Diffusion models excel in high-fidelity image generation but face scalability limits due to transformers’ quadratic attention complexity. Plug-and-play token reduction methods like ToMeSD and ToFu reduce FLOPs by merging redundant tokens in generated images but rely on GPU-inefficient operations (e.g., sorting, scattered writes), introducing overheads that negate theoretical speedups when paired with optimized attention implementations (e.g., FlashAttention). To bridge this gap, we propose Token Merge with Attention (ToMA), an off-the-shelf method that redesigns token reduction for GPU-aligned efficiency, with three key contributions: 1) a reformulation of token merging as a submodular optimization problem to select diverse tokens; 2) merge/unmerge as an attention-like linear transformation via GPU-friendly matrix operations; and 3) exploiting latent locality and sequential redundancy (pattern reuse) to minimize overhead. ToMA reduces SDXL/Flux generation latency by 24%/23% (DINO $\Delta <$ 0.07), outperforming prior methods. This work bridges the gap between theoretical and practical efficiency for transformers in diffusion.

Contrastive learning (CL) has achieved remarkable success in addressing data sparsity issues in collaborative filtering (CF) for recommender systems (RSs). The key principle is to generate different augmented views given a user-item interaction graph. However, prior endeavors mainly focus on performing augmentation via stochastic functions, e.g., by injecting perturbations into different hidden dimensions uniformly. Without fine control, the hidden representations of augmentations may contain noisy dimensions that are harmful to CL and irrelevant to RSs. Removing dimension-specific noise is a challenging task due to the following two major bottlenecks. It is difficult to (i) distinguish different dimensions' efficacy for CL and (ii) bridge the semantic gap between CL and RSs. Overlooking these limitations may cause redundant, false-positive, and irrelevant noise in hidden dimensions of the augmented views.

With increasing concern over data privacy, split learning has become a widely used distributed machine learning paradigm in practice, where two participants (namely the non-label party and the label party) own raw features and raw labels respectively, and jointly train a model. Although no raw data is communicated between the two parties during model training, several works have demonstrated that data privacy, especially label privacy, is still vulnerable in split learning, and have proposed several defense algorithms against label attacks. However, the theoretical guarantee on the privacy preservation of these algorithms is limited. In this work, we propose a novel Private Split Learning Framework (PSLF). In PSLF, the label party shares only the gradients computed by flipped labels with the non-label party, which improves privacy preservation on raw labels, and meanwhile, we further design an extra sub-model from true labels to improve prediction accuracy. We also design a Flipped Multi-Label Generation mechanism (FMLG) based on randomized response for the label party to generate flipped labels. FMLG is proven differentially private and the label party could make a trade-off between privacy and utility by setting the DP budget. In addition, we design an upsampling method to further protect the labels against some existing attacks. We have evaluated PSLF over real-world datasets to demonstrate its effectiveness in protecting label privacy and achieving promising prediction accuracy.

Machine learning force fields (MLFF) have been proposed to accelerate molecular dynamics (MD) simulation, which finds widespread applications in chemistry and biomedical research. Even for the most data-efficient MLFFs, reaching chemical accuracy can require hundreds of frames of force and energy labels generated by expensive quantum mechanical algorithms, which may scale as $O(n^3)$ to $O(n^7)$, with $n$ proportional to the number of basis functions. To address this issue, we propose a multi-stage computational framework -- ASTEROID, which lowers the data cost of MLFFs by leveraging a combination of cheap inaccurate data and expensive accurate data. The motivation behind ASTEROID is that inaccurate data, though incurring large bias, can help capture the sophisticated structures of the underlying force field. Therefore, we first train a MLFF model on a large amount of inaccurate training data, employing a bias-aware loss function to prevent the model from overfitting the potential bias of this data. We then fine-tune the obtained model using a small amount of accurate training data, which preserves the knowledge learned from the inaccurate training data while significantly improving the model's accuracy. Moreover, we propose a variant of ASTEROID based on score matching for the setting where the inaccurate training data are unlabeled. Extensive experiments on MD datasets and downstream tasks validate the efficacy of ASTEROID. Our code and data are available at https://github.com/abukharin3/asteroid.

