Saurabh Garg

Saurabh Garg, Mehrdad Farajtabar, Hadi Pouransari et al. · utoronto

Keeping large foundation models up to date on latest data is inherently expensive. To avoid the prohibitive costs of constantly retraining, it is imperative to continually train these models. This problem is exacerbated by the lack of any large scale continual learning benchmarks or baselines. We introduce the first set of web-scale Time-Continual (TiC) benchmarks for training vision-language models: TiC-DataComp, TiC-YFCC, and TiC-Redcaps. TiC-DataComp, our largest dataset, contains over 12.7B timestamped image-text pairs spanning 9 years (2014-2022). We first use our benchmarks to curate various dynamic evaluations to measure temporal robustness of existing models. We show OpenAI's CLIP (trained on data up to 2020) loses $\approx 8\%$ zero-shot accuracy on our curated retrieval task from 2021-2022 compared with more recently trained models in OpenCLIP repository. We then study how to efficiently train models on time-continuous data. We demonstrate that a simple rehearsal-based approach that continues training from the last checkpoint and replays old data reduces compute by $2.5\times$ when compared to the standard practice of retraining from scratch. Code is available at https://github.com/apple/ml-tic-clip.

Zachary Novack, Julian McAuley, Zachary C. Lipton et al.

Open vocabulary models (e.g. CLIP) have shown strong performance on zero-shot classification through their ability generate embeddings for each class based on their (natural language) names. Prior work has focused on improving the accuracy of these models through prompt engineering or by incorporating a small amount of labeled downstream data (via finetuning). However, there has been little focus on improving the richness of the class names themselves, which can pose issues when class labels are coarsely-defined and are uninformative. We propose Classification with Hierarchical Label Sets (or CHiLS), an alternative strategy for zero-shot classification specifically designed for datasets with implicit semantic hierarchies. CHiLS proceeds in three steps: (i) for each class, produce a set of subclasses, using either existing label hierarchies or by querying GPT-3; (ii) perform the standard zero-shot CLIP procedure as though these subclasses were the labels of interest; (iii) map the predicted subclass back to its parent to produce the final prediction. Across numerous datasets with underlying hierarchical structure, CHiLS leads to improved accuracy in situations both with and without ground-truth hierarchical information. CHiLS is simple to implement within existing zero-shot pipelines and requires no additional training cost. Code is available at: https://github.com/acmi-lab/CHILS.

Pratyush Maini, Saurabh Garg, Zachary C. Lipton et al.

Researchers investigating example hardness have increasingly focused on the dynamics by which neural networks learn and forget examples throughout training. Popular metrics derived from these dynamics include (i) the epoch at which examples are first correctly classified; (ii) the number of times their predictions flip during training; and (iii) whether their prediction flips if they are held out. However, these metrics do not distinguish among examples that are hard for distinct reasons, such as membership in a rare subpopulation, being mislabeled, or belonging to a complex subpopulation. In this paper, we propose $second$-$split$ $forgetting$ $time$ (SSFT), a complementary metric that tracks the epoch (if any) after which an original training example is forgotten as the network is fine-tuned on a randomly held out partition of the data. Across multiple benchmark datasets and modalities, we demonstrate that $mislabeled$ examples are forgotten quickly, and seemingly $rare$ examples are forgotten comparatively slowly. By contrast, metrics only considering the first split learning dynamics struggle to differentiate the two. At large learning rates, SSFT tends to be robust across architectures, optimizers, and random seeds. From a practical standpoint, the SSFT can (i) help to identify mislabeled samples, the removal of which improves generalization; and (ii) provide insights about failure modes. Through theoretical analysis addressing overparameterized linear models, we provide insights into how the observed phenomena may arise. Code for reproducing our experiments can be found here: https://github.com/pratyushmaini/ssft

Saurabh Garg, Sivaraman Balakrishnan, Zachary C. Lipton

We introduce the problem of domain adaptation under Open Set Label Shift (OSLS) where the label distribution can change arbitrarily and a new class may arrive during deployment, but the class-conditional distributions p(x|y) are domain-invariant. OSLS subsumes domain adaptation under label shift and Positive-Unlabeled (PU) learning. The learner's goals here are two-fold: (a) estimate the target label distribution, including the novel class; and (b) learn a target classifier. First, we establish necessary and sufficient conditions for identifying these quantities. Second, motivated by advances in label shift and PU learning, we propose practical methods for both tasks that leverage black-box predictors. Unlike typical Open Set Domain Adaptation (OSDA) problems, which tend to be ill-posed and amenable only to heuristics, OSLS offers a well-posed problem amenable to more principled machinery. Experiments across numerous semi-synthetic benchmarks on vision, language, and medical datasets demonstrate that our methods consistently outperform OSDA baselines, achieving 10--25% improvements in target domain accuracy. Finally, we analyze the proposed methods, establishing finite-sample convergence to the true label marginal and convergence to optimal classifier for linear models in a Gaussian setup. Code is available at https://github.com/acmi-lab/Open-Set-Label-Shift.

Saurabh Garg, Nick Erickson, James Sharpnack et al.

Despite the emergence of principled methods for domain adaptation under label shift, their sensitivity to shifts in class conditional distributions is precariously under explored. Meanwhile, popular deep domain adaptation heuristics tend to falter when faced with label proportions shifts. While several papers modify these heuristics in attempts to handle label proportions shifts, inconsistencies in evaluation standards, datasets, and baselines make it difficult to gauge the current best practices. In this paper, we introduce RLSbench, a large-scale benchmark for relaxed label shift, consisting of $>$500 distribution shift pairs spanning vision, tabular, and language modalities, with varying label proportions. Unlike existing benchmarks, which primarily focus on shifts in class-conditional $p(x|y)$, our benchmark also focuses on label marginal shifts. First, we assess 13 popular domain adaptation methods, demonstrating more widespread failures under label proportion shifts than were previously known. Next, we develop an effective two-step meta-algorithm that is compatible with most domain adaptation heuristics: (i) pseudo-balance the data at each epoch; and (ii) adjust the final classifier with target label distribution estimate. The meta-algorithm improves existing domain adaptation heuristics under large label proportion shifts, often by 2--10\% accuracy points, while conferring minimal effect ($<$0.5\%) when label proportions do not shift. We hope that these findings and the availability of RLSbench will encourage researchers to rigorously evaluate proposed methods in relaxed label shift settings. Code is publicly available at https://github.com/acmi-lab/RLSbench.

