Despite the success of large language models (LLMs) on general-purpose tasks, their performance in highly specialized domains such as biomedicine remains unsatisfactory.A key limitation is the inability of LLMs to effectively leverage biomedical tools, which clinical experts and biomedical researchers rely on extensively in daily workflows.While recent general-domain tool-calling datasets have substantially improved the capabilities of LLM agents, existing efforts in the biomedical domain largely rely on in-context learning and restrict models to a small set of tools.To address this gap, we introduce BIOTOOL, a comprehensive biomedical tool-calling dataset designed for fine-tuning LLMs.BIOTOOL comprises 34 frequently used tools collected from the NCBI, Ensembl, and UniProt databases, along with 7,040 high-quality, human-verified query-API call pairs spanning variation, genomics, proteomics, evolution, and general biology.Fine-tuning a 4-billion-parameter LLM on BIOTOOL yields substantial improvements in biomedical tool-calling performance, outperforming cutting-edge commercial LLMs such as GPT-5.1.Furthermore, human expert evaluations demonstrate that integrating a BIOTOOLfine-tuned tool caller significantly improves downstream answer quality compared to the same LLM without tool usage, highlighting the effectiveness of BIOTOOL in enhancing the biomedical capabilities of LLMs.The full dataset and evaluation code are available at https://github.com/gxx27/BioTool.

Large Language Models (LLMs) are widely used for temporal prediction, but their reliance on pretraining data raises contamination concerns, as accurate predictions on pre-cutoff test data may reflect memorization rather than reasoning, leading to an overestimation of their generalization capability.With the recent emergence of prompting-based unlearning techniques, a natural question arises: Can LLMs be prompted to simulate an earlier knowledge cutoff?In this work, we investigate the capability of prompting to simulate earlier knowledge cutoff in LLMs.We construct three evaluation datasets to assess the extent to which LLMs can forget (1) direct factual knowledge, (2) semantic shifts, and (3) causally related knowledge.Results demonstrate that while prompt-based simulated knowledge cutoffs show effectiveness when directly queried with the information after that date, they struggle to induce forgetting when the forgotten content is not directly asked but causally related to the query.These findings highlight the need for more rigorous evaluation settings when applying LLMs for temporal prediction tasks.The full dataset and evaluation code are available at https: //github.com/gxx27/time_unlearn.

Reasoning has substantially improved the performance of large language models (LLMs) on complicated tasks. Central to the current reasoning studies, Process Reward Models (PRMs) offer a fine-grained evaluation of intermediate reasoning steps and guide the reasoning process. However, extending PRMs to multimodal large language models (MLLMs) introduces challenges. Since multimodal reasoning covers a wider range of tasks compared to text-only scenarios, the resulting distribution shift from the training to testing sets is more severe, leading to greater generalization difficulty. Training a reliable multimodal PRM, therefore, demands large and diverse datasets to ensure sufficient coverage. However, current multimodal reasoning datasets suffer from a marked quality imbalance, which degrades PRM performance and highlights the need for an effective data selection strategy. To address the issues, we introduce DreamPRM, a domain-reweighted training framework for multimodal PRMs which employs bi-level optimization. In the lower-level optimization, DreamPRM performs fine-tuning on multiple datasets with domain weights, allowing the PRM to prioritize high-quality reasoning signals and alleviating the impact of dataset quality imbalance. In the upper-level optimization, the PRM is evaluated on a separate meta-learning dataset; this feedback updates the domain weights through an aggregation loss function, thereby improving the generalization capability of trained PRM. Extensive experiments on multiple multimodal reasoning benchmarks covering both mathematical and general reasoning show that test-time scaling with DreamPRM consistently improves the performance of state-of-the-art MLLMs. Further comparisons reveal that DreamPRM's domain-reweighting strategy surpasses other data selection methods and yields higher accuracy gains than existing test-time scaling approaches. Notably, DreamPRM achieves a top-1 accuracy of 85.2% on the MathVista leaderboard using the o4-mini model, demonstrating strong generalization capability in complex multimodal reasoning tasks.

State-of-the-art medical multi-modal LLMs (med-MLLMs), such as LLaVA-Med and BioMedGPT, primarily depend on scaling model size and data volume, with training driven largely by autoregressive objectives. However, we reveal that this approach can lead to weak vision-language alignment, making these models overly dependent on costly instruction-following data. To address this, we introduce ExGra-Med, a novel multi-graph alignment framework that jointly aligns images, instruction responses, and extended captions in the latent space, advancing semantic grounding and cross-modal coherence. To scale to large LLMs (e.g., LLaMa-7B), we develop an efficient end-to-end training scheme using black-box gradient estimation, enabling fast and scalable optimization. Empirically, ExGra-Med matches LLaVA-Med’s performance using just 10\% of pre-training data, achieving a 20.13\% gain on VQA-RAD and approaching full-data performance. It also outperforms strong baselines like BioMedGPT and RadFM on visual chatbot and zero-shot classification tasks, demonstrating its promise for efficient, high-quality vision-language integration in medical AI.

In many image classification applications, the number of labeled training images is limited, which leads to model overfitting. To mitigate the lack of training data, deep generative models have been leveraged to generate synthetic training data. However, existing methods generate data for individual classes based on how much training data they have without considering their actual data needs. To address this limitation, we propose needs-aware image generation, which automatically identifies the different data needs of individual classes based on their classification performance and divides a limited data generation budget into these classes according to their needs. We propose a multi-level optimization based framework which performs four learning stages in an end-to-end manner. Experiments on both imbalanced and balanced classification datasets demonstrate the effectiveness of our proposed method.

Large language models generate high-quality responses with potential misinformation, underscoring the need for regulation by distinguishing AI-generated and human-written texts. Watermarking is pivotal in this context, which involves embedding hidden markers in texts during the LLM inference phase, which is imperceptible to humans. Achieving both the detectability of inserted watermarks and the semantic quality of generated texts is challenging. While current watermarking algorithms have made promising progress in this direction, there remains significant scope for improvement. To address these challenges, we introduce a novel multi-objective optimization (MOO) approach for watermarking that utilizes lightweight networks to generate token-specific watermarking logits and splitting ratios. By leveraging MOO to optimize for both detection and semantic objective functions, our method simultaneously achieves detectability and semantic integrity. Experimental results show that our method outperforms current watermarking techniques in enhancing the detectability of texts generated by LLMs while maintaining their semantic coherence. Our code is available at https://github.com/mignonjia/TS_watermark.

