HashEvict: A Pre-Attention KV Cache Eviction Strategy using Locality-Sensitive Hashing

Transformer-based large language models (LLMs) use the key-value (KV) cache to significantly accelerate inference by storing the key and value embeddings of past tokens. However, this cache consumes significant GPU memory. In this work, we introduce HashEvict, an algorithm that uses locality-sensitive hashing (LSH) to compress the KV cache. HashEvict quickly locates tokens in the cache that are cosine dissimilar to the current query token. This is achieved by computing the Hamming distance between binarized Gaussian projections of the current token query and cached token keys, with a projection length much smaller than the embedding dimension. We maintain a lightweight binary structure in GPU memory to facilitate these calculations. Unlike existing compression strategies that compute attention to determine token retention, HashEvict makes these decisions pre-attention, thereby reducing computational costs. Additionally, HashEvict is dynamic - at every decoding step, the key and value of the current token replace the embeddings of a token expected to produce the lowest attention score. We demonstrate that HashEvict can compress the KV cache by 30%-70% while maintaining high performance across reasoning, multiple-choice, long-context retrieval and summarization tasks.

Learning Normal Flow Directly From Event Neighborhoods

Event-based motion field estimation is an important task. However, current optical flow methods face challenges: learning-based approaches, often frame-based and relying on CNNs, lack cross-domain transferability, while model-based methods, though more robust, are less accurate. To address the limitations of optical flow estimation, recent works have focused on normal flow, which can be more reliably measured in regions with limited texture or strong edges. However, existing normal flow estimators are predominantly model-based and suffer from high errors. In this paper, we propose a novel supervised point-based method for normal flow estimation that overcomes the limitations of existing event learning-based approaches. Using a local point cloud encoder, our method directly estimates per-event normal flow from raw events, offering multiple unique advantages: 1) It produces temporally and spatially sharp predictions. 2) It supports more diverse data augmentation, such as random rotation, to improve robustness across various domains. 3) It naturally supports uncertainty quantification via ensemble inference, which benefits downstream tasks. 4) It enables training and inference on undistorted data in normalized camera coordinates, improving transferability across cameras. Extensive experiments demonstrate our method achieves better and more consistent performance than state-of-the-art methods when transferred across different datasets. Leveraging this transferability, we train our model on the union of datasets and release it for public use. Finally, we introduce an egomotion solver based on a maximum-margin problem that uses normal flow and IMU to achieve strong performance in challenging scenarios.

VioPose: Violin Performance 4D Pose Estimation by Hierarchical Audiovisual Inference

Musicians delicately control their bodies to generate music. Sometimes, their motions are too subtle to be captured by the human eye. To analyze how they move to produce the music, we need to estimate precise 4D human pose (3D pose over time). However, current state-of-the-art (SoTA) visual pose estimation algorithms struggle to produce accurate monocular 4D poses because of occlusions, partial views, and human-object interactions. They are limited by the viewing angle, pixel density, and sampling rate of the cameras and fail to estimate fast and subtle movements, such as in the musical effect of vibrato. We leverage the direct causal relationship between the music produced and the human motions creating them to address these challenges. We propose VioPose: a novel multimodal network that hierarchically estimates dynamics. High-level features are cascaded to low-level features and integrated into Bayesian updates. Our architecture is shown to produce accurate pose sequences, facilitating precise motion analysis, and outperforms SoTA. As part of this work, we collected the largest and the most diverse calibrated violin-playing dataset, including video, sound, and 3D motion capture poses. Project page: is available at https://sj-yoo.info/viopose/.

