Prevailing Multimodal Large Language Models (MLLMs) encode the input image(s) as vision tokens and feed them into the language backbone, similar to how Large Language Models (LLMs) process the text tokens. However, the number of vision tokens increases quadratically as the image resolutions, leading to huge computational costs. In this paper, we consider improving MLLM’s efficiency from two scenarios, (I) Reducing computational cost without degrading the performance. (II) Improving the performance with given budgets. We start with our main finding that the ranking of each vision token sorted by attention scores is similar in each layer except the first layer. Based on it, we assume that the number of essential top vision tokens does not increase along layers. Accordingly, for Scenario I, we propose a greedy search algorithm (G-Search) to find the least number of vision tokens to keep at each layer from the shallow to the deep. Interestingly, G-Search is able to reach the optimal reduction strategy based on our assumption. For Scenario II, based on the reduction strategy from G-Search, we design a parametric sigmoid function (P-Sigmoid) to guide the reduction at each layer of the MLLM, whose parameters are optimized by Bayesian Optimization. Extensive experiments demonstrate that our approach can significantly accelerate those popular MLLMs, e.g. LLaVA, and InternVL2 models, by more than 2⇥ without performance drops. Our approach also far outperforms other token reduction methods when budgets are limited, achieving a better trade-off between efficiency and effectiveness.

Despite the rapid integration of video perception capabilities into Large Multimodal Models (LMMs), what drives their video perception remains poorly understood. Consequently, many design decisions in this domain are made without proper justification or analysis. The high computational cost of training and evaluating such models and limited open research hinder the development of video-LMMs. To address this, we present a comprehensive study that helps uncover what effectively drives video understanding in LMMs. We begin by critically examining the primary contributors to the high computational requirements associated with video-LMM research and discover Scaling Consistency, wherein design and training decisions made on smaller models and datasets (up to a critical size) effectively transfer to larger models. Leveraging these insights, we explored many video-specific aspects of video-LMMs, including video sampling, architectures, data composition, training schedules, and more. Guided by these findings, we introduce Apollo, a state-of-the-art family of LMMs that achieve superior performance across different model sizes. Our models process over 1-hour videos efficiently, with the 3B parameter variant outperforming most existing 7B models. Apollo-7B is state-of-the-art compared to 7B LMMs with a 70.9 on MLVU, and 63.3 on Video-MME.

Text-to-Video generation has achieved remarkable progress with the rise of diffusion models. In this work, we introduce Cached Memory-Guided Video Generation (Corgi), aiming to generate multi-scene videos with arbi-trary number of video clips, conditioned on input images and instruction prompts. This is a challenging task, as tra-ditional T2V methods often struggle to maintain the quality of longer videos due to the difficulties in preserving visual context from earlier scenes. We address this by introducing a cached memory mechanism that stores the key frames. Our multi-scene video generation process is explicitly con-ditioned on the cached memories to avoid forgetting the vi-sual appearance of target subjects. Corgi shows significant improvement in multi-scene video generation compared to the prior art, with up to 59.2% in long-term consistency and 7.6% in diversity.

Large Vision-Language Foundation Models (VLFM), such as CLIP, ALIGN and Florence, are trained on large-scale datasets of image-caption pairs and achieve superior transferability and robustness on downstream tasks, but they are difficult to use in many practical applications due to their large size, high latency and fixed architectures. Unfortunately, recent work shows training a small custom VLFM for resource-limited applications is currently very difficult using public and smaller-scale data. In this paper, we introduce a new distillation mechanism (DIME-FM) that allows us to transfer the knowledge contained in large VLFMs to smaller, customized foundation models using a relatively small amount of inexpensive, unpaired images and sentences. We transfer the knowledge from the pre-trained CLIP-ViT-L/14 model to a ViT-B/32 model, with only 40M public images and 28.4M unpaired public sentences. The resulting model "Distill-ViT-B/32" rivals the CLIP-ViT-B/32 model pre-trained on its private WiT dataset (400M image-text pairs): Distill-ViT-B/32 achieves similar results in terms of zero-shot and linear-probing performance on both Ima-geNet and the ELEVATER (20 image classification tasks) benchmarks. It also displays comparable robustness when evaluated on five datasets with natural distribution shifts from ImageNet. Please refer to our project page for code and more details.

We present a novel algorithm for transferring artistic styles of semantically meaningful local regions of an image onto local regions of a target video while preserving its photorealism. Local regions may be selected either fully automatically from an image, through using video segmentation algorithms, or from casual user guidance such as scribbles. Our method, based on a deep neural network architecture inspired by recent work in photorealistic style transfer, is real-time and works on arbitrary inputs without runtime optimization once trained on a diverse dataset of artistic styles. By augmenting our video dataset with noisy semantic labels and jointly optimizing over style, content, mask, and temporal losses, our method can cope with a variety of imperfections in the input and produce temporally coherent videos without visual artifacts. We demonstrate our method on a variety of style images and target videos, including the ability to transfer different styles onto multiple objects simultaneously, and smoothly transition between styles in time.