The Segment Anything Model (SAM), a foundation model pretrained on millions of images and segmentation masks, has significantly advanced semantic segmentation, a fundamental task in computer vision. Despite its strengths, SAM encounters two major challenges. Firstly, it struggles with segmenting specific objects autonomously, as it relies on users to manually input prompts like points or bounding boxes to identify targeted objects. Secondly, SAM faces challenges in excelling at specific downstream tasks, like medical imaging, due to a disparity between the distribution of its pretraining data, which predominantly consists of general-domain images, and the data used in downstream tasks. Current solutions to these problems, which involve finetuning SAM, often lead to overfitting, a notable issue in scenarios with very limited data, like in medical imaging. To overcome these limitations, we introduce BLO-SAM, which finetunes SAM based on bi-level optimization (BLO). Our approach allows for automatic image segmentation without the need for manual prompts, by optimizing a learnable prompt embedding. Furthermore, it significantly reduces the risk of overfitting by training the model’s weight parameters and the prompt embedding on two separate subsets of the training dataset, each at a different level of optimization. We apply BLO-SAM to diverse semantic segmentation tasks in general and medical domains. The results demonstrate BLO-SAM’s superior performance over various state-of-the-art image semantic segmentation methods. The code of BLO-SAM is available at https://github.com/importZL/BLO-SAM.

Sai Ashish Somayajula, Youwei Liang, Li Zhang, Abhishek Singh, Pengtao Xie. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Ruiyi Zhang, Rushi Qiang, Sai Ashish Somayajula, Pengtao Xie. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Increasing the throughput of the Transformer architecture, a foundational component used in numerous state-of-the-art models for vision and language tasks (e.g., GPT, LLaVa), is an important problem in machine learning. One recent and effective strategy is to merge token representations within Transformer models, aiming to reduce computational and memory requirements while maintaining accuracy. Prior work has proposed algorithms based on Bipartite Soft Matching (BSM), which divides tokens into distinct sets and merges the top $k$ similar tokens. However, these methods have significant drawbacks, such as sensitivity to token-splitting strategies and damage to informative tokens in later layers. This paper presents a novel paradigm called PiToMe, which prioritizes the preservation of informative tokens using an additional metric termed the \textit{energy score}. This score identifies large clusters of similar tokens as high-energy, indicating potential candidates for merging, while smaller (unique and isolated) clusters are considered as low-energy and preserved. Experimental findings demonstrate that PiToMe saved from 40-60\% FLOPs of the base models while exhibiting superior off-the-shelf performance on image classification (0.5\% average performance drop of ViT-MAEH compared to 2.6\% as baselines), image-text retrieval (0.3\% average performance drop of Clip on Flick30k compared to 4.5\% as others), and analogously in visual questions answering with LLaVa-7B. Furthermore, PiToMe is theoretically shown to preserve intrinsic spectral properties to the original token space under mild conditions.

Abstract In many NLP applications, to mitigate data deficiency in a target task, source data is collected to help with target model training. Existing transfer learning methods either select a subset of source examples that are close to the target domain or try to adapt all source examples into the target domain, then use selected or adapted source examples to train the target model. These methods either incur significant information loss or bear the risk that after adaptation, source examples which are originally already in the target domain may be outside the target domain. To address the limitations of these methods, we propose a four-level optimization based framework which simultaneously selects and adapts source data. Our method can automatically identify in-domain and out-of-domain source examples and apply example-specific processing methods: selection for in-domain examples and adaptation for out-of-domain examples. Experiments on various datasets demonstrate the effectiveness of our proposed method.

Collecting large-scale medical datasets with fully annotated samples for training of deep networks is prohibitively expensive, especially for 3D volume data. Recent breakthroughs in self-supervised learning (SSL) offer the ability to overcome the lack of labeled training samples by learning feature representations from unlabeled data. However, most current SSL techniques in the medical field have been designed for either 2D images or 3D volumes. In practice, this restricts the capability to fully leverage unlabeled data from numerous sources, which may include both 2D and 3D data. Additionally, the use of these pre-trained networks is constrained to downstream tasks with compatible data dimensions. In this paper, we propose a novel framework for unsupervised joint learning on 2D and 3D data modalities. Given a set of 2D images or 2D slices extracted from 3D volumes, we construct an SSL task based on a 2D contrastive clustering problem for distinct classes. The 3D volumes are exploited by computing vectored embedding at each slice and then assembling a holistic feature through deformable self-attention mechanisms in Transformer, allowing incorporating long-range dependencies between slices inside 3D volumes. These holistic features are further utilized to define a novel 3D clustering agreement-based SSL task and masking embedding prediction inspired by pre-trained language models. Experiments on downstream tasks, such as 3D brain segmentation, lung nodule detection, 3D heart structures segmentation, and abnormal chest X-ray detection, demonstrate the effectiveness of our joint 2D and 3D SSL approach. We improve plain 2D Deep-ClusterV2 and SwAV by a significant margin and also surpass various modern 2D and 3D SSL approaches.

The integration of machine learning models in various real-world applications is becoming more prevalent to assist humans in their daily decision-making tasks as a result of recent advancements in this field. However, it has been discovered that there is a tradeoff between the accuracy and fairness of these decision-making tasks. In some cases, these AI systems can be unfair by exhibiting bias or discrimination against certain social groups, which can have severe consequences in real life. Inspired by one of the most well-known human learning skills called grouping, we address this issue by proposing a novel machine learning (ML) framework where the ML model learns to group a diverse set of problems into distinct subgroups to solve each subgroup using its specific sub-model. Our proposed framework involves three stages of learning, which are formulated as a three-level optimization problem: 1) grouping problems into subgroups, 2) learning group-specific sub-models for problem-solving, and 3) updating group assignments of training examples by minimizing validation loss. These three learning stages are performed end-to-end in a joint manner using gradient descent. To improve fairness and accuracy, we develop an efficient optimization algorithm to solve this three-level optimization problem. To further decrease the risk of overfitting in small datasets using our LBG method, we incorporate domain adaptation techniques in the second stage of training. We further apply our method to differentiable neural architecture search (NAS) methods.