Elan Rosenfeld, Saurabh Garg

We derive an (almost) guaranteed upper bound on the error of deep neural networks under distribution shift using unlabeled test data. Prior methods either give bounds that are vacuous in practice or give estimates that are accurate on average but heavily underestimate error for a sizeable fraction of shifts. In particular, the latter only give guarantees based on complex continuous measures such as test calibration -- which cannot be identified without labels -- and are therefore unreliable. Instead, our bound requires a simple, intuitive condition which is well justified by prior empirical works and holds in practice effectively 100% of the time. The bound is inspired by $\mathcal{H}Δ\mathcal{H}$-divergence but is easier to evaluate and substantially tighter, consistently providing non-vacuous guarantees. Estimating the bound requires optimizing one multiclass classifier to disagree with another, for which some prior works have used sub-optimal proxy losses; we devise a "disagreement loss" which is theoretically justified and performs better in practice. We expect this loss can serve as a drop-in replacement for future methods which require maximizing multiclass disagreement. Across a wide range of benchmarks, our method gives valid error bounds while achieving average accuracy comparable to competitive estimation baselines. Code is publicly available at https://github.com/erosenfeld/disagree_discrep .

Kundan Krishna, Saurabh Garg, Jeffrey P. Bigham et al.

For most natural language processing tasks, the dominant practice is to finetune large pretrained transformer models (e.g., BERT) using smaller downstream datasets. Despite the success of this approach, it remains unclear to what extent these gains are attributable to the massive background corpora employed for pretraining versus to the pretraining objectives themselves. This paper introduces a large-scale study of self-pretraining, where the same (downstream) training data is used for both pretraining and finetuning. In experiments addressing both ELECTRA and RoBERTa models and 10 distinct downstream classification datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around $10\times$--$500\times$ less data), outperforming the latter on $7$ and $5$ datasets, respectively. Surprisingly, these task-specific pretrained models often perform well on other tasks, including the GLUE benchmark. Besides classification tasks, self-pretraining also provides benefits on structured output prediction tasks such as span based question answering and commonsense inference, often providing more than $50\%$ of the performance boosts provided by pretraining on the BookWiki corpus. Our results hint that in many scenarios, performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the use of external pretraining data in massive amounts. These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data.

Manley Roberts, Pranav Mani, Saurabh Garg et al.

What sorts of structure might enable a learner to discover classes from unlabeled data? Traditional approaches rely on feature-space similarity and heroic assumptions on the data. In this paper, we introduce unsupervised learning under Latent Label Shift (LLS), where we have access to unlabeled data from multiple domains such that the label marginals $p_d(y)$ can shift across domains but the class conditionals $p(\mathbf{x}|y)$ do not. This work instantiates a new principle for identifying classes: elements that shift together group together. For finite input spaces, we establish an isomorphism between LLS and topic modeling: inputs correspond to words, domains to documents, and labels to topics. Addressing continuous data, we prove that when each label's support contains a separable region, analogous to an anchor word, oracle access to $p(d|\mathbf{x})$ suffices to identify $p_d(y)$ and $p_d(y|\mathbf{x})$ up to permutation. Thus motivated, we introduce a practical algorithm that leverages domain-discriminative models as follows: (i) push examples through domain discriminator $p(d|\mathbf{x})$; (ii) discretize the data by clustering examples in $p(d|\mathbf{x})$ space; (iii) perform non-negative matrix factorization on the discrete data; (iv) combine the recovered $p(y|d)$ with the discriminator outputs $p(d|\mathbf{x})$ to compute $p_d(y|x) \; \forall d$. With semi-synthetic experiments, we show that our algorithm can leverage domain information to improve upon competitive unsupervised classification methods. We reveal a failure mode of standard unsupervised classification methods when feature-space similarity does not indicate true groupings, and show empirically that our method better handles this case. Our results establish a deep connection between distribution shift and topic modeling, opening promising lines for future work.

Zachary Novack, Simran Kaur, Tanya Marwah et al.

A number of competing hypotheses have been proposed to explain why small-batch Stochastic Gradient Descent (SGD)leads to improved generalization over the full-batch regime, with recent work crediting the implicit regularization of various quantities throughout training. However, to date, empirical evidence assessing the explanatory power of these hypotheses is lacking. In this paper, we conduct an extensive empirical evaluation, focusing on the ability of various theorized mechanisms to close the small-to-large batch generalization gap. Additionally, we characterize how the quantities that SGD has been claimed to (implicitly) regularize change over the course of training. By using micro-batches, i.e. disjoint smaller subsets of each mini-batch, we empirically show that explicitly penalizing the gradient norm or the Fisher Information Matrix trace, averaged over micro-batches, in the large-batch regime recovers small-batch SGD generalization, whereas Jacobian-based regularizations fail to do so. This generalization performance is shown to often be correlated with how well the regularized model's gradient norms resemble those of small-batch SGD. We additionally show that this behavior breaks down as the micro-batch size approaches the batch size. Finally, we note that in this line of inquiry, positive experimental findings on CIFAR10 are often reversed on other datasets like CIFAR100, highlighting the need to test hypotheses on a wider collection of datasets.

Renjie Li, Xinyi Wang, Guan Huang et al.

