In recent advancements in unsupervised visual representation learning, the Joint-Embedding Predictive Architecture (JEPA) has emerged as a significant method for extracting visual features from unlabeled imagery through an innovative masking strategy. Despite its success, two primary limitations have been identified: the inefficacy of Exponential Moving Average (EMA) from I-JEPA in preventing entire collapse and the inadequacy of I-JEPA prediction in accurately learning the mean of patch representations. Addressing these challenges, this study introduces a novel framework, namely C-JEPA (Contrastive-JEPA), which integrates the Image-based Joint-Embedding Predictive Architecture with the Variance-Invariance-Covariance Regularization (VICReg) strategy. This integration is designed to effectively learn the variance/covariance for preventing entire collapse and ensuring invariance in the mean of augmented views, thereby overcoming the identified limitations. Through empirical and theoretical evaluations, our work demonstrates that C-JEPA significantly enhances the stability and quality of visual representation learning. When pre-trained on the ImageNet-1K dataset, C-JEPA exhibits rapid and improved convergence in both linear probing and fine-tuning performance metrics.

Large recommendation models (LRMs) are fundamental to the multi-billion dollar online advertising industry, processing massive datasets of hundreds of billions of examples before transitioning to continuous online training to adapt to rapidly changing user behavior [1]. The massive scale of data directly impacts both computational costs and the speed at which new methods can be evaluated (R&D velocity). This paper presents actionable principles and high-level frameworks to guide practitioners in optimizing training data requirements. These strategies have been successfully deployed in Google's largest Ads CTR prediction models [1, 2] and are broadly applicable beyond LRMs. We outline the concept of data convergence, describe methods to accelerate this convergence, and finally, detail how to optimally balance training data volume with model size.

Domain-specific adaptation is critical to maximizing the performance of pre-trained language models (PLMs) on one or multiple targeted tasks, especially under resource-constrained use cases, such as edge devices. However, existing methods often struggle to balance domain-specific performance, retention of general knowledge, and efficiency for training and inference. To address these challenges, we propose Modular Domain Experts (MoDE). MoDE is a mixture-of-experts architecture that augments a general PLMs with modular, domain-specialized experts. These experts are trained independently and composed together via a lightweight training process. In contrast to standard low-rank adaptation methods, each MoDE expert consists of several transformer layers which scale better with more training examples and larger parameter counts. Our evaluation demonstrates that MoDE achieves comparable target performances to full parameter fine-tuning while achieving 1.65% better retention performance. Moreover, MoDE's architecture enables flexible sharding configurations and improves training speeds by up to 38% over state-of-the-art distributed training configurations.

Distribution shifts on graphs -- the discrepancies in data distribution between training and employing a graph machine learning model -- are ubiquitous and often unavoidable in real-world scenarios. These shifts may severely deteriorate model performance, posing significant challenges for reliable graph machine learning. Consequently, there has been a surge in research on graph machine learning under distribution shifts, aiming to train models to achieve satisfactory performance on out-of-distribution (OOD) test data. In our survey, we provide an up-to-date and forward-looking review of deep graph learning under distribution shifts. Specifically, we cover three primary scenarios: graph OOD generalization, training-time graph OOD adaptation, and test-time graph OOD adaptation. We begin by formally formulating the problems and discussing various types of distribution shifts that can affect graph learning, such as covariate shifts and concept shifts. To provide a better understanding of the literature, we systematically categorize the existing models based on our proposed taxonomy and investigate the adopted techniques behind. We also summarize commonly used datasets in this research area to facilitate further investigation. Finally, we point out promising research directions and the corresponding challenges to encourage further study in this vital domain. Additionally, we provide a continuously updated reading list at https://github.com/kaize0409/Awesome-Graph-OOD.

Differential privacy (DP) is applied when fine-tuning pre-trained large language models (LLMs) to limit leakage of training examples. While most DP research has focused on improving a model's privacy-utility tradeoff, some find that DP can be unfair to or biased against underrepresented groups. In this work, we show the impact of DP on bias in LLMs through empirical analysis. Differentially private training can increase the model bias against protected groups w.r.t AUC-based bias metrics. DP makes it more difficult for the model to differentiate between the positive and negative examples from the protected groups and other groups in the rest of the population. Our results also show that the impact of DP on bias is not only affected by the privacy protection level but also the underlying distribution of the dataset.