Finding the optimal pass sequence of compilation can lead to a significant reduction in program size. Prior works on compilation pass ordering have two major drawbacks. They either require an excessive budget (in terms of the number of compilation passes) at compile time or fail to generalize to unseen programs. In this work, instead of predicting passes sequentially, we directly learn a policy on the pass sequence space, which outperforms the default -Oz flag by an average of 4.5% over a large collection (4683) of unseen code repositories from diverse domains across 14 datasets. To achieve this, we first identify a small set (termed coreset) of pass sequences that generally optimize the size of most programs. Then, a policy is learned to pick the optimal sequences by predicting the normalized values of the pass sequences in the coreset. Our results demonstrate that existing human-designed compiler passes can be improved with a simple yet effective technique that leverages pass sequence space which contains dense rewards, while approaches operating on the individual pass space may suffer from issues of sparse reward, and do not generalize well to held-out programs from different domains. Website: https://rlcompopt.github.io.

In humans' classroom learning, many effective study techniques (e.g., the Feynman technique, peer questioning, etc.) have been developed to improve learning outcomes. We are interested in investigating whether these techniques can inspire the development of ML training strategies to improve bi-level optimization (BLO) based methods. Towards this goal, we develop a general framework, Skillearn, which consists of basic elements such as learners, interaction functions, learning stages, etc. These elements can be flexibly configured to create various training strategies, each emulating a study technique of humans. In case studies, we apply Skillearn to create new training strategies, by emulating the Feynman technique and peer questioning, which are two broadly adopted techniques in humans' classroom learning. These training strategies are used for improving two BLO based applications including neural architecture search and data weighting. Experiments on various datasets demonstrate the effectiveness of our methods.

Despite its flexibility to learn diverse inductive biases in machine learning programs, meta learning (i.e.,\ learning to learn) has long been recognized to suffer from poor scalability due to its tremendous compute/memory costs, training instability, and a lack of efficient distributed training support. In this work, we focus on making scalable meta learning practical by introducing SAMA, which combines advances in both implicit differentiation algorithms and systems. Specifically, SAMA is designed to flexibly support a broad range of adaptive optimizers in the base level of meta learning programs, while reducing computational burden by avoiding explicit computation of second-order gradient information, and exploiting efficient distributed training techniques implemented for first-order gradients. Evaluated on multiple large-scale meta learning benchmarks, SAMA showcases up to 1.7/4.8x increase in throughput and 2.0/3.8x decrease in memory consumption respectively on single-/multi-GPU setups compared to other baseline meta learning algorithms. Furthermore, we show that SAMA-based data optimization leads to consistent improvements in text classification accuracy with BERT and RoBERTa large language models, and achieves state-of-the-art results in both small- and large-scale data pruning on image classification tasks, demonstrating the practical applicability of scalable meta learning across language and vision domains.

Obtaining large pre-trained models that can be fine-tuned to new tasks with limited annotated samples has remained an open challenge for medical imaging data. While pre-trained networks on ImageNet and vision-language foundation models trained on web-scale data are the prevailing approaches, their effectiveness on medical tasks is limited due to the significant domain shift between natural and medical images. To bridge this gap, we introduce LVM-Med, the first family of deep networks trained on large-scale medical datasets. We have collected approximately 1.3 million medical images from 55 publicly available datasets, covering a large number of organs and modalities such as CT, MRI, X-ray, and Ultrasound. We benchmark several state-of-the-art self-supervised algorithms on this dataset and propose a novel self-supervised contrastive learning algorithm using a graph-matching formulation. The proposed approach makes three contributions: (i) it integrates prior pair-wise image similarity metrics based on local and global information; (ii) it captures the structural constraints of feature embeddings through a loss function constructed through a combinatorial graph-matching objective, and (iii) it can be trained efficiently end-to-end using modern gradient-estimation techniques for black-box solvers. We thoroughly evaluate the proposed LVM-Med on 15 downstream medical tasks ranging from segmentation and classification to object detection, and both for the in and out-of-distribution settings. LVM-Med empirically outperforms a number of state-of-the-art supervised, self-supervised, and foundation models. For challenging tasks such as Brain Tumor Classification or Diabetic Retinopathy Grading, LVM-Med improves previous vision-language models trained on 1 billion masks by 6-7% while using only a ResNet-50.

Driverdrowsiness is an important cause of traffic accidents. Many studies using computer vision techniques to detect driver drowsiness states, such as slow blinking, yawning, and nodding, have demonstrated excellent potential. Although existing studies have made significant progress, the number of samples in the training corpora is small, which makes it difficult for a model to learn effective drowsiness representations from images or videos. To address this issue, we develop an isotropic self-supervised learning (IsoSSL) approach to learn powerful representations of images without relying on human-provided annotations and propose an IsoSSL-MoCo model by combining IsoSSL with momentum contrast (MoCo). To exploit the complementarity of multimodal data, an attention-based multimodal fusion model is also proposed to fuse features from the eye, mouth, and optical flow of the head. Specifically, we first use the IsoSSL-MoCo model to pretrain the image encoders for the three modalities in other datasets. Then, these encoders are fine-tuned and integrated into the proposed fusion model. The feature vectors generated by the image encoders of the three modalities are fed into the recursive layer to extract temporal information. To capture the importance degrees of the effects of temporal features from the three modalities on drowsiness detection, an attention mechanism is introduced to automatically weigh the feature vectors from the recursive layer to improve detection accuracy. Finally, a vector representation is generated by the attention layer and is used to detect driver drowsiness states. Experimental results based on two challenging datasets show that our method outperforms the baseline methods and the latest existing methods.