Deep supervision, or known as 'intermediate supervision' or 'auxiliary supervision', is to add supervision at hidden layers of a neural network. This technique has been increasingly applied in deep neural network learning systems for various computer vision applications recently. There is a consensus that deep supervision helps improve neural network performance by alleviating the gradient vanishing problem, as one of the many strengths of deep supervision. Besides, in different computer vision applications, deep supervision can be applied in different ways. How to make the most use of deep supervision to improve network performance in different applications has not been thoroughly investigated. In this paper, we provide a comprehensive in-depth review of deep supervision in both theories and applications. We propose a new classification of different deep supervision networks, and discuss advantages and limitations of current deep supervision networks in computer vision applications.

Asbjørn Munk, Stefano Cerri, Vardan Nersesjan et al.

Clinical deployment of automated brain MRI analysis faces a fundamental challenge: clinical data is heterogeneous and noisy, and high-quality labels are prohibitively costly to obtain. Self-supervised learning (SSL) can address this by leveraging the vast amounts of unlabeled data produced in clinical workflows to train robust \textit{foundation models} that adapt out-of-domain with minimal supervision. However, the development of foundation models for brain MRI has been limited by small pretraining datasets and in-domain benchmarking focused on high-quality, research-grade data. To address this gap, we organized the FOMO25 challenge as a satellite event at MICCAI 2025. FOMO25 provided participants with a large pretraining dataset, FOMO60K, and evaluated models on data sourced directly from clinical workflows in few-shot and out-of-domain settings. Tasks covered infarct classification, meningioma segmentation, and brain age regression, and considered both models trained on FOMO60K (method track) and any data (open track). Nineteen foundation models from sixteen teams were evaluated using a standardized containerized pipeline. Results show that (a) self-supervised pretraining improves generalization on clinical data under domain shift, with the strongest models trained \textit{out-of-domain} surpassing supervised baselines trained \textit{in-domain}. (b) No single pretraining objective benefits all tasks: MAE favors segmentation, hybrid reconstruction-contrastive objectives favor classification, and (c) strong performance was achieved by small pretrained models, and improvements from scaling model size and training duration did not yield reliable benefits.

Jianping Yao, Son N. Tran, Saurabh Garg et al.

Deep learning plays an important role in modern agriculture, especially in plant pathology using leaf images where convolutional neural networks (CNN) are attracting a lot of attention. While numerous reviews have explored the applications of deep learning within this research domain, there remains a notable absence of an empirical study to offer insightful comparisons due to the employment of varied datasets in the evaluation. Furthermore, a majority of these approaches tend to address the problem as a singular prediction task, overlooking the multifaceted nature of predicting various aspects of plant species and disease types. Lastly, there is an evident need for a more profound consideration of the semantic relationships that underlie plant species and disease types. In this paper, we start our study by surveying current deep learning approaches for plant identification and disease classification. We categorise the approaches into multi-model, multi-label, multi-output, and multi-task, in which different backbone CNNs can be employed. Furthermore, based on the survey of existing approaches in plant pathology and the study of available approaches in machine learning, we propose a new model named Generalised Stacking Multi-output CNN (GSMo-CNN). To investigate the effectiveness of different backbone CNNs and learning approaches, we conduct an intensive experiment on three benchmark datasets Plant Village, Plant Leaves, and PlantDoc. The experimental results demonstrate that InceptionV3 can be a good choice for a backbone CNN as its performance is better than AlexNet, VGG16, ResNet101, EfficientNet, MobileNet, and a custom CNN developed by us. Interestingly, empirical results support the hypothesis that using a single model can be comparable or better than using two models. Finally, we show that the proposed GSMo-CNN achieves state-of-the-art performance on three benchmark datasets.

Jianping Yao, Son N. Tran, Samantha Sawyer et al.

The growing demand for sustainable development brings a series of information technologies to help agriculture production. Especially, the emergence of machine learning applications, a branch of artificial intelligence, has shown multiple breakthroughs which can enhance and revolutionize plant pathology approaches. In recent years, machine learning has been adopted for leaf disease classification in both academic research and industrial applications. Therefore, it is enormously beneficial for researchers, engineers, managers, and entrepreneurs to have a comprehensive view about the recent development of machine learning technologies and applications for leaf disease detection. This study will provide a survey in different aspects of the topic including data, techniques, and applications. The paper will start with publicly available datasets. After that, we summarize common machine learning techniques, including traditional (shallow) learning, deep learning, and augmented learning. Finally, we discuss related applications. This paper would provide useful resources for future study and application of machine learning for smart agriculture in general and leaf disease classification in particular.

Renjie Li, Chun Yu Lao, Rebecca St. George et al.

Objective The coordination of human movement directly reflects function of the central nervous system. Small deficits in movement are often the first sign of an underlying neurological problem. The objective of this research is to develop a new end-to-end, deep learning-based system, Rapid-Motion-Track (RMT) that can track the fastest human movement accurately when webcams or laptop cameras are used. Materials and Methods We applied RMT to finger tapping, a well-validated test of motor control that is one of the most challenging human motions to track with computer vision due to the small keypoints of digits and the high velocities that are generated. We recorded 160 finger tapping assessments simultaneously with a standard 2D laptop camera (30 frames/sec) and a high-speed wearable sensor-based 3D motion tracking system (250 frames/sec). RMT and a range of DLC models were applied to the video data with tapping frequencies up to 8Hz to extract movement features. Results The movement features (e.g. speed, rhythm, variance) identified with the new RMT system exhibited very high concurrent validity with the gold-standard measurements (97.3\% of RMT measures were within +/-0.5Hz of the Optotrak measures), and outperformed DLC and other advanced computer vision tools (around 88.2\% of DLC measures were within +/-0.5Hz of the Optotrak measures). RMT also accurately tracked a range of other rapid human movements such as foot tapping, head turning and sit-to -stand movements. Conclusion: With the ubiquity of video technology in smart devices, the RMT method holds potential to transform access and accuracy of human movement assessment.

