We tackle the problem of zero-shot cross-lingual transfer in NLP tasks via the use of language adapters (LAs). Most of the earlier works have explored training with adapter of a single source (often English), and testing either using the target LA or LA of another related language. Training target LA requires unlabeled data, which may not be readily available for low resource unseen languages: those that are neither seen by the underlying multilingual language model (e.g., mBERT), nor do we have any (labeled or unlabeled) data for them. We posit that for more effective cross-lingual transfer, instead of just one source LA, we need to leverage LAs of multiple (linguistically or geographically related) source languages, both at train and test-time - which we investigate via our novel neural architecture, ZGUL. Extensive experimentation across four language groups, covering 15 unseen target languages, demonstrates improvements of up to 3.2 average F1 points over standard fine-tuning and other strong baselines on POS tagging and NER tasks. We also extend ZGUL to settings where either (1) some unlabeled data or (2) few-shot training examples are available for the target language. We find that ZGUL continues to outperform baselines in these settings too.

Despite its practical importance across a wide range of modalities, recent advances in self-supervised learning (SSL) have been primarily focused on a few well-curated domains, e.g., vision and language, often relying on their domain-specific knowledge. For example, Masked Auto-Encoder (MAE) has become one of the popular architectures in these domains, but less has explored its potential in other modalities. In this paper, we develop MAE as a unified, modality-agnostic SSL framework. In turn, we argue meta-learning as a key to interpreting MAE as a modality-agnostic learner, and propose enhancements to MAE from the motivation to jointly improve its SSL across diverse modalities, coined MetaMAE as a result. Our key idea is to view the mask reconstruction of MAE as a meta-learning task: masked tokens are predicted by adapting the Transformer meta-learner through the amortization of unmasked tokens. Based on this novel interpretation, we propose to integrate two advanced meta-learning techniques. First, we adapt the amortized latent of the Transformer encoder using gradient-based meta-learning to enhance the reconstruction. Then, we maximize the alignment between amortized and adapted latents through task contrastive learning which guides the Transformer encoder to better encode the task-specific knowledge. Our experiment demonstrates the superiority of MetaMAE in the modality-agnostic SSL benchmark (called DABS), significantly outperforming prior baselines. Code is available at https://github.com/alinlab/MetaMAE.

Segmentation and measurement of cardiac chambers is critical in cardiac ultrasound but is laborious and poorly reproducible. Neural networks can assist, but supervised approaches require the same laborious manual annotations. We built a pipeline for self-supervised (no manual labels) segmentation combining computer vision, clinical domain knowledge, and deep learning. We trained on 450 echocardiograms (93,000 images) and tested on 8,393 echocardiograms (4,476,266 images; mean 61 years, 51% female), using the resulting segmentations to calculate biometrics. We also tested against external images from an additional 10,030 patients with available manual tracings of the left ventricle. r2 between clinically measured and pipeline-predicted measurements were similar to reported inter-clinician variation and comparable to supervised learning across several different measurements (r2 0.56-0.84). Average accuracy for detecting abnormal chamber size and function was 0.85 (range 0.71-0.97) compared to clinical measurements. A subset of test echocardiograms (n=553) had corresponding cardiac MRIs, where MRI is the gold standard. Correlation between pipeline and MRI measurements was similar to that between clinical echocardiogram and MRI. Finally, the pipeline accurately segments the left ventricle with an average Dice score of 0.89 (95% CI [0.89]) in the external, manually labeled dataset. Our results demonstrate a manual-label free, clinically valid, and highly scalable method for segmentation from ultrasound, a noisy but globally important imaging modality.

In recommendation, graph-based Collaborative Filtering (CF) methods mitigate the data sparsity by introducing Graph Contrastive Learning (GCL). However, the random negative sampling strategy in these GCL-based CF models neglects the semantic structure of users (items), which not only introduces false negatives (negatives that are similar to anchor user (item)) but also ignores the potential positive samples. To tackle the above issues, we propose Topology-aware Debiased Self-supervised Graph Learning (TDSGL) for recommendation, which constructs contrastive pairs according to the semantic similarity between users (items). Specifically, since the original user-item interaction data commendably reflects the purchasing intent of users and certain characteristics of items, we calculate the semantic similarity between users (items) on interaction data. Then, given a user (item), we construct its negative pairs by selecting users (items) which embed different semantic structures to ensure the semantic difference between the given user (item) and its negatives. Moreover, for a user (item), we design a feature extraction module that converts other semantically similar users (items) into an auxiliary positive sample to acquire a more informative representation. Experimental results show that the proposed model outperforms the state-of-the-art models significantly on three public datasets. Our model implementation codes are available at https://github.com/malajikuai/TDSGL.

We present DALE, a novel and effective generative Data Augmentation framework for low-resource LEgal NLP. DALE addresses the challenges existing frameworks pose in generating effective data augmentations of legal documents - legal language, with its specialized vocabulary and complex semantics, morphology, and syntax, does not benefit from data augmentations that merely rephrase the source sentence. To address this, DALE, built on an Encoder-Decoder Language Model, is pre-trained on a novel unsupervised text denoising objective based on selective masking - our masking strategy exploits the domain-specific language characteristics of templatized legal documents to mask collocated spans of text. Denoising these spans helps DALE acquire knowledge about legal concepts, principles, and language usage. Consequently, it develops the ability to generate coherent and diverse augmentations with novel contexts. Finally, DALE performs conditional generation to generate synthetic augmentations for low-resource Legal NLP tasks. We demonstrate the effectiveness of DALE on 13 datasets spanning 6 tasks and 4 low-resource settings. DALE outperforms all our baselines, including LLMs, qualitatively and quantitatively, with improvements of 1%-50%.