What kinds of and how much data is necessary for language models to induce grammatical knowledge to judge sentence acceptability? Recent language models still have much room for improvement in their data efficiency compared to humans. This paper investigates whether language models efficiently use indirect data (indirect evidence), from which they infer sentence acceptability. In contrast, humans use indirect evidence efficiently, which is considered one of the inductive biases contributing to efficient language acquisition. To explore this question, we introduce the Wug InDirect Evidence Test (WIDET), a dataset consisting of training instances inserted into the pre-training data and evaluation instances. We inject synthetic instances with newly coined wug words into pretraining data and explore the model's behavior on evaluation data that assesses grammatical acceptability regarding those words. We prepare the injected instances by varying their levels of indirectness and quantity. Our experiments surprisingly show that language models do not induce grammatical knowledge even after repeated exposure to instances with the same structure but differing only in lexical items from evaluation instances in certain language phenomena. Our findings suggest a potential direction for future research: developing models that use latent indirect evidence to induce grammatical knowledge.

Forecasting rail congestion is crucial for efficient mobility in transport systems. We present rail congestion forecasting using reports from passengers collected through a transit application. Although reports from passengers have received attention from researchers, ensuring a sufficient volume of reports is challenging due to passenger's reluctance. The limited number of reports results in the sparsity of the congestion label, which can be an issue in building a stable prediction model. To address this issue, we propose a semi-supervised method for congestion forecasting for trains, or SURCONFORT. Our key idea is twofold: firstly, we adopt semi-supervised learning to leverage sparsely labeled data and many unlabeled data. Secondly, in order to complement the unlabeled data from nearby stations, we design a railway network-oriented graph and apply the graph to semi-supervised graph regularization. Empirical experiments with actual reporting data show that SURCONFORT improved the forecasting performance by 14.9% over state-of-the-art methods under the label sparsity.

This paper demonstrates that spatial information can be used to learn interpretable representations in medical images using Self-Supervised Learning (SSL). Our proposed method, ISImed, is based on the observation that medical images exhibit a much lower variability among different images compared to classic data vision benchmarks. By leveraging this resemblance of human body structures across multiple images, we establish a self-supervised objective that creates a latent representation capable of capturing its location in the physical realm. More specifically, our method involves sampling image crops and creating a distance matrix that compares the learned representation vectors of all possible combinations of these crops to the true distance between them. The intuition is, that the learned latent space is a positional encoding for a given image crop. We hypothesize, that by learning these positional encodings, comprehensive image representations have to be generated. To test this hypothesis and evaluate our method, we compare our learned representation with two state-of-the-art SSL benchmarking methods on two publicly available medical imaging datasets. We show that our method can efficiently learn representations that capture the underlying structure of the data and can be used to transfer to a downstream classification task.

Fine-tuning Large Language Models (LLMs) typically involves updating at least a few billions of parameters. A more parameter-efficient approach is Prompt Tuning (PT), which updates only a few learnable tokens, and differently, In-Context Learning (ICL) adapts the model to a new task by simply including examples in the input without any training. When applying optimization-based methods, such as fine-tuning and PT for few-shot learning, the model is specifically adapted to the small set of training examples, whereas ICL leaves the model unchanged. This distinction makes traditional learning methods more prone to overfitting; in contrast, ICL is less sensitive to the few-shot scenario. While ICL is not prone to overfitting, it does not fully extract the information that exists in the training examples. This work introduces Context-aware Prompt Tuning (CPT), a method inspired by ICL, PT, and adversarial attacks. We build on the ICL strategy of concatenating examples before the input, but we extend this by PT-like learning, refining the context embedding through iterative optimization to extract deeper insights from the training examples. We carefully modify specific context tokens, considering the unique structure of input and output formats. Inspired by adversarial attacks, we adjust the input based on the labels present in the context, focusing on minimizing, rather than maximizing, the loss. Moreover, we apply a projected gradient descent algorithm to keep token embeddings close to their original values, under the assumption that the user-provided data is inherently valuable. Our method has been shown to achieve superior accuracy across multiple classification tasks using various LLM models.

Multivariate time-series data in fields like healthcare and industry are informative but challenging due to high dimensionality and lack of labels. Recent self-supervised learning methods excel in learning rich representations without labels but struggle with disentangled embeddings and inductive bias issues like transformation-invariance. To address these challenges, we introduce TimeDRL, a framework for multivariate time-series representation learning with dual-level disentangled embeddings. TimeDRL features: (i) disentangled timestamp-level and instance-level embeddings using a [CLS] token strategy; (ii) timestamp-predictive and instance-contrastive tasks for representation learning; and (iii) avoidance of augmentation methods to eliminate inductive biases. Experiments on forecasting and classification datasets show TimeDRL outperforms existing methods, with further validation in semi-supervised settings with limited labeled data.