We propose a memory efficient method, named Stochastic Backpropagation (SBP), for training deep neural networks on videos. It is based on the finding that gradients from incomplete execution for backpropagation can still effectively train the models with minimal accuracy loss, which attributes to the high redundancy of video. SBP keeps all forward paths but randomly and independently removes the backward paths for each network layer in each training step. It reduces the GPU memory cost by eliminating the need to cache activation values corresponding to the dropped backward paths, whose amount can be controlled by an adjustable keep-ratio. Experiments show that SBP can be applied to a wide range of models for video tasks, leading to up to 80.0% GPU memory saving and 10% training speedup with less than 1% accuracy drop on action recognition and temporal action detection.

We propose a novel one-stage Transformer-based semantic and spatial refined transformer (SSRT) to solve the Human-Object Interaction detection task, which requires to localize humans and objects, and predicts their interactions. Differently from previous Transformer-based HOI approaches, which mostly focus at improving the design of the decoder outputs for the final detection, SSRT introduces two new modules to help select the most relevant object-action pairs within an image and refine the queries' representation using rich semantic and spatial features. These enhancements lead to state-of-the-art results on the two most popular HOI benchmarks: V-COCO and HICO-DET.

We propose TubeR: a simple solution for spatio-temporal video action detection. Different from existing methods that depend on either an offline actor detector or hand-designed actor-positional hypotheses like proposals or anchors, we propose to directly detect an action tubelet in a video by simultaneously performing action localization and recognition from a single representation. TubeR learns a set of tubelet-queries and utilizes a tubelet-attention module to model the dynamic spatio-temporal nature of a video clip, which effectively reinforces the model capacity compared to using actor-positional hypotheses in the spatio-temporal space. For videos containing transitional states or scene changes, we propose a context aware classification head to utilize short-term and long-term context to strengthen action classification, and an action switch regression head for detecting the precise temporal action extent. TubeR directly produces action tubelets with variable lengths and even maintains good results for long video clips. TubeR outperforms the previous state-of-the-art on commonly used action detection datasets AVA, UCF101-24 and JHMDB51-21. Code will be available on GluonCV(https://cv.gluon.ai/).

In this paper, we provide an in-depth study of Stochastic Backpropagation (SBP) when training deep neural networks for standard image classification and object detection tasks. During backward propagation, SBP calculates gradients by using only a subset of feature maps to save GPU memory and computational cost. We interpret SBP as an efficient way to implement stochastic gradient decent by performing backpropagation dropout, which leads to significant memory saving and training run-time reduction, with a minimal impact on the overall model accuracy. We offer best practices to apply SBP for training image recognition models, which can be adopted in learning a wide range of deep neural networks. Experiments on image classification and object detection show that SBP can save up to 40% of GPU memory with less than 1% accuracy degradation. Code is available at: https://github.com/amazon-research/stochastic-backpropagation

In the world of action recognition research, one primary focus has been on how to construct and train networks to model the spatial-temporal volume of an input video. These methods typically uniformly sample a segment of an input clip (along the temporal dimension). However, not all parts of a video are equally important to determine the action in the clip. In this work, we focus instead on learning where to extract features, so as to focus on the most informative parts of the video. We propose a method called the non-uniform temporal aggregation (NUTA), which aggregates features only from informative temporal segments. We also introduce a synchronization method that allows our NUTA features to be temporally aligned with traditional uniformly sampled video features, so that both local and clip-level features can be combined. Our model has achieved state-of-the-art performance on four widely used large-scale action-recognition datasets (Kinetics400, Kinetics700, Something-something V2 and Charades). In addition, we have created a visualization to illustrate how the proposed NUTA method selects only the most relevant parts of a video clip.

Temporal action segmentation is a task to classify each frame in the video with an action label. However, it is quite expensive to annotate every frame in a large corpus of videos to construct a comprehensive supervised training dataset. Thus in this work we propose an unsupervised method, namely SSCAP, that operates on a corpus of unlabeled videos and predicts a likely set of temporal segments across the videos. SSCAP leverages Self-Supervised learning to extract distinguishable features and then applies a novel Co-occurrence Action Parsing algorithm to not only capture the correlation among sub-actions under-lying the structure of activities, but also estimate the temporal path of the sub-actions in an accurate and general way. We evaluate on both classic datasets (Breakfast, 50Sal-ads) and the emerging fine-grained action dataset (Fine-Gym) with more complex activity structures and similar sub-actions. Results show that SSCAP achieves state-of-the-art performance on all datasets and can even outperform some weakly-supervised approaches, demonstrating its effectiveness and generalizability.