Event detection in power systems aims to identify triggers and event types, which helps relevant personnel respond to emergencies promptly and facilitates the optimization of power supply strategies. However, the limited length of short electrical record texts causes severe information sparsity, and numerous domain-specific terminologies of power systems makes it difficult to transfer knowledge from language models pre-trained on general-domain texts. Traditional event detection approaches primarily focus on the general domain and ignore these two problems in the power system domain. To address the above issues, we propose a Multi-Channel graph neural network utilizing Type information for Event Detection in power systems, named MC-TED , leveraging a semantic channel and a topological channel to enrich information interaction from short texts. Concretely, the semantic channel refines textual representations with semantic similarity, building the semantic information interaction among potential event-related words. The topological channel generates a relation-type-aware graph modeling word dependencies, and a word-type-aware graph integrating part-of-speech tags. To further reduce errors worsened by professional terminologies in type analysis, a type learning mechanism is designed for updating the representations of both the word type and relation type in the topological channel. In this way, the information sparsity and professional term occurrence problems can be alleviated by enabling interaction between topological and semantic information. Furthermore, to address the lack of labeled data in power systems, we built a Chinese event detection dataset based on electrical Power Event texts, named PoE . In experiments, our model achieves compelling results not only on the PoE dataset, but on general-domain event detection datasets including ACE 2005 and MAVEN.

Learning from one's mistakes is an effective human learning technique where the learners focus more on the topics where mistakes were made, so as to deepen their understanding. In this paper, we investigate if this human learning strategy can be applied in machine learning. We propose a novel machine learning method called Learning From Mistakes (LFM), wherein the learner improves its ability to learn by focusing more on the mistakes during revision. We formulate LFM as a three-stage optimization problem: 1) learner learns; 2) learner re-learns focusing on the mistakes, and; 3) learner validates its learning. We develop an efficient algorithm to solve the LFM problem. We apply the LFM framework to neural architecture search on CIFAR-10, CIFAR-100, and Imagenet. Experimental results strongly demonstrate the effectiveness of our model.

Knowledge distillation has shown great effectiveness for improving neural architecture search (NAS). Mutual knowledge distillation (MKD), where a group of models mutually generate knowledge to train each other, has achieved promising results in many applications. In existing MKD methods, mutual knowledge distillation is performed between models without scrutiny: a worse-performing model is allowed to generate knowledge to train a better-performing model, which may lead to collective failures. To address this problem, we propose a performance-aware MKD (PAMKD) approach for NAS, where knowledge generated by model$A$is allowed to train model$B$only if the performance of$A$is better than B. We propose a three-level optimization framework to formulate PAMKD, where three learning stages are performed end-to-end: 1) each model trains an initial model independently; 2) the initial models are evaluated on a validation set and better-performing models generate knowledge to train worse-performing models; 3) architectures are updated by minimizing a validation loss. Experimental results on a variety of datasets demonstrate that our method is effective.

Graph neural architecture search has shown great potentials for automatically designing graph neural network (GNN) architectures for graph classification tasks. However, when there is a distribution shift between training and testing graphs, the existing approaches fail to deal with the problem of adapting to unknown test graph structures since they only search for a fixed architecture for all graphs. To solve this problem, we propose a novel GRACES model which is able to generalize under distribution shifts through tailoring a customized GNN architecture suitable for each graph instance with unknown distribution. Specifically, we design a self-supervised disentangled graph encoder to characterize invariant factors hidden in diverse graph structures. Then, we propose a prototype-based architecture customization strategy to generate the most suitable GNN architecture weights in a continuous space for each graph instance. We further propose a customized super-network to share weights among different architectures for the sake of efficient training. Extensive experiments on both synthetic and real-world datasets demonstrate that our proposed GRACES model can adapt to diverse graph structures and achieve state-of-the-art performance for graph classification tasks under distribution shifts.

In deep learning applications, image understanding is a crucial task, where several techniques such as image captioning and visual question answering have been widely studied to improve and evaluate the performances of deep neural networks (DNN) in this area. In image captioning, models have encoder-decoder architectures, where the encoders take the input images, produce embeddings, and feed them into the decoders to generate textual descriptions. Designing a proper image captioning encoder-decoder architecture manually is a difficult challenge due to the complexity of recognizing the critical objects of the input images and their relationships to generate caption descriptions. To address this issue, we propose a three-level optimization method that employs differentiable architecture search strategies to seek the most suitable architecture for image captioning automatically. Our optimization framework involves three stages, which are performed end-to-end. In the first stage, an image captioning model learns and updates the weights of its encoder and decoder to create image captions. At the next stage, the trained encoder-decoder generates a pseudo image captioning dataset from unlabeled images, and the predictive model trains on the generated dataset to update its weights. Finally, the trained model validates its performance on the validation set and updates the encoder-decoder architecture by minimizing the validation loss. Experiments and studies on the COCO image captions datasets demonstrate that our method performs significantly better than the baselines and can achieve state-of-the-art results in image understanding tasks.

Recently a wide variety of NAS methods have been proposed and achieved considerable success in automatically identifying highly-performing architectures of neural networks for the sake of reducing the reliance on human experts. Existing NAS methods ignore the fact that different input data elements (e.g., image pixels) have different importance (or saliency) in determining the prediction outcome. They treat all data elements as being equally important and therefore lead to suboptimal performance. To address this problem, we propose an end-to-end framework which dynamically detects saliency of input data, reweights data using saliency maps, and searches architectures on saliency-reweighted data. Our framework is based on four-level optimization, which performs four learning stages in a unified way. At the first stage, a model is trained with its architecture tentatively fixed. At the second stage, saliency maps are generated using the trained model. At the third stage, the model is retrained on saliency-reweighted data. At the fourth stage, the model is evaluated on a validation set and the architecture is updated by minimizing the validation loss. Experiments on several datasets demonstrate the effectiveness of our framework.

Abstract Self-supervised learning (SSL) methods such as Word2vec, BERT, and GPT have shown great effectiveness in language understanding. Contrastive learning, as a recent SSL approach, has attracted increasing attention in NLP. Contrastive learning learns data representations by predicting whether two augmented data instances are generated from the same original data example. Previous contrastive learning methods perform data augmentation and contrastive learning separately. As a result, the augmented data may not be optimal for contrastive learning. To address this problem, we propose a four-level optimization framework that performs data augmentation and contrastive learning end-to-end, to enable the augmented data to be tailored to the contrastive learning task. This framework consists of four learning stages, including training machine translation models for sentence augmentation, pretraining a text encoder using contrastive learning, finetuning a text classification model, and updating weights of translation data by minimizing the validation loss of the classification model, which are performed in a unified way. Experiments on datasets in the GLUE benchmark (Wang et al., 2018a) and on datasets used in Gururangan et al. (2020) demonstrate the effectiveness of our method.