End-to-end retrieval models, such as Tree-based Models (TDM) and Deep Retrieval (DR), have attracted a lot of attention, but they cannot handle cold-start and long-tail item recommendation scenarios well. Specifically, DR learns a compact indexing structure, enabling efficient and accurate retrieval for large recommendation systems. However, it is discovered that DR largely fails on retrieving cold-start and long-tail items. This is because DR only utilizes user-item interaction data, which is rare and often noisy for cold-start and long-tail items. Besides, end-to-end retrieval models are unable to make use of the rich item content features. To address this issue while maintaining the efficiency of DR indexing structure, we propose Full Index Deep Retrieval (FIDR) that learns indices for the full corpus items, including cold-start and long-tail items. In addition to the original structure in DR (called User Structure in FIDR) that learns with user-item interaction data (e.g., clicks), we add an Item Structure to embed items directly based on item content features (e.g., categories). With joint efforts of User Structure and Item Structure, FIDR makes cold-start items retrievable and also improves the recommendation quality of long-tail items. To our best knowledge, FIDR is the first to solve the cold-start and long-tail recommendation problem for the end-to-end retrieval models. Through extensive experiments on three real-world datasets, we demonstrate that FIDR can effectively recommend cold-start as well as long-tail items, and largely promote overall recommendation performance without sacrificing inference efficiency. According to the experiments, the recall of FIDR is improved by 8.8%~11.9%, while the inference of FIDR is as efficient as DR.

Graph Neural Network (GNN) based recommender systems have been attracting more and more attention in recent years due to their excellent performance in accuracy. Representing user-item interactions as a bipartite graph, a GNN model generates user and item representations by aggregating embeddings of their neighbors. However, such an aggregation procedure often accumulates information purely based on the graph structure, overlooking the redundancy of the aggregated neighbors and resulting in poor diversity of the recommended list. In this paper, we propose diversifying GNN-based recommender systems by directly improving the embedding generation procedure. Particularly, we utilize the following three modules: submodular neighbor selection to find a subset of diverse neighbors to aggregate for each GNN node, layer attention to assign attention weights for each layer, and loss reweighting to focus on the learning of items belonging to long-tail categories. Blending the three modules into GNN, we present DGRec (Diversified GNN-based Recommender System) for diversified recommendation. Experiments on real-world datasets demonstrate that the proposed method can achieve the best diversity while keeping the accuracy comparable to state-of-the-art GNN-based recommender systems. We open source DGRec at https://github.com/YangLiangwei/DGRec.

Continual learning is an emerging research challenge in machine learning that addresses the problem where models quickly fit the most recently trained-on data but suffer from catastrophic forgetting of previous data due to distribution shifts --- it does this by maintaining a small historical replay buffer in replay-based methods. To avoid these problems, this paper proposes a method, ``Retrospective Adversarial Replay (RAR)'', that synthesizes adversarial samples near the forgetting boundary. RAR perturbs a buffered sample towards its nearest neighbor drawn from the current task in a latent representation space. By replaying such samples, we are able to refine the boundary between previous and current tasks, hence combating forgetting and reducing bias towards the current task. To mitigate the severity of a small replay buffer, we develop a novel MixUp-based strategy to increase replay variation by replaying mixed augmentations. Combined with RAR, this achieves a holistic framework that helps to alleviate catastrophic forgetting. We show that this excels on broadly-used benchmarks and outperforms other continual learning baselines especially when only a small buffer is available. We conduct a thorough ablation study over each key component as well as a hyperparameter sensitivity analysis to demonstrate the effectiveness and robustness of RAR.

We study a novel curriculum learning scheme where in each round, samples are selected to achieve the greatest progress and fastest learning speed towards the ground-truth on all available samples. Inspired by an analysis of optimization dynamics under gradient flow for both regression and classification, the problem reduces to selecting training samples by a score computed from samples’ residual and linear temporal dynamics. It encourages the model to focus on the samples at learning frontier, i.e., those with large loss but fast learning speed. The scores in discrete time can be estimated via already-available byproducts of training, and thus require a negligible amount of extra computation. We discuss the properties and potential advantages of the proposed dynamics optimization via current deep learning theory and empirical study. By integrating it with cyclical training of neural networks, we introduce "dynamics-optimized curriculum learning (DoCL)", which selects the training set for each step by weighted sampling based on the scores. On nine different datasets, DoCL significantly outperforms random mini-batch SGD and recent curriculum learning methods both in terms of efficiency and final performance.