Bregman divergences generalize measures such as the squared Euclidean distance and the KL divergence, and arise throughout many areas of machine learning. In this paper, we focus on the problem of approximating an arbitrary Bregman divergence from supervision, and we provide a well-principled approach to analyzing such approximations. We develop a formulation and algorithm for learning arbitrary Bregman divergences based on approximating their underlying convex generating function via a piecewise linear function. We provide theoretical approximation bounds using our parameterization and show that the generalization error $O_p(m^{-1/2})$ for metric learning using our framework matches the known generalization error in the strictly less general Mahalanobis metric learning setting. We further demonstrate empirically that our method performs well in comparison to existing metric learning methods, particularly for clustering and ranking problems.

We propose a novel deep metric learning method by revisiting the learning to rank approach. Our method, named FastAP, optimizes the rank-based Average Precision measure, using an approximation derived from distance quantization. FastAP has a low complexity compared to existing methods, and is tailored for stochastic gradient descent. To fully exploit the benefits of the ranking formulation, we also propose a new minibatch sampling scheme, as well as a simple heuristic to enable large-batch training. On three few-shot image retrieval datasets, FastAP consistently outperforms competing methods, which often involve complex optimization heuristics or costly model ensembles.

Conventional unsupervised domain adaptation (UDA) assumes that training data are sampled from a single domain. This neglects the more practical scenario where training data are collected from multiple sources, requiring multi-source domain adaptation. We make three major contributions towards addressing this problem. First, we collect and annotate by far the largest UDA dataset, called DomainNet, which contains six domains and about 0.6 million images distributed among 345 categories, addressing the gap in data availability for multi-source UDA research. Second, we propose a new deep learning approach, Moment Matching for Multi-Source Domain Adaptation (M3SDA), which aims to transfer knowledge learned from multiple labeled source domains to an unlabeled target domain by dynamically aligning moments of their feature distributions. Third, we provide new theoretical insights specifically for moment matching approaches in both single and multiple source domain adaptation. Extensive experiments are conducted to demonstrate the power of our new dataset in benchmarking state-of-the-art multi-source domain adaptation methods, as well as the advantage of our proposed model. Dataset and Code are available at http://ai.bu.edu/M3SDA/.

Astronomers and telescope operators must make decisions about what to observe given limited telescope time. To optimize this decision-making process, we present a batch, cost-sensitive, active learning approach that exploits structure in the unlabeled dataset, accounts for label uncertainty, and minimizes annotation costs. We first cluster the unlabeled instances in feature space. We next introduce an uncertainty-reducing selection criterion that encourages the batch of selected instances to span multiple clusters, in addition to taking into account annotation cost. Finally, we extend this criterion to incorporate the fact that nearby astronomical objects may be observed at the same time. On two large astronomical data sets, our approach balances the trade-offs among FOV, aperture, and time cost and, therefore, helps astronomers design effective experiments.

BACKGROUND: Molecular interactions between proteins and DNA molecules underlie many cellular processes, including transcriptional regulation, chromosome replication, and nucleosome positioning. Computational analyses of protein-DNA interactions rely on experimental data characterizing known protein-DNA interactions structurally and biochemically. While many databases exist that contain either structural or biochemical data, few integrate these two data sources in a unified fashion. Such integration is becoming increasingly critical with the rapid growth of structural and biochemical data, and the emergence of algorithms that rely on the synthesis of multiple data types to derive computational models of molecular interactions. DESCRIPTION: We have developed an integrated affinity-structure database in which the experimental and quantitative DNA binding affinities of helix-turn-helix proteins are mapped onto the crystal structures of the corresponding protein-DNA complexes. This database provides access to: (i) protein-DNA structures, (ii) quantitative summaries of protein-DNA binding affinities using position weight matrices, and (iii) raw experimental data of protein-DNA binding instances. Critically, this database establishes a correspondence between experimental structural data and quantitative binding affinity data at the single basepair level. Furthermore, we present a novel alignment algorithm that structurally aligns the protein-DNA complexes in the database and creates a unified residue-level coordinate system for comparing the physico-chemical environments at the interface between complexes. Using this unified coordinate system, we compute the statistics of atomic interactions at the protein-DNA interface of helix-turn-helix proteins. We provide an interactive website for visualization, querying, and analyzing this database, and a downloadable version to facilitate programmatic analysis. CONCLUSIONS: This database will facilitate the analysis of protein-DNA interactions and the development of programmatic computational methods that capitalize on integration of structural and biochemical datasets. The database can be accessed at http://ProteinDNA.hms.harvard.edu.