Dheeraj Baby, Saurabh Garg, Tzu-Ching Yen et al.

This paper focuses on supervised and unsupervised online label shift, where the class marginals $Q(y)$ varies but the class-conditionals $Q(x|y)$ remain invariant. In the unsupervised setting, our goal is to adapt a learner, trained on some offline labeled data, to changing label distributions given unlabeled online data. In the supervised setting, we must both learn a classifier and adapt to the dynamically evolving class marginals given only labeled online data. We develop novel algorithms that reduce the adaptation problem to online regression and guarantee optimal dynamic regret without any prior knowledge of the extent of drift in the label distribution. Our solution is based on bootstrapping the estimates of \emph{online regression oracles} that track the drifting proportions. Experiments across numerous simulated and real-world online label shift scenarios demonstrate the superior performance of our proposed approaches, often achieving 1-3\% improvement in accuracy while being sample and computationally efficient. Code is publicly available at https://github.com/acmi-lab/OnlineLabelShift.

Saurabh Garg, Sivaraman Balakrishnan, Zachary C. Lipton et al.

Real-world machine learning deployments are characterized by mismatches between the source (training) and target (test) distributions that may cause performance drops. In this work, we investigate methods for predicting the target domain accuracy using only labeled source data and unlabeled target data. We propose Average Thresholded Confidence (ATC), a practical method that learns a threshold on the model's confidence, predicting accuracy as the fraction of unlabeled examples for which model confidence exceeds that threshold. ATC outperforms previous methods across several model architectures, types of distribution shifts (e.g., due to synthetic corruptions, dataset reproduction, or novel subpopulations), and datasets (Wilds, ImageNet, Breeds, CIFAR, and MNIST). In our experiments, ATC estimates target performance $2$-$4\times$ more accurately than prior methods. We also explore the theoretical foundations of the problem, proving that, in general, identifying the accuracy is just as hard as identifying the optimal predictor and thus, the efficacy of any method rests upon (perhaps unstated) assumptions on the nature of the shift. Finally, analyzing our method on some toy distributions, we provide insights concerning when it works. Code is available at https://github.com/saurabhgarg1996/ATC_code/.

Daniel P. Jeong, Saurabh Garg, Zachary C. Lipton et al.

Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pretraining on publicly available biomedical corpora. These works typically claim that such domain-adaptive pretraining (DAPT) improves performance on downstream medical tasks, such as answering medical licensing exam questions. In this paper, we compare seven public "medical" LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting regime for medical question-answering (QA) tasks. For instance, across the tasks and model pairs we consider in the 3-shot setting, medical LLMs only outperform their base models in 12.1% of cases, reach a (statistical) tie in 49.8% of cases, and are significantly worse than their base models in the remaining 38.2% of cases. Our conclusions are based on (i) comparing each medical model head-to-head, directly against the corresponding base model; (ii) optimizing the prompts for each model separately; and (iii) accounting for statistical uncertainty in comparisons. While these basic practices are not consistently adopted in the literature, our ablations show that they substantially impact conclusions. Our findings suggest that state-of-the-art general-domain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.

Saurabh Garg, Amrith Setlur, Zachary Chase Lipton et al. · cmu

Self-training and contrastive learning have emerged as leading techniques for incorporating unlabeled data, both under distribution shift (unsupervised domain adaptation) and when it is absent (semi-supervised learning). However, despite the popularity and compatibility of these techniques, their efficacy in combination remains unexplored. In this paper, we undertake a systematic empirical investigation of this combination, finding that (i) in domain adaptation settings, self-training and contrastive learning offer significant complementary gains; and (ii) in semi-supervised learning settings, surprisingly, the benefits are not synergistic. Across eight distribution shift datasets (e.g., BREEDs, WILDS), we demonstrate that the combined method obtains 3--8% higher accuracy than either approach independently. We then theoretically analyze these techniques in a simplified model of distribution shift, demonstrating scenarios under which the features produced by contrastive learning can yield a good initialization for self-training to further amplify gains and achieve optimal performance, even when either method alone would fail.

Daniel P. Jeong, Pranav Mani, Saurabh Garg et al.

Several recent works seek to adapt general-purpose large language models (LLMs) and vision-language models (VLMs) for medical applications through continued pretraining on publicly available biomedical corpora. These works typically claim that such domain-adaptive pretraining improves performance on various downstream medical tasks, such as answering medical exam questions. In this paper, we compare ten "medical" LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting and supervised fine-tuning regimes for medical question answering (QA). For instance, on clinical-note-based QA tasks in the 3-shot setting, medical LLMs outperform their base models in only 26.7% of cases, reach a (statistical) tie in 16.7% of cases, and perform significantly worse in the remaining 56.7% of cases. Our conclusions are based on (i) comparing each medical model directly against its base model; (ii) optimizing the prompts for each model separately in zero-/few-shot prompting; and (iii) accounting for statistical uncertainty in comparisons. Our findings suggest that state-of-the-art general-domain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.

Alexander H. Liu, Andy Ehrenberg, Andy Lo et al. · deepmind

We present Voxtral Mini and Voxtral Small, two multimodal audio chat models. Voxtral is trained to comprehend both spoken audio and text documents, achieving state-of-the-art performance across a diverse range of audio benchmarks, while preserving strong text capabilities. Voxtral Small outperforms a number of closed-source models, while being small enough to run locally. A 32K context window enables the model to handle audio files up to 40 minutes in duration and long multi-turn conversations. We also contribute three benchmarks for evaluating speech understanding models on knowledge and trivia. Both Voxtral models are released under Apache 2.0 license.

Rishabh Ranjan, Saurabh Garg, Mrigank Raman et al.