Extremum Seeking Controlled Wiggling for Tactile Insertion

When humans perform insertion tasks such as inserting a cup into a cupboard, routing a cable, or key insertion, they wiggle the object and observe the process through tactile and proprioceptive feedback. While recent advances in tactile sensors have resulted in tactile-based approaches, there has not been a generalized formulation based on wiggling similar to human behavior. Thus, we propose an extremum-seeking control law that can insert four keys into four types of locks without control parameter tuning despite significant variation in lock type. The resulting model-free formulation wiggles the end effector pose to maximize insertion depth while minimizing strain as measured by a GelSight Mini tactile sensor that grasps a key. The algorithm achieves a 71\% success rate over 120 randomly initialized trials with uncertainty in both translation and orientation. Over 240 deterministically initialized trials, where only one translation or rotation parameter is perturbed, 84\% of trials succeeded. Given tactile feedback at 13 Hz, the mean insertion time for these groups of trials are 262 and 147 seconds respectively.

FeelAnyForce: Estimating Contact Force Feedback from Tactile Sensation for Vision-Based Tactile Sensors

In this paper, we tackle the problem of estimating 3D contact forces using vision-based tactile sensors. In particular, our goal is to estimate contact forces over a large range (up to 15 N) on any objects while generalizing across different vision-based tactile sensors. Thus, we collected a dataset of over 200K indentations using a robotic arm that pressed various indenters onto a GelSight Mini sensor mounted on a force sensor and then used the data to train a multi-head transformer for force regression. Strong generalization is achieved via accurate data collection and multi-objective optimization that leverages depth contact images. Despite being trained only on primitive shapes and textures, the regressor achieves a mean absolute error of 4\% on a dataset of unseen real-world objects. We further evaluate our approach's generalization capability to other GelSight mini and DIGIT sensors, and propose a reproducible calibration procedure for adapting the pre-trained model to other vision-based sensors. Furthermore, the method was evaluated on real-world tasks, including weighing objects and controlling the deformation of delicate objects, which relies on accurate force feedback. Project webpage: http://prg.cs.umd.edu/FeelAnyForce

Maelstrom Networks

Artificial Neural Networks has struggled to devise a way to incorporate working memory into neural networks. While the ``long term'' memory can be seen as the learned weights, the working memory consists likely more of dynamical activity, that is missing from feed-forward models. Current state of the art models such as transformers tend to ``solve'' this by ignoring working memory entirely and simply process the sequence as an entire piece of data; however this means the network cannot process the sequence in an online fashion, and leads to an immense explosion in memory requirements. Here, inspired by a combination of controls, reservoir computing, deep learning, and recurrent neural networks, we offer an alternative paradigm that combines the strength of recurrent networks, with the pattern matching capability of feed-forward neural networks, which we call the \textit{Maelstrom Networks} paradigm. This paradigm leaves the recurrent component - the \textit{Maelstrom} - unlearned, and offloads the learning to a powerful feed-forward network. This allows the network to leverage the strength of feed-forward training without unrolling the network, and allows for the memory to be implemented in new neuromorphic hardware. It endows a neural network with a sequential memory that takes advantage of the inductive bias that data is organized causally in the temporal domain, and imbues the network with a state that represents the agent's ``self'', moving through the environment. This could also lead the way to continual learning, with the network modularized and ``'protected'' from overwrites that come with new data. In addition to aiding in solving these performance problems that plague current non-temporal deep networks, this also could finally lead towards endowing artificial networks with a sense of ``self''.

A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture

We propose VecKM, a novel local point cloud geometry encoder that is descriptive, efficient and robust to noise. VecKM leverages a unique approach by vectorizing a kernel mixture to represent the local point clouds. Such representation is descriptive and robust to noise, which is supported by two theorems that confirm its ability to reconstruct and preserve the similarity of the local shape. Moreover, VecKM is the first successful attempt to reduce the computation and memory costs from $O(n^2+nKd)$ to $O(nd)$ by sacrificing a marginal constant factor, where $n$ is the size of the point cloud and $K$ is neighborhood size. The efficiency is primarily due to VecKM's unique factorizable property that eliminates the need of explicitly grouping points into neighborhoods. In the normal estimation task, VecKM demonstrates not only 100x faster inference speed but also strongest descriptiveness and robustness compared with existing popular encoders. In classification and segmentation tasks, integrating VecKM as a preprocessing module achieves consistently better performance than the PointNet, PointNet++, and point transformer baselines, and runs consistently faster by up to 10x.