In the robust submodular partitioning problem, we aim to allocate a set of items into $m$ blocks, so that the evaluation of the minimum block according to a submodular function is maximized. Robust submodular partitioning promotes the diversity of every block in the partition. It has many applications in machine learning, e.g., partitioning data for distributed training so that the gradients computed on every block are consistent. We study an extension of the robust submodular partition problem with additional constraints (e.g., cardinality, multiple matroids, and/or knapsack) on every block. For example, when partitioning data for distributed training, we can add a constraint that the number of samples of each class is the same in each partition block, ensuring data balance. We present two classes of algorithms, i.e., Min-Block Greedy based algorithms (with an $\Omega(1/m)$ bound), and Round-Robin Greedy based algorithms (with a constant bound) and show that under various constraints, they still have good approximation guarantees. Interestingly, while normally the latter runs in only weakly polynomial time, we show that using the two together yields strongly polynomial running time while preserving the approximation guarantee. Lastly, we apply the algorithms on a real-world machine learning data partitioning problem showing good results.

Semi-supervised learning (SSL) leverages unlabeled data when training a model with insufficient labeled data. A common strategy for SSL is to enforce the consistency of model outputs between similar samples, e.g., neighbors or data augmentations of the same sample. However, model outputs can vary dramatically on unlabeled data over different training stages, e.g., when using large learning rates. This can introduce harmful noises and inconsistent objectives over time that may lead to concept drift and catastrophic forgetting. In this paper, we study the dynamics of neural net outputs in SSL and show that selecting and using first the unlabeled samples with more consistent outputs over the course of training (i.e., "time-consistency") can improve the final test accuracy and save computation. Under the time-consistent data selection, we design an SSL objective composed of two self-supervised losses, i.e., a consistency loss between a sample and its augmentation, and a contrastive loss encouraging different samples to have different outputs. Our approach achieves SOTA on several SSL benchmarks with much fewer computations.

A good teacher can adjust the curriculum based on students' learning history. By analogy, in this paper, we study the dynamics of a deep neural network's (DNN) performance on individual samples during its learning process. The observed properties allow us to develop an adaptive curriculum that leads to faster learning of more accurate models. We introduce dynamic instance hardness (DIH), the exponential moving average of a sample's instantaneous hardness (e.g., a loss, or a change in outputs) over the training history. A low DIH indicates that a model retains knowledge about a sample over time, and implies a flat loss landscape for that sample. Moreover, for DNNs, we find that a sample's DIH early in training predicts its DIH in later stages. Hence, we can train a model using samples with higher DIH and safely ignore those with lower DIH. This motivates a DIH guided curriculum learning (DIHCL). Compared to existing CL methods: (1) DIH is more stable over time than using only instantaneous hardness, which is noisy due to stochastic training and DNN's non-smoothness; (2) DIHCL is computationally inexpensive since it uses only a byproduct of back-propagation and thus does not require extra inference. On 11 datasets, DIHCL significantly outperforms random mini-batch SGD and recent CL methods in terms of efficiency and final performance.

We propose a general and efficient hierarchical robust partitioning framework to generate a deterministic sequence of mini-batches, one that offers assurances of being high quality, unlike a randomly drawn sequence. We compare our deterministically generated mini-batch sequences to randomly generated sequences; we show that, on a variety of deep learning tasks, the deterministic sequences significantly beat the mean and worst case performance of the random sequences, and often outperforms the best of the random sequences. Our theoretical contributions include a new algorithm for the robust submodular partition problem subject to cardinality constraints (which is used to construct mini-batch sequences), and show in general that the algorithm is fast and has good theoretical guarantees; we also show a more efficient hierarchical variant of the algorithm with similar guarantees under mild assumptions.

The gradient of a deep neural network (DNN) w.r.t. the input provides information that can be used to explain the output prediction in terms of the input features and has been widely studied to assist in interpreting DNNs. In a linear model (i.e., g(x) = wx + b), the gradient corresponds to the weights w. Such a model can reasonably locally-linearly approximate a smooth nonlinear DNN, and hence the weights of this local model are the gradient. The bias b, however, is usually overlooked in attribution methods. In this paper, we observe that since the bias in a DNN also has a non-negligible contribution to the correctness of predictions, it can also play a significant role in understanding DNN behavior. We propose a backpropagation-type algorithm “bias back-propagation (BBp)” that starts at the output layer and iteratively attributes the bias of each layer to its input nodes as well as combining the resulting bias term of the previous layer. Together with the backpropagation of the gradient generating w, we can fully recover the locally linear model g(x) = wx + b. In experiments, we show that BBp can generate complementary and highly interpretable explanations.

