Vision Transformers (ViTs) have achieved state-of-the-art performance in offline video action detection, but their reliance on processing fixed-size clips with full spatiotemporal attention makes them computationally expensive and ill-suited for real-time streaming applications due to massive computational redundancy. This paper introduces a novel framework to adapt these powerful offline models into efficient, online student models through knowledge distillation. We propose two causal, streaming-friendly attention architectures that replace the full self-attention mechanism: (1) a lightweight Temporal Shift Attention that integrates past context with minimal overhead, and (2) a Decomposed Spatial-Temporal Attention that combines intraframe spatial attention with an LSTM for temporal modeling. Both architectures utilize caching to eliminate redundant operations on a frame-by-frame basis. To maximize knowledge transfer, we introduce an uncertainty-guided distillation process, which formulates the training as a multi-task learning problem. Our resulting models demonstrate significant efficiency gains, achieving up to a 4x improvement in latency and throughput compared to the original offline teacher while ensuring state-of-the-art performance for online methods. Our work provides a practical and effective methodology for deploying high-accuracy transformer models in latency-sensitive, real-world video analysis systems.

Human actions often involve complex interactions across several inter-related objects in the scene. However, existing approaches to fine-grained video understanding or visual relationship detection often rely on single object representation or pairwise object relationships. Furthermore, learning interactions across multiple objects in hundreds of frames for video is computationally infeasible and performance may suffer since a large combinatorial space has to be modeled. In this paper, we propose to efficiently learn higher-order interactions between arbitrary subgroups of objects for fine-grained video understanding. We demonstrate that modeling object interactions significantly improves accuracy for both action recognition and video captioning, while saving more than 3-times the computation over traditional pairwise relationships. The proposed method is validated on two large-scale datasets: Kinetics and ActivityNet Captions. Our SINet and SINet-Caption achieve state-of-the-art performances on both datasets even though the videos are sampled at a maximum of 1 FPS. To the best of our knowledge, this is the first work modeling object interactions on open domain large-scale video datasets, and we additionally model higher-order object interactions which improves the performance with low computational costs.

Virtually every molecular biologist has searched a protein or DNA sequence database to find sequences that are evolutionarily related to a given query. Pairwise sequence comparison methods--i.e., measures of similarity between query and target sequences--provide the engine for sequence database search and have been the subject of 30 years of computational research. For the difficult problem of detecting remote evolutionary relationships between protein sequences, the most successful pairwise comparison methods involve building local models (e.g., profile hidden Markov models) of protein sequences. However, recent work in massive data domains like web search and natural language processing demonstrate the advantage of exploiting the global structure of the data space. Motivated by this work, we present a large-scale algorithm called ProtEmbed, which learns an embedding of protein sequences into a low-dimensional "semantic space." Evolutionarily related proteins are embedded in close proximity, and additional pieces of evidence, such as 3D structural similarity or class labels, can be incorporated into the learning process. We find that ProtEmbed achieves superior accuracy to widely used pairwise sequence methods like PSI-BLAST and HHSearch for remote homology detection; it also outperforms our previous RankProp algorithm, which incorporates global structure in the form of a protein similarity network. Finally, the ProtEmbed embedding space can be visualized, both at the global level and local to a given query, yielding intuition about the structure of protein sequence space.

We present a new learning strategy for classification problems in which train and/or test data suffer from missing features. In previous work, instances are represented as vectors from some feature space and one is forced to impute missing values or to consider an instance-specific subspace. In contrast, our method considers instances as sets of (feature,value) pairs which naturally handle the missing value case. Building onto this framework, we propose a classification strategy for sets. Our proposal maps (feature,value) pairs into an embedding space and then non-linearly combines the set of embedded vectors. The embedding and the combination parameters are learned jointly on the final classification objective. This simple strategy allows great flexibility in encoding prior knowledge about the features in the embedding step and yields advantageous results compared to alternative solutions over several datasets.

UNLABELLED: We present a large-scale implementation of the Rankprop protein homology ranking algorithm in the form of an openly accessible web server. We use the NRDB40 PSI-BLAST all-versus-all protein similarity network of 1.1 million proteins to construct the graph for the Rankprop algorithm, whereas previously, results were only reported for a database of 108 000 proteins. We also describe two algorithmic improvements to the original algorithm, including propagation from multiple homologs of the query and better normalization of ranking scores, that lead to higher accuracy and to scores with a probabilistic interpretation. AVAILABILITY: The Rankprop web server and source code are available at http://rankprop.gs.washington.edu

BACKGROUND: Predicting a protein's structural or functional class from its amino acid sequence or structure is a fundamental problem in computational biology. Recently, there has been considerable interest in using discriminative learning algorithms, in particular support vector machines (SVMs), for classification of proteins. However, because sufficiently many positive examples are required to train such classifiers, all SVM-based methods are hampered by limited coverage. RESULTS: In this study, we develop a hybrid machine learning approach for classifying proteins, and we apply the method to the problem of assigning proteins to structural categories based on their sequences or their 3D structures. The method combines a full-coverage but lower accuracy nearest neighbor method with higher accuracy but reduced coverage multiclass SVMs to produce a full coverage classifier with overall improved accuracy. The hybrid approach is based on the simple idea of "punting" from one method to another using a learned threshold. CONCLUSION: In cross-validated experiments on the SCOP hierarchy, the hybrid methods consistently outperform the individual component methods at all levels of coverage. Code and data sets are available at http://noble.gs.washington.edu/proj/sabretooth.

BACKGROUND: Predicting a protein's structural class from its amino acid sequence is a fundamental problem in computational biology. Much recent work has focused on developing new representations for protein sequences, called string kernels, for use with support vector machine (SVM) classifiers. However, while some of these approaches exhibit state-of-the-art performance at the binary protein classification problem, i.e. discriminating between a particular protein class and all other classes, few of these studies have addressed the real problem of multi-class superfamily or fold recognition. Moreover, there are only limited software tools and systems for SVM-based protein classification available to the bioinformatics community. RESULTS: We present a new multi-class SVM-based protein fold and superfamily recognition system and web server called SVM-Fold, which can be found at http://svm-fold.c2b2.columbia.edu. Our system uses an efficient implementation of a state-of-the-art string kernel for sequence profiles, called the profile kernel, where the underlying feature representation is a histogram of inexact matching k-mer frequencies. We also employ a novel machine learning approach to solve the difficult multi-class problem of classifying a sequence of amino acids into one of many known protein structural classes. Binary one-vs-the-rest SVM classifiers that are trained to recognize individual structural classes yield prediction scores that are not comparable, so that standard "one-vs-all" classification fails to perform well. Moreover, SVMs for classes at different levels of the protein structural hierarchy may make useful predictions, but one-vs-all does not try to combine these multiple predictions. To deal with these problems, our method learns relative weights between one-vs-the-rest classifiers and encodes information about the protein structural hierarchy for multi-class prediction. In large-scale benchmark results based on the SCOP database, our code weighting approach significantly improves on the standard one-vs-all method for both the superfamily and fold prediction in the remote homology setting and on the fold recognition problem. Moreover, our code weight learning algorithm strongly outperforms nearest-neighbor methods based on PSI-BLAST in terms of prediction accuracy on every structure classification problem we consider. CONCLUSION: By combining state-of-the-art SVM kernel methods with a novel multi-class algorithm, the SVM-Fold system delivers efficient and accurate protein fold and superfamily recognition.