Adam has been shown to outperform gradient descent on large language models by a larger margin than on other tasks, but it is unclear why. We show that a key factor in this performance gap is the heavy-tailed class imbalance found in language tasks.

Indigenous communities face ongoing challenges in preserving their cultural heritage, particularly in the face of systemic marginalization and urban development. In Brazil, the Museu Nacional dos Povos Indigenas through the Tainacan platform hosts the country's largest online collection of Indigenous objects and iconographies, providing a critical resource for cultural engagement. Using publicly available data from this repository, we present a data-driven initiative that applies artificial intelligence to enhance accessibility, interpretation, and exploration. We develop two semantic pipelines: a visual pipeline that models image-based similarity and a textual pipeline that captures semantic relationships from item descriptions. These embedding spaces are projected into two dimensions and integrated into an interactive visualization tool we also developed. In addition to similarity-based navigation, users can explore the collection through temporal and geographic lenses, enabling both semantic and contextualized perspectives. The system supports curatorial tasks, aids public engagement, and reveals latent connections within the collection. This work demonstrates how AI can ethically contribute to cultural preservation practices.

Learning-based solutions for long-tailed recognition face difficulties in generalizing on balanced test datasets. Due to imbalanced data prior, the learned \textit{a posteriori} distribution is biased toward the most frequent (head) classes, leading to an inferior performance on the least frequent (tail) classes. In general, the performance can be improved by removing such a bias by eliminating the effect of imbalanced prior modeled using the number of class samples (frequencies). We first observe that the \textit{effective prior} on the classes, learned by the model at the end of the training, can differ from the empirical prior obtained using class frequencies. Thus, we propose a novel approach to accurately model the effective prior of a trained model using \textit{a posteriori} probabilities. We propose to correct the imbalanced prior by adjusting the predicted \textit{a posteriori} probabilities (Prior2Posterior: P2P) using the calculated prior in a post-hoc manner after the training, and show that it can result in improved model performance. We present theoretical analysis showing the optimality of our approach for models trained with naive cross-entropy loss as well as logit adjusted loss. Our experiments show that the proposed approach achieves new state-of-the-art (SOTA) on several benchmark datasets from the long-tail literature in the category of logit adjustment methods. Further, the proposed approach can be used to inspect any existing method to capture the \textit{effective prior} and remove any residual bias to improve its performance, post-hoc, without model retraining. We also show that by using the proposed post-hoc approach, the performance of many existing methods can be improved further.

Severe data imbalance naturally exists among web-scale vision-language datasets. Despite this, we find CLIP pre-trained thereupon exhibits notable robustness to the data imbalance compared to supervised learning, and demonstrates significant effectiveness in learning generalizable representations. With an aim to investigate the reasons behind this finding, we conduct controlled experiments to study various underlying factors, and reveal that CLIP's pretext task forms a dynamic classification problem wherein only a subset of classes is present in training. This isolates the bias from dominant classes and implicitly balances the learning signal. Furthermore, the robustness and discriminability of CLIP improve with more descriptive language supervision, larger data scale, and broader open-world concepts, which are inaccessible to supervised learning. Our study not only uncovers the mechanisms behind CLIP's generalizability beyond data imbalance but also provides transferable insights for the research community. The findings are validated in both supervised and self-supervised learning, enabling models trained on imbalanced data to achieve CLIP-level performance on diverse recognition tasks.

Labeling data via rules-of-thumb and minimal label supervision is central to Weak Supervision, a paradigm subsuming subareas of machine learning such as crowdsourced learning and semi-supervised ensemble learning. By using this labeled data to train modern machine learning methods, the cost of acquiring large amounts of hand labeled data can be ameliorated. Approaches to combining the rules-of-thumb falls into two camps, reflecting different ideologies of statistical estimation. The most common approach, exemplified by the Dawid-Skene model, is based on probabilistic modeling. The other, developed in the work of Balsubramani-Freund and others, is adversarial and game-theoretic. We provide a variety of statistical results for the adversarial approach under log-loss: we characterize the form of the solution, relate it to logistic regression, demonstrate consistency, and give rates of convergence. On the other hand, we find that probabilistic approaches for the same model class can fail to be consistent. Experimental results are provided to corroborate the theoretical results.

Adam has been shown to outperform gradient descent in optimizing large language transformers empirically, and by a larger margin than on other tasks, but it is unclear why this happens. We show that the heavy-tailed class imbalance found in language modeling tasks leads to difficulties in the optimization dynamics. When training with gradient descent, the loss associated with infrequent words decreases slower than the loss associated with frequent ones. As most samples come from relatively infrequent words, the average loss decreases slowly with gradient descent. On the other hand, Adam and sign-based methods do not suffer from this problem and improve predictions on all classes. To establish that this behavior is indeed caused by class imbalance, we show empirically that it persist through different architectures and data types, on language transformers, vision CNNs, and linear models. We further study this phenomenon on a linear classification with cross-entropy loss, showing that heavy-tailed class imbalance leads to ill-conditioning, and that the normalization used by Adam can counteract it.

Integrating supervised contrastive loss to cross entropy-based communication has recently been proposed as a solution to address the long-tail learning problem. However, when the class imbalance ratio is high, it requires adjusting the supervised contrastive loss to support the tail classes, as the conventional contrastive learning is biased towards head classes by default. To this end, we present Rebalanced Contrastive Learning (RCL), an efficient means to increase the long tail classification accuracy by addressing three main aspects: 1. Feature space balancedness - Equal division of the feature space among all the classes, 2. Intra-Class compactness - Reducing the distance between same-class embeddings, 3. Regularization - Enforcing larger margins for tail classes to reduce overfitting. RCL adopts class frequency-based SoftMax loss balancing to supervised contrastive learning loss and exploits scalar multiplied features fed to the contrastive learning loss to enforce compactness. We implement RCL on the Balanced Contrastive Learning (BCL) Framework, which has the SOTA performance. Our experiments on three benchmark datasets demonstrate the richness of the learnt embeddings and increased top-1 balanced accuracy RCL provides to the BCL framework. We further demonstrate that the performance of RCL as a standalone loss also achieves state-of-the-art level accuracy.

Pruning has emerged as a powerful technique for compressing deep neural networks, reducing memory usage and inference time without significantly affecting overall performance. However, the nuanced ways in which pruning impacts model behavior are not well understood, particularly for long-tailed, multi-label datasets commonly found in clinical settings. This knowledge gap could have dangerous implications when deploying a pruned model for diagnosis, where unexpected model behavior could impact patient well-being. To fill this gap, we perform the first analysis of pruning's effect on neural networks trained to diagnose thorax diseases from chest X-rays (CXRs). On two large CXR datasets, we examine which diseases are most affected by pruning and characterize class "forgettability" based on disease frequency and co-occurrence behavior. Further, we identify individual CXRs where uncompressed and heavily pruned models disagree, known as pruning-identified exemplars (PIEs), and conduct a human reader study to evaluate their unifying qualities. We find that radiologists perceive PIEs as having more label noise, lower image quality, and higher diagnosis difficulty. This work represents a first step toward understanding the impact of pruning on model behavior in deep long-tailed, multi-label medical image classification.