2vec 2.0 model by 18% and by 23%, respectively, leading to the pre-training of a 155M wav2vec 2.0 model finished within one week with 4 Tesla A100 GPUs. Code is available at https://github.com/SamsungLabs/SummaryMixing.

Research on token-level reference-free hallucination detection has predominantly focused on English, primarily due to the scarcity of robust datasets in other languages. This has hindered systematic investigations into the effectiveness of cross-lingual transfer for this important NLP application. To address this gap, we introduce ANHALTEN, a new evaluation dataset that extends the English hallucination detection dataset to German. To the best of our knowledge, this is the first work that explores cross-lingual transfer for token-level reference-free hallucination detection. ANHALTEN contains gold annotations in German that are parallel (i.e., directly comparable to the original English instances). We benchmark several prominent cross-lingual transfer approaches, demonstrating that larger context length leads to better hallucination detection in German, even without succeeding context. Importantly, we show that the sample-efficient few-shot transfer is the most effective approach in most setups. This highlights the practical benefits of minimal annotation effort in the target language for reference-free hallucination detection. Aiming to catalyze future research on cross-lingual token-level reference-free hallucination detection, we make ANHALTEN publicly available: https://github.com/janekh24/anhalten

Dynamic Algorithm Configuration (DAC) addresses the challenge of dynamically setting hyperparameters of an algorithm for a diverse set of instances rather than focusing solely on individual tasks. Agents trained with Deep Reinforcement Learning (RL) offer a pathway to solve such settings. However, the limited generalization performance of these agents has significantly hindered the application in DAC. Our hypothesis is that a potential bias in the training instances limits generalization capabilities. We take a step towards mitigating this by selecting a representative subset of training instances to overcome overrepresentation and then retraining the agent on this subset to improve its generalization performance. For constructing the meta-features for the subset selection, we particularly account for the dynamic nature of the RL agent by computing time series features on trajectories of actions and rewards generated by the agent's interaction with the environment. Through empirical evaluations on the Sigmoid and CMA-ES benchmarks from the standard benchmark library for DAC, called DACBench, we discuss the potentials of our selection technique compared to training on the entire instance set. Our results highlight the efficacy of instance selection in refining DAC policies for diverse instance spaces.

Padding is often used in tuning LLM models by adding special tokens to shorter training examples to match the length of the longest sequence in each batch. While this ensures uniformity for batch processing, it introduces inefficiencies by including irrelevant padding tokens in the computation and wastes GPU resources. On the other hand, the Hugging Face SFT trainer offers the option to use packing to combine multiple training examples up to the maximum sequence length. This allows for maximal utilization of GPU resources. However, without proper masking of each packed training example, attention will not be computed correctly when using SFT trainer. We enable and then analyse packing and Flash Attention with proper attention masking of each example and show the benefits of this training paradigm.

Credit card fraud has significant implications at both an individual and societal level, making effective prevention essential. Current methods rely heavily on feature engineering and labeled information, both of which have significant limitations. In this work, we present GraphGuard, a novel contrastive self-supervised graph-based framework for detecting fraudulent credit card transactions. We conduct experiments on a real-world dataset and a synthetic dataset. Our results provide a promising initial direction for exploring the effectiveness of graph-based self-supervised approaches for credit card fraud detection.

Recent advances in measuring hardness-wise properties of data guide language models in sample selection within low-resource scenarios. However, class-specific properties are overlooked for task setup and learning. How will these properties influence model learning and is it generalizable across datasets? To answer this question, this work formally initiates the concept of $\textit{class-wise hardness}$. Experiments across eight natural language understanding (NLU) datasets demonstrate a consistent hardness distribution across learning paradigms, models, and human judgment. Subsequent experiments unveil a notable challenge in measuring such class-wise hardness with instance-level metrics in previous works. To address this, we propose $\textit{GeoHard}$ for class-wise hardness measurement by modeling class geometry in the semantic embedding space. $\textit{GeoHard}$ surpasses instance-level metrics by over 59 percent on $\textit{Pearson}$'s correlation on measuring class-wise hardness. Our analysis theoretically and empirically underscores the generality of $\textit{GeoHard}$ as a fresh perspective on data diagnosis. Additionally, we showcase how understanding class-wise hardness can practically aid in improving task learning.

Knowledge distillation (KD) is an effective method for transferring knowledge from a large, well-trained teacher model to a smaller, more efficient student model. Despite its success, one of the main challenges in KD is ensuring the efficient transfer of complex knowledge while maintaining the student's computational efficiency. Unlike previous works that applied contrastive objectives promoting explicit negative instances, we introduce Relational Representation Distillation (RRD). Our approach leverages pairwise similarities to explore and reinforce the relationships between the teacher and student models. Inspired by self-supervised learning principles, it uses a relaxed contrastive loss that focuses on similarity rather than exact replication. This method aligns the output distributions of teacher samples in a large memory buffer, improving the robustness and performance of the student model without the need for strict negative instance differentiation. Our approach demonstrates superior performance on CIFAR-100, outperforming traditional KD techniques and surpassing 13 state-of-the-art methods. It also transfers successfully to other datasets like Tiny ImageNet and STL-10. The code will be made public soon.

In open-set semi-supervised learning (OSSL), we consider unlabeled datasets that may contain unknown classes. Existing OSSL methods often use the softmax confidence for classifying data as in-distribution (ID) or out-of-distribution (OOD). Additionally, many works for OSSL rely on ad-hoc thresholds for ID/OOD classification, without considering the statistics of the problem. We propose a new score for ID/OOD classification based on angles in feature space between data and an ID subspace. Moreover, we propose an approach to estimate the conditional distributions of scores given ID or OOD data, enabling probabilistic predictions of data being ID or OOD. These components are put together in a framework for OSSL, termed \emph{ProSub}, that is experimentally shown to reach SOTA performance on several benchmark problems. Our code is available at https://github.com/walline/prosub.