Trained models are often composed with post-hoc transforms such as temperature scaling (TS), ensembling and stochastic weight averaging (SWA) to improve performance, robustness, uncertainty estimation, etc. However, such transforms are typically applied only after the base models have already been finalized by standard means. In this paper, we challenge this practice with an extensive empirical study. In particular, we demonstrate a phenomenon that we call post-hoc reversal, where performance trends are reversed after applying post-hoc transforms. This phenomenon is especially prominent in high-noise settings. For example, while base models overfit badly early in training, both ensembling and SWA favor base models trained for more epochs. Post-hoc reversal can also prevent the appearance of double descent and mitigate mismatches between test loss and test error seen in base models. Preliminary analyses suggest that these transforms induce reversal by suppressing the influence of mislabeled examples, exploiting differences in their learning dynamics from those of clean examples. Based on our findings, we propose post-hoc selection, a simple technique whereby post-hoc metrics inform model development decisions such as early stopping, checkpointing, and broader hyperparameter choices. Our experiments span real-world vision, language, tabular and graph datasets. On an LLM instruction tuning dataset, post-hoc selection results in >1.5x MMLU improvement compared to naive selection.

Roozbeh Bazargani, Saqib Abdullah Basar, Daniel Daly-Grafstein et al.

The human spine is a complex structure composed of 33 vertebrae. It holds the body and is important for leading a healthy life. The spine is vulnerable to age-related degenerations that can be identified through magnetic resonance imaging (MRI). In this paper we propose a novel computer-vison-based deep learning method to estimate spine age using images from over 18,000 MRI series. Data are restricted to subjects with only age-related spine degeneration. Eligibility criteria are created by identifying common age-based clusters of degenerative spine conditions using uniform manifold approximation and projection (UMAP) and hierarchical density-based spatial clustering of applications with noise (HDBSCAN). Model selection is determined using a detailed ablation study on data size, loss, and the effect of different spine regions. We evaluate the clinical utility of our model by calculating the difference between actual spine age and model-predicted age, the spine age gap (SAG), and examining the association between these differences and spine degenerative conditions and lifestyle factors. We find that SAG is associated with conditions including disc bulges, disc osteophytes, spinal stenosis, and fractures, as well as lifestyle factors like smoking and physically demanding work, and thus may be a useful biomarker for measuring overall spine health.

Yewon Byun, Sanket Vaibhav Mehta, Saurabh Garg et al. · deepmind

In many real-world settings, regulations and economic incentives permit the sharing of models but not data across institutional boundaries. In such scenarios, practitioners might hope to adapt models to new domains, without losing performance on previous domains (so-called catastrophic forgetting). While any single model may struggle to achieve this goal, learning an ensemble of domain-specific experts offers the potential to adapt more closely to each individual institution. However, a core challenge in this context is determining which expert to deploy at test time. In this paper, we propose Generate to Discriminate (G2D), a domain-incremental continual learning method that leverages synthetic data to train a domain-discriminator that routes samples at inference time to the appropriate expert. Surprisingly, we find that leveraging synthetic data in this capacity is more effective than using the samples to \textit{directly} train the downstream classifier (the more common approach to leveraging synthetic data in the lifelong learning literature). We observe that G2D outperforms competitive domain-incremental learning methods on tasks in both vision and language modalities, providing a new perspective on the use of synthetic data in the lifelong learning literature.

Amrith Setlur, Saurabh Garg, Xinyang Geng et al.

Training on model-generated synthetic data is a promising approach for finetuning LLMs, but it remains unclear when it helps or hurts. In this paper, we investigate this question for math reasoning via an empirical study, followed by building a conceptual understanding of our observations. First, we find that while the typical approach of finetuning a model on synthetic correct or positive problem-solution pairs generated by capable models offers modest performance gains, sampling more correct solutions from the finetuned learner itself followed by subsequent fine-tuning on this self-generated data $\textbf{doubles}$ the efficiency of the same synthetic problems. At the same time, training on model-generated positives can amplify various spurious correlations, resulting in flat or even inverse scaling trends as the amount of data increases. Surprisingly, we find that several of these issues can be addressed if we also utilize negative responses, i.e., model-generated responses that are deemed incorrect by a final answer verifier. Crucially, these negatives must be constructed such that the training can appropriately recover the utility or advantage of each intermediate step in the negative response. With this per-step scheme, we are able to attain consistent gains over only positive data, attaining performance similar to amplifying the amount of synthetic data by $\mathbf{8 \times}$. We show that training on per-step negatives can help to unlearn spurious correlations in the positive data, and is equivalent to advantage-weighted reinforcement learning (RL), implying that it inherits robustness benefits of RL over imitating positive data alone.

Jeffrey Li, Alex Fang, Georgios Smyrnis et al.

We introduce DataComp for Language Models (DCLM), a testbed for controlled dataset experiments with the goal of improving language models. As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations. Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing at model scales ranging from 412M to 7B parameters. As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set. The resulting dataset, DCLM-Baseline enables training a 7B parameter language model from scratch to 64% 5-shot accuracy on MMLU with 2.6T training tokens. Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6.6 percentage point improvement on MMLU while being trained with 40% less compute. Our baseline model is also comparable to Mistral-7B-v0.3 and Llama 3 8B on MMLU (63% & 66%), and performs similarly on an average of 53 natural language understanding tasks while being trained with 6.6x less compute than Llama 3 8B. Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation.

Gal Kaplun, Nikhil Ghosh, Saurabh Garg et al.