We discuss three novel insights about dropout for DNNs with ReLUs: 1) dropout encourages each local linear piece of a DNN to be trained on data points from nearby regions; 2) the same dropout rate results in different (effective) deactivation rates for layers with different portions of ReLU-deactivated neurons; and 3) the rescaling factor of dropout causes a normalization inconsistency between training and test when used together with batch normalization. The above leads to three simple but nontrivial modifications resulting in our method “jumpout.” Jumpout samples the dropout rate from a monotone decreasing distribution (e.g., the right half of a Gaussian), so each local linear piece is trained, with high probability, to work better for data points from nearby than more distant regions. Jumpout moreover adaptively normalizes the dropout rate at each layer and every training batch, so the effective deactivation rate on the activated neurons is kept the same. Furthermore, it rescales the outputs for a better trade-off that keeps both the variance and mean of neurons more consistent between training and test phases, thereby mitigating the incompatibility between dropout and batch normalization. Jumpout significantly improves the performance of different neural nets on CIFAR10, CIFAR100, Fashion-MNIST, STL10, SVHN, ImageNet-1k, etc., while introducing negligible additional memory and computation costs.

We study a new method (``Diverse Ensemble Evolution (DivE$^2$)'') to train an ensemble of machine learning models that assigns data to models at each training epoch based on each model's current expertise and an intra- and inter-model diversity reward. DivE$^2$ schedules, over the course of training epochs, the relative importance of these characteristics; it starts by selecting easy samples for each model, and then gradually adjusts towards the models having specialized and complementary expertise on subsets of the training data, thereby encouraging high accuracy of the ensemble. We utilize an intra-model diversity term on data assigned to each model, and an inter-model diversity term on data assigned to pairs of models, to penalize both within-model and cross-model redundancy. We formulate the data-model marriage problem as a generalized bipartite matching, represented as submodular maximization subject to two matroid constraints. DivE$^2$ solves a sequence of continuous-combinatorial optimizations with slowly varying objectives and constraints. The combinatorial part handles the data-model marriage while the continuous part updates model parameters based on the assignments. In experiments, DivE$^2$ outperforms other ensemble training methods under a variety of model aggregation techniques, while also maintaining competitive efficiency.

Deep neural networks have achieved great successes on various machine learning tasks, however, there are many open fundamental questions to be answered. In this paper, we tackle the problem of quantifying the quality of learned wights of different networks with possibly different architectures, going beyond considering the final classification error as the only metric. We introduce \emphExtended Data Jacobian Matrix to help analyze properties of networks of various structures, finding that, the spectrum of the extended data jacobian matrix is a strong discriminating factor for networks of different structures and performance. Based on such observation, we propose a novel regularization method, which manages to improve the network performance comparably to dropout, which in turn verifies the observation.

UNLABELLED: Tandem mass spectrometry (MS/MS) is the dominant high throughput technology for identifying and quantifying proteins in complex biological samples. Analysis of the tens of thousands of fragmentation spectra produced by an MS/MS experiment begins by assigning to each observed spectrum the peptide that is hypothesized to be responsible for generating the spectrum. This assignment is typically done by searching each spectrum against a database of peptides. To our knowledge, all existing MS/MS search engines compute scores individually between a given observed spectrum and each possible candidate peptide from the database. In this work, we use a trellis, a data structure capable of jointly representing a large set of candidate peptides, to avoid redundantly recomputing common sub-computations among different candidates. We show how trellises may be used to significantly speed up existing scoring algorithms, and we theoretically quantify the expected speedup afforded by trellises. Furthermore, we demonstrate that compact trellis representations of whole sets of peptides enables efficient discriminative learning of a dynamic Bayesian network for spectrum identification, leading to greatly improved spectrum identification accuracy. CONTACT: [email protected] or [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

We investigate two novel mixed robust/average-case submodular data partitioning problems that we collectively call Submodular Partitioning. These problems generalize purely robust instances of the problem, namely max-min submodular fair allocation (SFA) and \emph{min-max submodular load balancing} (SLB), and also average-case instances, that is the submodular welfare problem (SWP) and submodular multiway partition (SMP). While the robust versions have been studied in the theory community, existing work has focused on tight approximation guarantees, and the resultant algorithms are not generally scalable to large real-world applications. This contrasts the average case instances, where most of the algorithms are scalable. In the present paper, we bridge this gap, by proposing several new algorithms (including greedy, majorization-minimization, minorization-maximization, and relaxation algorithms) that not only scale to large datasets but that also achieve theoretical approximation guarantees comparable to the state-of-the-art. We moreover provide new scalable algorithms that apply to additive combinations of the robust and average-case objectives. We show that these problems have many applications in machine learning (ML), including data partitioning and load balancing for distributed ML, data clustering, and image segmentation. We empirically demonstrate the efficacy of our algorithms on real-world problems involving data partitioning for distributed optimization (of convex and deep neural network objectives), and also purely unsupervised image segmentation.