Abstract Text augmentation is an effective technique in alleviating overfitting in NLP tasks. In existing methods, text augmentation and downstream tasks are mostly performed separately. As a result, the augmented texts may not be optimal to train the downstream model. To address this problem, we propose a three-level optimization framework to perform text augmentation and the downstream task end-to- end. The augmentation model is trained in a way tailored to the downstream task. Our framework consists of three learning stages. A text summarization model is trained to perform data augmentation at the first stage. Each summarization example is associated with a weight to account for its domain difference with the text classification data. At the second stage, we use the model trained at the first stage to perform text augmentation and train a text classification model on the augmented texts. At the third stage, we evaluate the text classification model trained at the second stage and update weights of summarization examples by minimizing the validation loss. These three stages are performed end-to-end. We evaluate our method on several text classification datasets where the results demonstrate the effectiveness of our method. Code is available at https://github.com/Sai-Ashish/End-to-End-Text-Augmentation.

Interpretable machine learning has gained much attention recently. Briefness and comprehensiveness are necessary in order to provide a large amount of information concisely when explaining a black-box decision system. However, existing interpretable machine learning methods fail to consider briefness and comprehensiveness simultaneously, leading to redundant explanations. We propose the variational information bottleneck for interpretation, VIBI, a system-agnostic interpretable method that provides a brief but comprehensive explanation. VIBI adopts an information theoretic principle, information bottleneck principle, as a criterion for finding such explanations. For each instance, VIBI selects key features that are maximally compressed about an input (briefness), and informative about a decision made by a black-box system on that input (comprehensive). We evaluate VIBI on three datasets and compare with state-of-the-art interpretable machine learning methods in terms of both interpretability and fidelity evaluated by human and quantitative metrics.

Graph classification is a widely studied problem and has broad applications. In many real-world problems, the number of labeled graphs available for training classification models is limited, which renders these models prone to overfitting. To address this problem, we propose two approaches based on contrastive self-supervised learning (CSSL) to alleviate overfitting. In the first approach, we use CSSL to pretrain graph encoders on widely-available unlabeled graphs without relying on human-provided labels, then finetune the pretrained encoders on labeled graphs. In the second approach, we develop a regularizer based on CSSL, and solve the supervised classification task and the unsupervised CSSL task simultaneously. To perform CSSL on graphs, given a collection of original graphs, we perform data augmentation to create augmented graphs out of the original graphs. An augmented graph is created by consecutively applying a sequence of graph alteration operations. A contrastive loss is defined to learn graph encoders by judging whether two augmented graphs are from the same original graph. Experiments on various graph classification datasets demonstrate the effectiveness of our proposed methods. The code is available at https://github.com/UCSD-AI4H/GraphSSL.

In deep learning applications, the architectures of deep neural networks are crucial in achieving high accuracy. Many methods have been proposed to search for high-performance neural architectures automatically. However, these searched architectures are prone to adversarial attacks. A small perturbation of the input data can render the architecture to change prediction outcomes significantly. To address this problem, we propose methods to perform differentiable search of robust neural architectures. In our methods, two differentiable metrics are defined to measure architectures’ robustness, based on certified lower bound and Jacobian norm bound. Then we search for robust architectures by maximizing the robustness metrics. Different from previous approaches which aim to improve architectures’ robustness in an implicit way: performing adversarial training and injecting random noise, our methods explicitly and directly maximize robustness metrics to harvest robust architectures. On CIFAR-10, ImageNet, and MNIST, we perform game-based evaluation and verification-based evaluation on the robustness of our methods. The experimental results show that our methods 1) are more robust to various norm-bound attacks than several robust NAS baselines; 2) are more accurate than baselines when there are no attacks; 3) have significantly higher certified lower bounds than baselines.

Abstract To develop commonsense-grounded NLP applications, a comprehensive and accurate commonsense knowledge graph (CKG) is needed. It is time-consuming to manually construct CKGs and many research efforts have been devoted to the automatic construction of CKGs. Previous approaches focus on generating concepts that have direct and obvious relationships with existing concepts and lack an capability to generate unobvious concepts. In this work, we aim to bridge this gap. We propose a general graph-to-paths pretraining framework that leverages high-order structures in CKGs to capture high-order relationships between concepts. We instantiate this general framework to four special cases: long path, path-to-path, router, and graph-node-path. Experiments on two datasets demonstrate the effectiveness of our methods. The code will be released via the public GitHub repository.

Abstract Text classification is a widely studied problem and has broad applications. In many real-world problems, the number of texts for training classification models is limited, which renders these models prone to overfitting. To address this problem, we propose SSL-Reg, a data-dependent regularization approach based on self-supervised learning (SSL). SSL (Devlin et al., 2019a) is an unsupervised learning approach that defines auxiliary tasks on input data without using any human-provided labels and learns data representations by solving these auxiliary tasks. In SSL-Reg, a supervised classification task and an unsupervised SSL task are performed simultaneously. The SSL task is unsupervised, which is defined purely on input texts without using any human- provided labels. Training a model using an SSL task can prevent the model from being overfitted to a limited number of class labels in the classification task. Experiments on 17 text classification datasets demonstrate the effectiveness of our proposed method. Code is available at https://github.com/UCSD-AI4H/SSReg.

Adversarial training is a useful approach to promote the learning of transferable representations across the source and target domains, which has been widely applied for domain adaptation (DA) tasks based on deep neural networks. Until very recently, existing adversarial domain adaptation (ADA) methods ignore the useful information from the label space, which is an important factor accountable for the complicated data distributions associated with different semantic classes. Especially, the inter-class semantic relationships have been rarely considered and discussed in the current work of transfer learning. In this paper, we propose a novel relationship-aware adversarial domain adaptation (RADA) algorithm, which first utilizes a single multi-class domain discriminator to enforce the learning of inter-class dependency structure during domain-adversarial training and then aligns this structure with the inter-class dependencies that are characterized from training the label predictor on source domain. Specifically, we impose a regularization term to penalize the structure discrepancy between the inter-class dependencies respectively estimated from domain discriminator and label predictor. Through this alignment, our proposed method makes the adversarial domain adaptation aware of the class relationships. Empirical studies show that the incorporation of class relationships significantly improves the performance on benchmark datasets.