In machine learning, we traditionally evaluate the performance of a single model, averaged over a collection of test inputs. In this work, we propose a new approach: we measure the performance of a collection of models when evaluated on a $\textit{single input point}$. Specifically, we study a point's $\textit{profile}$: the relationship between models' average performance on the test distribution and their pointwise performance on this individual point. We find that profiles can yield new insights into the structure of both models and data -- in and out-of-distribution. For example, we empirically show that real data distributions consist of points with qualitatively different profiles. On one hand, there are "compatible" points with strong correlation between the pointwise and average performance. On the other hand, there are points with weak and even $\textit{negative}$ correlation: cases where improving overall model accuracy actually $\textit{hurts}$ performance on these inputs. We prove that these experimental observations are inconsistent with the predictions of several simplified models of learning proposed in prior work. As an application, we use profiles to construct a dataset we call CIFAR-10-NEG: a subset of CINIC-10 such that for standard models, accuracy on CIFAR-10-NEG is $\textit{negatively correlated}$ with accuracy on CIFAR-10 test. This illustrates, for the first time, an OOD dataset that completely inverts "accuracy-on-the-line" (Miller, Taori, Raghunathan, Sagawa, Koh, Shankar, Liang, Carmon, and Schmidt 2021)

Renjie Li, Son Tran, Saurabh Garg et al.

Keypoint detection plays an important role in a wide range of applications. However, predicting keypoints of small objects such as human hands is a challenging problem. Recent works fuse feature maps of deep Convolutional Neural Networks (CNNs), either via multi-level feature integration or multi-resolution aggregation. Despite achieving some success, the feature fusion approaches increase the complexity and the opacity of CNNs. To address this issue, we propose a novel CNN model named Multi-Scale Deep Supervision Network (P-MSDSNet) that learns feature maps at different scales with deep supervisions to produce attention maps for adaptive feature propagation from layers to layers. P-MSDSNet has a multi-stage architecture which makes it scalable while its deep supervision with spatial attention improves transparency to the feature learning at each stage. We show that P-MSDSNet outperforms the state-of-the-art approaches on benchmark datasets while requiring fewer number of parameters. We also show the application of P-MSDSNet to quantify finger tapping hand movements in a neuroscience study.

Saurabh Garg, Yifan Wu, Alex Smola et al.

Given only positive examples and unlabeled examples (from both positive and negative classes), we might hope nevertheless to estimate an accurate positive-versus-negative classifier. Formally, this task is broken down into two subtasks: (i) Mixture Proportion Estimation (MPE) -- determining the fraction of positive examples in the unlabeled data; and (ii) PU-learning -- given such an estimate, learning the desired positive-versus-negative classifier. Unfortunately, classical methods for both problems break down in high-dimensional settings. Meanwhile, recently proposed heuristics lack theoretical coherence and depend precariously on hyperparameter tuning. In this paper, we propose two simple techniques: Best Bin Estimation (BBE) (for MPE); and Conditional Value Ignoring Risk (CVIR), a simple objective for PU-learning. Both methods dominate previous approaches empirically, and for BBE, we establish formal guarantees that hold whenever we can train a model to cleanly separate out a small subset of positive examples. Our final algorithm (TED)$^n$, alternates between the two procedures, significantly improving both our mixture proportion estimator and classifier

Khizar Hameed, Saurabh Garg, Muhammad Bilal Amin et al.

In a clone node attack, an attacker attempted to physically capture the devices to gather sensitive information to conduct various insider attacks. Several solutions for detecting clone node attacks on IoT networks have been presented in the viewpoints above. These solutions are focused on specific system designs, processes, and feature sets and act as a high-level abstraction of underlying system architectures based on a few performance requirements. However, critical features like formal analysis, modelling, and verification are frequently overlooked in existing proposed solutions aimed at verifying the correctness and robustness of systems in order to ensure that no problematic scenarios or anomalies exist. This paper presents a formal analysis, modelling, and verification of our existing proposed clone node attack detection scheme in IoT. Firstly, we modelled the architectural components of the proposed scheme using High-Level Petri Nets (HLPNs) and then mapped them using their specified functionalities. Secondly, we defined and analysed the behavioural properties of the proposed scheme using Z specification language. Furthermore, we used the Satisfiability Modulo Theories Library (SMT-Lib) and the Z3 Solver to validate and demonstrate the overall functionality of the proposed scheme. Finally, in addition to modelling and analysis, this work employs Coloured Petri Nets (CPNs), which combine Petri Nets with a high-level programming language, making them more suitable for large-scale system modelling. To perform the simulations in CPN, we used both timed and untimed models, where timed models are used to evaluate performance, and untimed models are used to validate logical validity.

Khizar Hameed, Saurabh Garg, Muhammad Bilal Amin et al.

The rapidly expanding nature of the Internet of Things (IoT) networks is beginning to attract interest across a range of applications, including smart homes, smart transportation, smart health, and industrial contexts. This cutting-edge technology enables individuals to track and control their integrated environment in real-time and remotely via a thousand IoT devices comprised of sensors and actuators that actively participate in sensing, processing, storing, and sharing information. Nonetheless, IoT devices are frequently deployed in hostile environments, wherein adversaries attempt to capture and breach them in order to seize control of the entire network. One such example of potentially malicious behaviour is the cloning of IoT devices, in which an attacker can physically capture the devices, obtain some sensitive information, duplicate the devices, and intelligently deploy them in desired locations to conduct various insider attacks. A device cloning attack on IoT networks is a significant security concern since it allows for selective forwarding, sink-hole, and black-hole attacks. To address this issue, this paper provides an efficient scheme for detecting clone node attacks on IoT networks that makes use of semantic information about IoT devices known as context information sensed from the deployed environment to locate them securely. We design a location proof mechanism by combining location proofs and batch verification of the extended elliptic curve digital signature technique to accelerate the verification process at selected trusted nodes. We demonstrate the security of our scheme and its resilience to secure clone node attack detection by conducting a comprehensive security analysis. The performance of our proposed scheme provides a high degree of detection accuracy with minimal detection time and significantly reduces the computation, communication and storage overhead.

Khizar Hameed, Mutaz Barika, Saurabh Garg et al.

