Large Language Models (LLMs) are increasingly used in applications requiring long context lengths, but the key-value (KV) cache often becomes a memory bottleneck on GPUs as context grows. To address this, we propose Commutative Vector Quantization (CommVQ) to significantly reduce memory usage for long-context LLM inference. We first introduce additive quantization with a lightweight encoder and codebook to compress the KV cache, which can be decoded via simple matrix multiplication. To further reduce computational costs during decoding, we design the codebook to be commutative with Rotary Position Embedding (RoPE) and train it using an Expectation-Maximization (EM) algorithm. This enables efficient integration of decoding into the self-attention mechanism. Our approach achieves high accuracy with additive quantization and low overhead via the RoPE-commutative codebook. Experiments on long-context benchmarks and GSM8K show that our method reduces FP16 KV cache size by 87.5% with 2-bit quantization, while outperforming state-of-the-art KV cache quantization methods. Notably, it enables 1-bit KV cache quantization with minimal accuracy loss, allowing a LLaMA-3.1 8B model to run with a 128K context length on a single RTX 4090 GPU. The source code is available at: https://github.com/UMass-Embodied-AGI/CommVQ.

Large, pretrained models are commonly finetuned with imagery that is heavily augmented to mimic different conditions and scales, with the resulting models used for various tasks with imagery from a range of spatial scales. Such models overlook scale-specific information in the data for scale-dependent domains, such as remote sensing. In this paper, we present Scale-MAE, a pretraining method that explicitly learns relationships between data at different, known scales throughout the pretraining process. Scale-MAE pre-trains a network by masking an input image at a known input scale, where the area of the Earth covered by the image determines the scale of the ViT positional encoding, not the image resolution. Scale-MAE encodes the masked image with a standard ViT backbone, and then decodes the masked image through a bandpass filter to reconstruct low/high frequency images at lower/higher scales. We find that tasking the network with reconstructing both low/high frequency images leads to robust multiscale representations for remote sensing imagery. Scale-MAE achieves an average of a 2.4 − 5.6% non-parametric kNN classification improvement across eight remote sensing datasets compared to current state-of-the-art and obtains a 0.9 mIoU to 1.7 mIoU improvement on the SpaceNet building segmentation transfer task for a range of evaluation scales.

Invariant risk minimization (IRM) has recently emerged as a promising alternative for domain generalization. Nevertheless, the loss function is difficult to optimize for nonlinear classifiers and the original optimization objective could fail when pseudo-invariant features and geometric skews exist. Inspired by IRM, in this paper we propose a novel formulation for domain generalization, dubbed invariant information bottleneck (IIB). IIB aims at minimizing invariant risks for nonlinear classifiers and simultaneously mitigating the impact of pseudo-invariant features and geometric skews. Specifically, we first present a novel formulation for invariant causal prediction via mutual information. Then we adopt the variational formulation of the mutual information to develop a tractable loss function for nonlinear classifiers. To overcome the failure modes of IRM, we propose to minimize the mutual information between the inputs and the corresponding representations. IIB significantly outperforms IRM on synthetic datasets, where the pseudo-invariant features and geometric skews occur, showing the effectiveness of proposed formulation in overcoming failure modes of IRM. Furthermore, experiments on DomainBed show that IIB outperforms 13 baselines by 0.9% on average across 7 real datasets.

Recent self-supervised pretraining methods for object detection largely focus on pretraining the backbone of the object detector, neglecting key parts of detection architecture. Instead, we introduce DETReg, a new self-supervised method that pretrains the entire object detection network, including the object localization and embedding components. During pretraining, DETReg predicts object localizations to match the localizations from an unsupervised region proposal generator and simultaneously aligns the corresponding feature embeddings with embeddings from a self-supervised image encoder. We implement DETReg using the DETR family of detectors and show that it improves over competitive baselines when finetuned on COCO, PASCAL VOC, and Airbus Ship benchmarks. In low-data regimes, including semi-supervised and few-shot learning settings, DETReg establishes many state-of-the-art results, e.g., on COCO we see a +6.0 AP improvement for 10-shot detection and over 2 AP improvements when training with only 1 % of the labels.11Code: https://www.amirbar.net/detreg/.