Guangtao Zeng, Wenmian Yang, Zeqian Ju, Yue Yang, Sicheng Wang, Ruisi Zhang, Meng Zhou, Jiaqi Zeng, Xiangyu Dong, Ruoyu Zhang, Hongchao Fang, Penghui Zhu, Shu Chen, Pengtao Xie. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.

The International Classification of Diseases (ICD) is a list of classification codes for the diagnoses. Automatic ICD coding is a multi-label text classification problem with noisy clinical document inputs and long-tailed label distribution, making it difficult for fine-grained classification on both frequent and zero-shot codes at the same time, i.e. generalized zero-shot ICD coding. In this paper, we propose a latent feature generation framework to improve the prediction on unseen codes without compromising the performance on seen codes. Our framework generates semantically meaningful features for zero-shot codes by exploiting ICD code hierarchical structure and reconstructing the code-relevant keywords with a novel cycle architecture. To the best of our knowledge, this is the first adversarial generative model for generalized zero-shot learning on multi-label text classification. Extensive experiments demonstrate the effectiveness of our approach. On the public MIMIC-III dataset, our methods improve the F1 score from nearly 0 to 20.91% for the zero-shot codes, and increase the AUC score by 3% (absolute improvement) from previous state of the art. Code is available at https://github.com/csong27/gzsl_text.

State-of-the-art approaches to causal discovery usually assume a fixed underlying causal model. However, it is often the case that causal models vary across domains or subjects, due to possibly omitted factors that affect the quantitative causal effects. As a typical example, causal connectivity in the brain network has been reported to vary across individuals, with significant differences across groups of people, such as autistics and typical controls. In this paper, we develop a unified framework for causal discovery and mechanism-based group identification. In particular, we propose a specific and shared causal model (SSCM), which takes into account the variabilities of causal relations across individuals/groups and leverages their commonalities to achieve statistically reliable estimation. The learned SSCM gives the specific causal knowledge for each individual as well as the general trend over the population. In addition, the estimated model directly provides the group information of each individual. Experimental results on synthetic and real-world data demonstrate the efficacy of the proposed method.

The International Classification of Diseases (ICD) provides a hierarchy of diagnostic codes for classifying diseases.Medical coding -which assigns a subset of ICD codes to a patient visit -is a mandatory process that is crucial for patient care and billing.Manual coding is time-consuming, expensive, and errorprone.In this paper, we build a neural architecture for automated coding.It takes the diagnosis descriptions (DDs) of a patient as inputs and selects the most relevant ICD codes.This architecture contains four major ingredients: (1) tree-ofsequences LSTM encoding of code descriptions (CDs), (2) adversarial learning for reconciling the different writing styles of DDs and CDs, (3) isotonic constraints for incorporating the importance order among the assigned codes, and (4) attentional matching for performing many-toone and one-to-many mappings from DDs to CDs.We demonstrate the effectiveness of the proposed methods on a clinical datasets with 59K patient visits.

Medical imaging is widely used in clinical practice for diagnosis and treatment.Report-writing can be error-prone for unexperienced physicians, and timeconsuming and tedious for experienced physicians.To address these issues, we study the automatic generation of medical imaging reports.This task presents several challenges.First, a complete report contains multiple heterogeneous forms of information, including findings and tags.Second, abnormal regions in medical images are difficult to identify.Third, the reports are typically long, containing multiple sentences.To cope with these challenges, we (1) build a multi-task learning framework which jointly performs the prediction of tags and the generation of paragraphs, (2) propose a co-attention mechanism to localize regions containing abnormalities and generate narrations for them, (3) develop a hierarchical LSTM model to generate long paragraphs.We demonstrate the effectiveness of the proposed methods on two publicly available datasets.

Numerous existing works have shown that, key to the efficiency of distributed machine learning (ML) is proper system and algorithm co-design: system design should be tailored to the unique mathematical properties of ML algorithms, and algorithms can be re-designed to better exploit the system architecture. While existing research has made attempts along this direction, many algorithmic and system properties that are characteristic of ML problems remain to be explored. Through an exploration of system-algorithm co-design, we build a new decentralized system Orpheus to support distributed training of a general class of ML models whose parameters are represented with large matrices. Training such models at scale is challenging: transmitting and checkpointing large matrices incur substantial network traffic and disk IO, which aggravates the inconsistency among parameter replicas. To cope with these challenges, Orpheus jointly exploits system and algorithm designs which (1) reduce the size and number of network messages for efficient communication, 2) incrementally checkpoint vectors for light-weight and fine-grained fault tolerance without blocking computation, 3) improve the consistency among parameter copies via periodic centralized synchronization and parameter-replicas rotation. As a result of these co-designs, communication and fault tolerance costs are linear to both matrix dimension and number of machines in the network, as opposed to being quadratic in existing systems. And the improved parameter consistency accelerates algorithmic convergence. Empirically, we show our system outperforms several existing baseline systems on training several representative large-scale ML models.

Distance metric learning (DML), which learns a distance metric from labeled "similar" and "dissimilar" data pairs, is widely utilized. Recently, several works investigate orthogonality-promoting regularization (OPR), which encourages the projection vectors in DML to be close to being orthogonal, to achieve three effects: (1) high balancedness – achieving comparable performance on both frequent and infrequent classes; (2) high compactness – using a small number of projection vectors to achieve a "good" metric; (3) good generalizability – alleviating overfitting to training data. While showing promising results, these approaches suffer three problems. First, they involve solving non-convex optimization problems where achieving the global optimal is NP-hard. Second, it lacks a theoretical understanding why OPR can lead to balancedness. Third, the current generalization error analysis of OPR is not directly on the regularizer. In this paper, we address these three issues by (1) seeking convex relaxations of the original nonconvex problems so that the global optimal is guaranteed to be achievable; (2) providing a formal analysis on OPR’s capability of promoting balancedness; (3) providing a theoretical analysis that directly reveals the relationship between OPR and generalization performance. Experiments on various datasets demonstrate that our convex methods are more effective in promoting balancedness, compactness, and generalization, and are computationally more efficient, compared with the nonconvex methods.