Blockchain technology has taken on a leading role in today's industrial applications by providing salient features and showing significant performance since its beginning. Blockchain began its journey from the concept of cryptocurrency and is now part of a range of core applications to achieve resilience and automation between various tasks. With the integration of Blockchain technology into different industrial applications, many application designs, security and privacy challenges present themselves, posing serious threats to users and their data. Although several approaches have been proposed to address the specific security and privacy needs of targeted applications with functional parameters, there is still a need for a research study on the application, security and privacy challenges, and requirements of Blockchain-based industrial applications, along with possible security threats and countermeasures. This study presents a state-of-the-art survey of Blockchain-based Industry 4.0 applications, focusing on crucial application and security and privacy requirements, as well as corresponding attacks on Blockchain systems with potential countermeasures. We also analyse and provide the classification of different security and privacy techniques used in these applications to enhance the advancement of security features. Furthermore, we highlight some open issues in industrial applications that help to design secure Blockchain-based applications as future directions.

Saurabh Garg, Sivaraman Balakrishnan, J. Zico Kolter et al.

To assess generalization, machine learning scientists typically either (i) bound the generalization gap and then (after training) plug in the empirical risk to obtain a bound on the true risk; or (ii) validate empirically on holdout data. However, (i) typically yields vacuous guarantees for overparameterized models. Furthermore, (ii) shrinks the training set and its guarantee erodes with each re-use of the holdout set. In this paper, we introduce a method that leverages unlabeled data to produce generalization bounds. After augmenting our (labeled) training set with randomly labeled fresh examples, we train in the standard fashion. Whenever classifiers achieve low error on clean data and high error on noisy data, our bound provides a tight upper bound on the true risk. We prove that our bound is valid for 0-1 empirical risk minimization and with linear classifiers trained by gradient descent. Our approach is especially useful in conjunction with deep learning due to the early learning phenomenon whereby networks fit true labels before noisy labels but requires one intuitive assumption. Empirically, on canonical computer vision and NLP tasks, our bound provides non-vacuous generalization guarantees that track actual performance closely. This work provides practitioners with an option for certifying the generalization of deep nets even when unseen labeled data is unavailable and provides theoretical insights into the relationship between random label noise and generalization.

Renjie Li, Xinyi Wang, Katherine Lawler et al.

With populations ageing, the number of people with dementia worldwide is expected to triple to 152 million by 2050. Seventy percent of cases are due to Alzheimer's disease (AD) pathology and there is a 10-20 year 'pre-clinical' period before significant cognitive decline occurs. We urgently need, cost effective, objective methods to detect AD, and other dementias, at an early stage. Risk factor modification could prevent 40% of cases and drug trials would have greater chances of success if participants are recruited at an earlier stage. Currently, detection of dementia is largely by pen and paper cognitive tests but these are time consuming and insensitive to pre-clinical phases. Specialist brain scans and body fluid biomarkers can detect the earliest stages of dementia but are too invasive or expensive for widespread use. With the advancement of technology, Artificial Intelligence (AI) shows promising results in assisting with detection of early-stage dementia. Existing AI-aided methods and potential future research directions are reviewed and discussed.

Saurabh Garg, Joshua Zhanson, Emilio Parisotto et al.

Modern policy gradient algorithms such as Proximal Policy Optimization (PPO) rely on an arsenal of heuristics, including loss clipping and gradient clipping, to ensure successful learning. These heuristics are reminiscent of techniques from robust statistics, commonly used for estimation in outlier-rich (``heavy-tailed'') regimes. In this paper, we present a detailed empirical study to characterize the heavy-tailed nature of the gradients of the PPO surrogate reward function. We demonstrate that the gradients, especially for the actor network, exhibit pronounced heavy-tailedness and that it increases as the agent's policy diverges from the behavioral policy (i.e., as the agent goes further off policy). Further examination implicates the likelihood ratios and advantages in the surrogate reward as the main sources of the observed heavy-tailedness. We then highlight issues arising due to the heavy-tailed nature of the gradients. In this light, we study the effects of the standard PPO clipping heuristics, demonstrating that these tricks primarily serve to offset heavy-tailedness in gradients. Thus motivated, we propose incorporating GMOM, a high-dimensional robust estimator, into PPO as a substitute for three clipping tricks. Despite requiring less hyperparameter tuning, our method matches the performance of PPO (with all heuristics enabled) on a battery of MuJoCo continuous control tasks.

Saurabh Garg, Yifan Wu, Sivaraman Balakrishnan et al.

Under label shift, the label distribution p(y) might change but the class-conditional distributions p(x|y) do not. There are two dominant approaches for estimating the label marginal. BBSE, a moment-matching approach based on confusion matrices, is provably consistent and provides interpretable error bounds. However, a maximum likelihood estimation approach, which we call MLLS, dominates empirically. In this paper, we present a unified view of the two methods and the first theoretical characterization of MLLS. Our contributions include (i) consistency conditions for MLLS, which include calibration of the classifier and a confusion matrix invertibility condition that BBSE also requires; (ii) a unified framework, casting BBSE as roughly equivalent to MLLS for a particular choice of calibration method; and (iii) a decomposition of MLLS's finite-sample error into terms reflecting miscalibration and estimation error. Our analysis attributes BBSE's statistical inefficiency to a loss of information due to coarse calibration. Experiments on synthetic data, MNIST, and CIFAR10 support our findings.

Abdullah Al-Noman Patwary, Anmin Fu, Ranesh Kumar Naha et al.

Fog computing is an emerging computing paradigm that has come into consideration for the deployment of IoT applications amongst researchers and technology industries over the last few years. Fog is highly distributed and consists of a wide number of autonomous end devices, which contribute to the processing. However, the variety of devices offered across different users are not audited. Hence, the security of Fog devices is a major concern in the Fog computing environment. Furthermore, mitigating and preventing those security measures is a research issue. Therefore, to provide the necessary security for Fog devices, we need to understand what the security concerns are with regards to Fog. All aspects of Fog security, which have not been covered by other literature works needs to be identified and need to be aggregate all issues in Fog security. It needs to be noted that computation devices consist of many ordinary users, and are not managed by any central entity or managing body. Therefore, trust and privacy is also a key challenge to gain market adoption for Fog. To provide the required trust and privacy, we need to also focus on authentication, threats and access control mechanisms as well as techniques in Fog computing. In this paper, we perform a survey and propose a taxonomy, which presents an overview of existing security concerns in the context of the Fog computing paradigm. We discuss the Blockchain-based solutions towards a secure Fog computing environment and presented various research challenges and directions for future research.