While self-supervised pretraining has proven beneficial for many computer vision tasks, it requires expensive and lengthy computation, large amounts of data, and is sensitive to data augmentation. Prior work demonstrates that models pretrained on datasets dissimilar to their target data, such as chest X-ray models trained on ImageNet, underperform models trained from scratch. Users that lack the resources to pretrain must use existing models with lower performance. This paper explores Hierarchical PreTraining (HPT), which decreases convergence time and improves accuracy by initializing the pretraining process with an existing pretrained model. Through experimentation on 16 diverse vision datasets, we show HPT converges up to 80× faster, improves accuracy across tasks, and improves the robustness of the self-supervised pretraining process to changes in the image augmentation policy or amount of pretraining data. Taken together, HPT provides a simple framework for obtaining better pretrained representations with less computational resources.

A common practice in unsupervised representation learning is to use labeled data to evaluate the quality of the learned representations. This supervised evaluation is then used to guide critical aspects of the training process such as selecting the data augmentation policy. However, guiding an unsupervised training process through supervised evaluations is not possible for real-world data that does not actually contain labels (which may be the case, for example, in privacy sensitive fields such as medical imaging). Therefore, in this work we show that evaluating the learned representations with a self-supervised image rotation task is highly correlated with a standard set of supervised evaluations (rank correlation > 0.94). We establish this correlation across hundreds of augmentation policies, training settings, and network architectures and provide an algorithm (SelfAugment) to automatically and efficiently select augmentation policies without using supervised evaluations. Despite not using any labeled data, the learned augmentation policies perform comparably with augmentation policies that were determined using exhaustive supervised evaluations.

We present Region Similarity Representation Learning (ReSim), a new approach to self-supervised representation learning for localization-based tasks such as object detection and segmentation. While existing work has largely focused on solely learning global representations for an entire image, ReSim learns both regional representations for localization as well as semantic image-level representations. ReSim operates by sliding a fixed-sized window across the overlapping area between two views (e.g., image crops), aligning these areas with their corresponding convolutional feature map regions, and then maximizing the feature similarity across views. As a result, ReSim learns spatially and semantically consistent feature representation throughout the convolutional feature maps of a neural network. A shift or scale of an image region, e.g., a shift or scale of an object, has a corresponding change in the feature maps; this allows downstream tasks to leverage these representations for localization. Through object detection, instance segmentation, and dense pose estimation experiments, we illustrate how ReSim learns representations which significantly improve the localization and classification performance compared to a competitive MoCo-v2 baseline: $ + 2.7AP_{75}^{bb}VOC$, $ + 1.1AP_{75}^{bb}COCO$, and +1.9 APmkCityscapes. Code and pre-trained models are released at: https://github.com/Tete-Xiao/ReSim

Increasingly, authors are publishing long informational talks, lectures, and distance-learning videos online. However, it is difficult to browse and skim the content of such videos using current timeline-based video players. Video digests are a new format for informational videos that afford browsing and skimming by segmenting videos into a chapter/section structure and providing short text summaries and thumbnails for each section. Viewers can navigate by reading the summaries and clicking on sections to access the corresponding point in the video. We present a set of tools to help authors create such digests using transcript-based interactions. With our tools, authors can manually create a video digest from scratch, or they can automatically generate a digest by applying a combination of algorithmic and crowdsourcing techniques and then manually refine it as needed. Feedback from first-time users suggests that our transcript-based authoring tools and automated techniques greatly facilitate video digest creation. In an evaluative crowdsourced study we find that given a short viewing time, video digests support browsing and skimming better than timeline-based or transcript-based video players.

Inference for latent feature models is inherently difficult as the inference space grows exponentially with the size of the input data and number of latent features. In this work, we use Kurihara & Wellings (2008)’s maximization-expectation framework to perform approximate MAP inference for linear-Gaussian latent feature models with an Indian Buffet Process (IBP) prior. This formulation yields a submodular function of the features that corresponds to a lower bound on the model evidence. By adding a constant to this function, we obtain a nonnegative submodular function that can be maximized via a greedy algorithm that obtains at least a 1/3-approximation to the optimal solution. Our inference method scales linearly with the size of the input data, and we show the efficacy of our method on the largest datasets currently analyzed using an IBP model.