Variable selection is a classic problem in machine learning (ML), widely used to find important explanatory factors, and improve generalization performance and interpretability of ML models. In this paper, we consider variable selection for models where multiple responses are to be predicted based on the same set of covariates. Since each response is relevant to a unique subset of covariates, we desire the selected variables for different responses have small overlap. We propose a regularizer that simultaneously encourage orthogonality and sparsity, which jointly brings in an effect of reducing overlap. We apply this regularizer to four model instances and develop efficient algorithms to solve the regularized problems. We provide a formal analysis on why the proposed regularizer can reduce generalization error. Experiments on both simulation studies and real-world datasets demonstrate the effectiveness of the proposed regularizer in selecting less-overlapped variables and improving generalization performance.

Reading comprehension (RC), aiming to understand natural texts and answer questions therein, is a challenging task.In this paper, we study the RC problem on the Stanford Question Answering Dataset (SQuAD).Observing from the training set that most correct answers are centered around constituents in the parse tree, we design a constituent-centric neural architecture where the generation of candidate answers and their representation learning are both based on constituents and guided by the parse tree.Under this architecture, the search space of candidate answers can be greatly reduced without sacrificing the coverage of correct answers and the syntactic, hierarchical and compositional structure among constituents can be well captured, which contributes to better representation learning of the candidate answers.On SQuAD, our method achieves the state of the art performance and the ablation study corroborates the effectiveness of individual modules.

We study large-scale multi-label classification (MLC) on two recently released datasets: Youtube-8M and Open Images that contain millions of data instances and thousands of classes. The unprecedented problem scale poses great challenges for MLC. First, finding out the correct label subset out of exponentially many choices incurs substantial ambiguity and uncertainty. Second, the large data-size and class-size entail considerable computational cost. To address the first challenge, we investigate two strategies: capturing label-correlations from the training data and incorporating label co-occurrence relations obtained from external knowledge, which effectively eliminate semantically inconsistent labels and provide contextual clues to differentiate visually ambiguous labels. Specifically, we propose a Deep Determinantal Point Process (DDPP) model which seamlessly integrates a DPP with deep neural networks (DNNs) and supports end-to-end multi-label learning and deep representation learning. The DPP is able to capture label-correlations of any order with a polynomial computational cost, while the DNNs learn hierarchical features of images/videos and capture the dependency between input data and labels. To incorporate external knowledge about label co-occurrence relations, we impose relational regularization over the kernel matrix in DDPP. To address the second challenge, we study an efficient low-rank kernel learning algorithm based on inducing point methods. Experiments on the two datasets demonstrate the efficacy and efficiency of the proposed methods.

The large model capacity of latent space models (LSMs) enables them to achieve great performance on various applications, but meanwhile renders LSMs to be prone to overfitting. Several recent studies investigate a new type of regularization approach, which encourages components in LSMs to be diverse, for the sake of alleviating overfitting. While they have shown promising empirical effectiveness, in theory why larger “diversity” results in less overfitting is still unclear. To bridge this gap, we propose a new diversity-promoting approach that is both theoretically analyzable and empirically effective. Specifically, we use near-orthogonality to characterize “diversity” and impose angular constraints (ACs) on the components of LSMs to promote diversity. A generalization error analysis shows that larger diversity results in smaller estimation error and larger approximation error. An efficient ADMM algorithm is developed to solve the constrained LSM problems. Experiments demonstrate that ACs improve generalization performance of LSMs and outperform other diversity-promoting approaches.

Latent space models (LSMs) provide a principled and effective way to extract hidden patterns from observed data. To cope with two challenges in LSMs: (1) how to capture infrequent patterns when pattern frequency is imbalanced and (2) how to reduce model size without sacrificing their expressiveness, several studies have been proposed to “diversify” LSMs, which design regularizers to encourage the components therein to be “diverse”. In light of the limitations of existing approaches, we design a new diversity-promoting regularizer by considering two factors: uncorrelation and evenness, which encourage the components to be uncorrelated and to play equally important roles in modeling data. Formally, this amounts to encouraging the covariance matrix of the components to have more uniform eigenvalues. We apply the regularizer to two LSMs and develop an efficient optimization algorithm. Experiments on healthcare, image and text data demonstrate the effectiveness of the regularizer.

In multi-class support vector machine (MSVM) for classification, one core issue is to regularize the coefficient vectors to reduce overfitting. Various regularizers have been proposed such as L2, L1, and trace norm. In this paper, we introduce a new type of regularization approach -- angular regularization, that encourages the coefficient vectors to have larger angles such that class regions can be widen to flexibly accommodate unseen samples. We propose a novel angular regularizer based on the singular values of the coefficient matrix, where the uniformity of singular values reduces the correlation among different classes and drives the angles between coefficient vectors to increase. In generalization error analysis, we show that decreasing this regularizer effectively reduces generalization error bound. On various datasets, we demonstrate the efficacy of the regularizer in reducing overfitting.

In learning latent variable models (LVMs), it is important to effectively capture infrequent patterns and shrink model size without sacrificing modeling power. Various studies have been done to “diversify” a LVM, which aim to learn a diverse set of latent components in LVMs. Most existing studies fall into a frequentist-style regularization framework, where the components are learned via point estimation. In this paper, we investigate how to “diversify” LVMs in the paradigm of Bayesian learning, which has advantages complementary to point estimation, such as alleviating overfitting via model averaging and quantifying uncertainty. We propose two approaches that have complementary advantages. One is to define diversity-promoting mutual angular priors which assign larger density to components with larger mutual angles based on Bayesian network and von Mises-Fisher distribution and use these priors to affect the posterior via Bayes rule. We develop two efficient approximate posterior inference algorithms based on variational inference and Markov chain Monte Carlo sampling. The other approach is to impose diversity-promoting regularization directly over the post-data distribution of components. These two methods are applied to the Bayesian mixture of experts model to encourage the “experts” to be diverse and experimental results demonstrate the effectiveness and efficiency of our methods.