Semi-supervised learning (SSL) has become popular in recent years because it allows the training of a model using a large amount of unlabeled data. However, one issue that many SSL methods face is the confirmation bias, which occurs when the model is overfitted to the small labeled training dataset and produces overconfident, incorrect predictions. To address this issue, we propose SequenceMatch, an efficient SSL method that utilizes multiple data augmentations. The key element of SequenceMatch is the inclusion of a medium augmentation for unlabeled data. By taking advantage of different augmentations and the consistency constraints between each pair of augmented examples, SequenceMatch helps reduce the divergence between the prediction distribution of the model for weakly and strongly augmented examples. In addition, SequenceMatch defines two different consistency constraints for high and low-confidence predictions. As a result, SequenceMatch is more data-efficient than ReMixMatch, and more time-efficient than both ReMixMatch ($\times4$) and CoMatch ($\times2$) while having higher accuracy. Despite its simplicity, SequenceMatch consistently outperforms prior methods on standard benchmarks, such as CIFAR-10/100, SVHN, and STL-10. It also surpasses prior state-of-the-art methods by a large margin on large-scale datasets such as ImageNet, with a 38.46\% error rate. Code is available at https://github.com/beandkay/SequenceMatch.

Recent studies on semi-supervised learning (SSL) have achieved great success. Despite their promising performance, current state-of-the-art methods tend toward increasingly complex designs at the cost of introducing more network components and additional training procedures. In this paper, we propose a simple method named Ensemble Projectors Aided for Semi-supervised Learning (EPASS), which focuses mainly on improving the learned embeddings to boost the performance of the existing contrastive joint-training semi-supervised learning frameworks. Unlike standard methods, where the learned embeddings from one projector are stored in memory banks to be used with contrastive learning, EPASS stores the ensemble embeddings from multiple projectors in memory banks. As a result, EPASS improves generalization, strengthens feature representation, and boosts performance. For instance, EPASS improves strong baselines for semi-supervised learning by 39.47\%/31.39\%/24.70\% top-1 error rate, while using only 100k/1\%/10\% of labeled data for SimMatch, and achieves 40.24\%/32.64\%/25.90\% top-1 error rate for CoMatch on the ImageNet dataset. These improvements are consistent across methods, network architectures, and datasets, proving the general effectiveness of the proposed methods. Code is available at https://github.com/beandkay/EPASS.

Recent research suggests that the feed-forward module within Transformers can be viewed as a collection of key-value memories, where the keys learn to capture specific patterns from the input based on the training examples. The values then combine the output from the 'memories' of the keys to generate predictions about the next token. This leads to an incremental process of prediction that gradually converges towards the final token choice near the output layers. This interesting perspective raises questions about how multilingual models might leverage this mechanism. Specifically, for autoregressive models trained on two or more languages, do all neurons (across layers) respond equally to all languages? No! Our hypothesis centers around the notion that during pretraining, certain model parameters learn strong language-specific features, while others learn more language-agnostic (shared across languages) features. To validate this, we conduct experiments utilizing parallel corpora of two languages that the model was initially pretrained on. Our findings reveal that the layers closest to the network's input or output tend to exhibit more language-specific behaviour compared to the layers in the middle.

Data augmentation techniques are widely used in low-resource automatic morphological inflection to overcome data sparsity. However, the full implications of these techniques remain poorly understood. In this study, we aim to shed light on the theoretical aspects of the prominent data augmentation strategy StemCorrupt (Silfverberg et al., 2017; Anastasopoulos and Neubig, 2019), a method that generates synthetic examples by randomly substituting stem characters in gold standard training examples. To begin, we conduct an information-theoretic analysis, arguing that StemCorrupt improves compositional generalization by eliminating spurious correlations between morphemes, specifically between the stem and the affixes. Our theoretical analysis further leads us to study the sample efficiency with which StemCorrupt reduces these spurious correlations. Through evaluation across seven typologically distinct languages, we demonstrate that selecting a subset of datapoints with both high diversity and high predictive uncertainty significantly enhances the data-efficiency of StemCorrupt. However, we also explore the impact of typological features on the choice of the data selection strategy and find that languages incorporating a high degree of allomorphy and phonological alternations derive less benefit from synthetic examples with high uncertainty. We attribute this effect to phonotactic violations induced by StemCorrupt, emphasizing the need for further research to ensure optimal performance across the entire spectrum of natural language morphology.

In this paper, we address the challenging problem of data association for underwater SLAM through a novel method for sonar image correspondence using learned features. We introduce SONIC (SONar Image Correspondence), a pose-supervised network designed to yield robust feature correspondence capable of withstanding viewpoint variations. The inherent complexity of the underwater environment stems from the dynamic and frequently limited visibility conditions, restricting vision to a few meters of often featureless expanses. This makes camera-based systems suboptimal in most open water application scenarios. Consequently, multibeam imaging sonars emerge as the preferred choice for perception sensors. However, they too are not without their limitations. While imaging sonars offer superior long-range visibility compared to cameras, their measurements can appear different from varying viewpoints. This inherent variability presents formidable challenges in data association, particularly for feature-based methods. Our method demonstrates significantly better performance in generating correspondences for sonar images which will pave the way for more accurate loop closure constraints and sonar-based place recognition. Code as well as simulated and real-world datasets will be made public to facilitate further development in the field.