Choreographing the Digital Canvas: A Machine Learning Approach to Artistic Performance

This paper introduces the concept of a design tool for artistic performances based on attribute descriptions. To do so, we used a specific performance of falling actions. The platform integrates a novel machine-learning (ML) model with an interactive interface to generate and visualize artistic movements. Our approach's core is a cyclic Attribute-Conditioned Variational Autoencoder (AC-VAE) model developed to address the challenge of capturing and generating realistic 3D human body motions from motion capture (MoCap) data. We created a unique dataset focused on the dynamics of falling movements, characterized by a new ontology that divides motion into three distinct phases: Impact, Glitch, and Fall. The ML model's innovation lies in its ability to learn these phases separately. It is achieved by applying comprehensive data augmentation techniques and an initial pose loss function to generate natural and plausible motion. Our web-based interface provides an intuitive platform for artists to engage with this technology, offering fine-grained control over motion attributes and interactive visualization tools, including a 360-degree view and a dynamic timeline for playback manipulation. Our research paves the way for a future where technology amplifies the creative potential of human expression, making sophisticated motion generation accessible to a wider artistic community.

MARVIS: Motion & Geometry Aware Real and Virtual Image Segmentation

Tasks such as autonomous navigation, 3D reconstruction, and object recognition near the water surfaces are crucial in marine robotics applications. However, challenges arise due to dynamic disturbances, e.g., light reflections and refraction from the random air-water interface, irregular liquid flow, and similar factors, which can lead to potential failures in perception and navigation systems. Traditional computer vision algorithms struggle to differentiate between real and virtual image regions, significantly complicating tasks. A virtual image region is an apparent representation formed by the redirection of light rays, typically through reflection or refraction, creating the illusion of an object's presence without its actual physical location. This work proposes a novel approach for segmentation on real and virtual image regions, exploiting synthetic images combined with domain-invariant information, a Motion Entropy Kernel, and Epipolar Geometric Consistency. Our segmentation network does not need to be re-trained if the domain changes. We show this by deploying the same segmentation network in two different domains: simulation and the real world. By creating realistic synthetic images that mimic the complexities of the water surface, we provide fine-grained training data for our network (MARVIS) to discern between real and virtual images effectively. By motion & geometry-aware design choices and through comprehensive experimental analysis, we achieve state-of-the-art real-virtual image segmentation performance in unseen real world domain, achieving an IoU over 78% and a F1-Score over 86% while ensuring a small computational footprint. MARVIS offers over 43 FPS (8 FPS) inference rates on a single GPU (CPU core). Our code and dataset are available here https://github.com/jiayi-wu-umd/MARVIS.

LEAP: LLM-Generation of Egocentric Action Programs

We introduce LEAP (illustrated in Figure 1), a novel method for generating video-grounded action programs through use of a Large Language Model (LLM). These action programs represent the motoric, perceptual, and structural aspects of action, and consist of sub-actions, pre- and post-conditions, and control flows. LEAP's action programs are centered on egocentric video and employ recent developments in LLMs both as a source for program knowledge and as an aggregator and assessor of multimodal video information. We apply LEAP over a majority (87\%) of the training set of the EPIC Kitchens dataset, and release the resulting action programs as a publicly available dataset here (https://drive.google.com/drive/folders/1Cpkw_TI1IIxXdzor0pOXG3rWJWuKU5Ex?usp=drive_link). We employ LEAP as a secondary source of supervision, using its action programs in a loss term applied to action recognition and anticipation networks. We demonstrate sizable improvements in performance in both tasks due to training with the LEAP dataset. Our method achieves 1st place on the EPIC Kitchens Action Recognition leaderboard as of November 17 among the networks restricted to RGB-input (see Supplementary Materials).