Personal photos are enjoying explosive growth with the popularity of photo-taking devices and social media. The vast amount of online photos largely exhibit users' interests, emotion and opinions. Mining user interests from personal photos can boost a number of utilities, such as advertising, interest based community detection and photo recommendation. In this paper, we study the problem of user interests mining from personal photos. We propose a User Image Latent Space Model to jointly model user interests and image contents. User interests are modeled as latent factors and each user is assumed to have a distribution over them. By inferring the latent factors and users' distributions, we can discover what the users are interested in. We model image contents with a four-level hierarchical structure where the layers correspond to themes, semantic regions, visual words and pixels respectively. Users' latent interests are embedded in the theme layer. Given image contents, users' interests can be discovered by doing posterior inference. We use variational inference to approximate the posteriors of latent variables and learn model parameters. Experiments on 180K Flickr photos demonstrate the effectiveness of our model.

Image clustering and visual codebook learning are two fundamental problems in computer vision and they are tightly related. On one hand, a good codebook can generate effective feature representations which largely affect clustering performance. On the other hand, class labels obtained from image clustering can serve as supervised information to guide codebook learning. Traditionally, these two processes are conducted separately and their correlation is generally ignored.In this paper, we propose a Double Layer Gaussian Mixture Model (DLGMM) to simultaneously perform image clustering and codebook learning. In DLGMM, two tasks are seamlessly coupled and can mutually promote each other. Cluster labels and codebook are jointly estimated to achieve the overall best performance. To incorporate the spatial coherence between neighboring visual patches, we propose a Spatially Coherent DLGMM which uses a Markov Random Field to encourage neighboring patches to share the same visual word label.We use variational inference to approximate the posterior of latent variables and learn model parameters.Experiments on two datasets demonstrate the effectiveness of two models.

Restricted Boltzmann Machine (RBM) has shown great effectiveness in document modeling. It utilizes hidden units to discover the latent topics and can learn compact semantic representations for documents which greatly facilitate document retrieval, clustering and classification. The popularity (or frequency) of topics in text corpora usually follow a power-law distribution where a few dominant topics occur very frequently while most topics (in the long-tail region) have low probabilities. Due to this imbalance, RBM tends to learn multiple redundant hidden units to best represent dominant topics and ignore those in the long-tail region, which renders the learned representations to be redundant and non-informative. To solve this problem, we propose Diversified RBM (DRBM) which diversifies the hidden units, to make them cover not only the dominant topics, but also those in the long-tail region. We define a diversity metric and use it as a regularizer to encourage the hidden units to be diverse. Since the diversity metric is hard to optimize directly, we instead optimize its lower bound and prove that maximizing the lower bound with projected gradient ascent can increase this diversity metric. Experiments on document retrieval and clustering demonstrate that with diversification, the document modeling power of DRBM can be greatly improved.

How can one build a distributed framework that allows efficient deployment of a wide spectrum of modern advanced machine learning (ML) programs for industrial-scale problems using Big Models (100s of billions of parameters) on Big Data (terabytes or petabytes)- Contemporary parallelization strategies employ fine-grained operations and scheduling beyond the classic bulk-synchronous processing paradigm popularized by MapReduce, or even specialized operators relying on graphical representations of ML programs. The variety of approaches tends to pull systems and algorithms design in different directions, and it remains difficult to find a universal platform applicable to a wide range of different ML programs at scale. We propose a general-purpose framework that systematically addresses data- and model-parallel challenges in large-scale ML, by leveraging several fundamental properties underlying ML programs that make them different from conventional operation-centric programs: error tolerance, dynamic structure, and nonuniform convergence; all stem from the optimization-centric nature shared in ML programs' mathematical definitions, and the iterative-convergent behavior of their algorithmic solutions. These properties present unique opportunities for an integrative system design, built on bounded-latency network synchronization and dynamic load-balancing scheduling, which is efficient, programmable, and enjoys provable correctness guarantees. We demonstrate how such a design in light of ML-first principles leads to significant performance improvements versus well-known implementations of several ML programs, allowing them to run in much less time and at considerably larger model sizes, on modestly-sized computer clusters.

This paper studies how to incorporate the external word correlation knowledge to improve the coherence of topic modeling.Existing topic models assume words are generated independently and lack the mechanism to utilize the rich similarity relationships among words to learn coherent topics.To solve this problem, we build a Markov Random Field (MRF) regularized Latent Dirichlet Allocation (LDA) model, which defines a MRF on the latent topic layer of LDA to encourage words labeled as similar to share the same topic label.Under our model, the topic assignment of each word is not independent, but rather affected by the topic labels of its correlated words.Similar words have better chance to be put into the same topic due to the regularization of MRF, hence the coherence of topics can be boosted.In addition, our model can accommodate the subtlety that whether two words are similar depends on which topic they appear in, which allows word with multiple senses to be put into different topics properly.We derive a variational inference method to infer the posterior probabilities and learn model parameters and present techniques to deal with the hardto-compute partition function in MRF.Experiments on two datasets demonstrate the effectiveness of our model.

What is a systematic way to efficiently apply a wide spectrum of advanced ML programs to industrial scale problems, using Big Models (up to 100 s of billions of parameters) on Big Data (up to terabytes or petabytes)? Modern parallelization strategies employ fine-grained operations and scheduling beyond the classic bulk-synchronous processing paradigm popularized by MapReduce, or even specialized graph-based execution that relies on graph representations of ML programs. The variety of approaches tends to pull systems and algorithms design in different directions, and it remains difficult to find a universal platform applicable to a wide range of ML programs at scale. We propose a general-purpose framework, Petuum, that systematically addresses data- and model-parallel challenges in large-scale ML, by observing that many ML programs are fundamentally optimization-centric and admit error-tolerant, iterative-convergent algorithmic solutions. This presents unique opportunities for an integrative system design, such as bounded-error network synchronization and dynamic scheduling based on ML program structure. We demonstrate the efficacy of these system designs versus well-known implementations of modern ML algorithms, showing that Petuum allows ML programs to run in much less time and at considerably larger model sizes, even on modestly-sized compute clusters.