Saurabh Garg, Tanmay Parekh, Preethi Jyothi

This work focuses on building language models (LMs) for code-switched text. We propose two techniques that significantly improve these LMs: 1) A novel recurrent neural network unit with dual components that focus on each language in the code-switched text separately 2) Pretraining the LM using synthetic text from a generative model estimated using the training data. We demonstrate the effectiveness of our proposed techniques by reporting perplexities on a Mandarin-English task and derive significant reductions in perplexity.

Erfan Aghasian, Saurabh Garg, James Montgomery

Recommender systems have become an integral part of many social networks and extract knowledge from a user's personal and sensitive data both explicitly, with the user's knowledge, and implicitly. This trend has created major privacy concerns as users are mostly unaware of what data and how much data is being used and how securely it is used. In this context, several works have been done to address privacy concerns for usage in online social network data and by recommender systems. This paper surveys the main privacy concerns, measurements and privacy-preserving techniques used in large-scale online social networks and recommender systems. It is based on historical works on security, privacy-preserving, statistical modeling, and datasets to provide an overview of the technical difficulties and problems associated with privacy preserving in online social networks.

Alexander Brown, Saurabh Garg, James Montgomery

Recording and analysing environmental audio recordings has become a common approach for monitoring the environment. A current problem with performing analyses of environmental recordings is interference from noise that can mask sounds of interest. This makes detecting these sounds more difficult and can require additional resources. While some work has been done to remove stationary noise from environmental recordings, there has been little effort to remove noise from non-stationary sources, such as rain, wind, engines, and animal vocalisations that are not of interest. In this paper, we address the challenge of filtering noise from rain and cicada choruses from recordings containing bird sound. We improve upon previously established classification approaches using acoustic indices and Mel Frequency Cepstral Coefficients (MFCCs) as acoustic features to detect these noise sources, approaching the problem with the motivation of removing these sounds. We investigate the use of acoustic indices, and machine learning classifiers to find the most effective filters. The approach we use enables users to set thresholds to increase or decrease the sensitivity of classification, based on the prediction probability outputted by classifiers. We also propose a novel approach to remove cicada choruses using band-pass filters Our threshold-based approach (Random Forest with Acoustic Indices and Mel Frequency Cepstral Coefficients (MFCCs)) for rain detection achieves an AUC of 0.9881 and is more accurate than existing approaches when set to the same sensitivities. We also detect cicada choruses in our training set with 100% accuracy using 10-folds cross validation. Our cicada filtering approach greatly increased the median signal to noise ratios of affected recordings from 0.53 for unfiltered audio to 1.86 to audio filtered by both the cicada filter and a stationary noise filter.

Alexander Brown, Saurabh Garg, James Montgomery

In this work, we examine the problem of efficiently preprocessing high volume bird acoustic data. We combine several existing preprocessing steps including noise reduction approaches into a single efficient pipeline by examining each process individually. We then utilise a distributed computing architecture to improve execution time. Using a master-slave model with data parallelisation, we developed a near-linear automated scalable system, capable of preprocessing bird acoustic recordings 21.76 times faster with 32 cores over 8 virtual machines, compared to a serial process. This work contributes to the research area of bioacoustic analysis, which is currently very active because of its potential to monitor animals quickly at low cost. Overcoming noise interference is a significant challenge in many bioacoustic studies, and the volume of data in these studies is increasing. Our work makes large scale bird acoustic analyses more feasible by parallelising important bird acoustic processing tasks to significantly reduce execution times.

Saurabh Garg, Tanmay Parekh, Preethi Jyothi

In this work, we present a simple and elegant approach to language modeling for bilingual code-switched text. Since code-switching is a blend of two or more different languages, a standard bilingual language model can be improved upon by using structures of the monolingual language models. We propose a novel technique called dual language models, which involves building two complementary monolingual language models and combining them using a probabilistic model for switching between the two. We evaluate the efficacy of our approach using a conversational Mandarin-English speech corpus. We prove the robustness of our model by showing significant improvements in perplexity measures over the standard bilingual language model without the use of any external information. Similar consistent improvements are also reflected in automatic speech recognition error rates.

Uma Sawant, Saurabh Garg, Soumen Chakrabarti et al.

In Web search, entity-seeking queries often trigger a special Question Answering (QA) system. It may use a parser to interpret the question to a structured query, execute that on a knowledge graph (KG), and return direct entity responses. QA systems based on precise parsing tend to be brittle: minor syntax variations may dramatically change the response. Moreover, KG coverage is patchy. At the other extreme, a large corpus may provide broader coverage, but in an unstructured, unreliable form. We present AQQUCN, a QA system that gracefully combines KG and corpus evidence. AQQUCN accepts a broad spectrum of query syntax, between well-formed questions to short `telegraphic' keyword sequences. In the face of inherent query ambiguities, AQQUCN aggregates signals from KGs and large corpora to directly rank KG entities, rather than commit to one semantic interpretation of the query. AQQUCN models the ideal interpretation as an unobservable or latent variable. Interpretations and candidate entity responses are scored as pairs, by combining signals from multiple convolutional networks that operate collectively on the query, KG and corpus. On four public query workloads, amounting to over 8,000 queries with diverse query syntax, we see 5--16% absolute improvement in mean average precision (MAP), compared to the entity ranking performance of recent systems. Our system is also competitive at entity set retrieval, almost doubling F1 scores for challenging short queries.