Most action recognition solutions rely on dense sampling to precisely cover the informative temporal clip. Extensively searching temporal region is expensive for a real-world application. In this work, we focus on improving the inference efficiency of current action recognition backbones on trimmed videos, and illustrate that an action model can accurately classify an action with a single pass over the video unlike the multi-clip sampling common with SOTA by learning to drop non-informative features. We present Selective Feature Compression (SFC), an action recognition inference strategy that greatly increases model inference efficiency without compromising accuracy. Different from previous works that compress kernel size and decrease the channel dimension, we propose to compress features along the spatio-temporal dimensions without the need to change backbone parameters. Our experiments on Kinetics-400, UCF101 and ActivityNet show that SFC is able to reduce inference speed by 6-7x and memory usage by 5-6x compared with the commonly used 30 crop dense sampling procedure, while also slightly improving Top1 Accuracy. We perform thorough quantitative and qualitative evaluation and show how our SFC learns to attend to important video regions for the task of action recognition.

We introduce Video Transformer (VidTr) with separable-attention for video classification. Comparing with commonly used 3D networks, VidTr is able to aggregate spatiotemporal information via stacked attentions and provide better performance with higher efficiency. We first introduce the vanilla video transformer and show that transformer module is able to perform spatio-temporal modeling from raw pixels, but with heavy memory usage. We then present VidTr which reduces the memory cost by 3.3× while keeping the same performance. To further optimize the model, we propose the standard deviation based topK pooling for attention (pooltopK_std), which reduces the computation by dropping non-informative features along temporal dimension. VidTr achieves state-of-the-art performance on five commonly used datasets with lower computational requirement, showing both the efficiency and effectiveness of our design. Finally, error analysis and visualization show that VidTr is especially good at predicting actions that require long-term temporal reasoning.

We present Long Short-term TRansformer (LSTR), a temporal modeling algorithm for online action detection, which employs a long- and short-term memory mechanism to model prolonged sequence data. It consists of an LSTR encoder that dynamically leverages coarse-scale historical information from an extended temporal window (e.g., 2048 frames spanning of up to 8 minutes), together with an LSTR decoder that focuses on a short time window (e.g., 32 frames spanning 8 seconds) to model the fine-scale characteristics of the data. Compared to prior work, LSTR provides an effective and efficient method to model long videos with fewer heuristics, which is validated by extensive empirical analysis. LSTR achieves state-of-the-art performance on three standard online action detection benchmarks, THUMOS'14, TVSeries, and HACS Segment. Code has been made available at: https://xumingze0308.github.io/projects/lstr.

To track security and compliance requirements and perform problem diagnosis, administrators of cloud computing systems need to monitor significant system changes occurring on the set of cloud instances under their supervision. Considering the large number of instances (virtual machines, containers) possibly operating under multiple configurations, this is a difficult-to-track process. Standard solutions to this problem rely on manually-created rules to identify changes. These techniques suffer from a limited scope, rely on domain expertise, and are time-consuming and error-prone. Recently, more streamlined approaches that automatically determine the type of individual system changes have been proposed, but these techniques assume that system states right before and after each individual change can be captured, a rather difficult requirement to enforce in real world usage. This paper proposes DeltaSherlock, a practical system change discovery framework that can capture system states on-demand and detect multiple system changes between them. We evaluate DeltaSherlock over 25,000 system changes caused by software installations collected from virtual machines (VMs) deployed over a commercial cloud. DeltaSherlock can accurately identify multiple software installations with 96.8% accuracy when supplied with a non-overlapping record of system changes and with 77.8% accuracy when supplied with random irregular observations possibly containing overlapping or incomplete changes.

To accommodate the increasing presence of volatile and intermittent renewable energy sources in power generation, independent system operators (ISO) offer opportunities for demand side regulation service (RS) so as to stabilize the grid load. These power market features allow the demand side to earn monetary credits by modulating its power consumption dynamically following an RS signal broadcast by ISO. This paper studies the capacities and benefits of a major potential demand side, the data center, to provide RS. We propose a dynamic control policy that modulates the data center power consumption in response to ISO requests by leveraging server power capping techniques and various server power states. Results demonstrate that using our policy, data centers can provide fast reserves in quantities that are substantial proportions (around 50%) of their average energy consumption, with no major deterioration in quality of service (QoS). By doing so, data centers decrease their energy costs around 50%, while providing the ISOs and the society in general with cost effective demand side reserves that render massive renewable generation adoption affordable.

Discovering and identifying system changes caused by events such as software installation and updates, configuration changes, and security patches are important functionalities for change management, security, compliance and problem diagnosis in emerging cloud platforms. Currently, most discovery tools use manually written rules, which require specific knowledge of software and systems. Approaches based on manually written rules are often fragile and require constant maintenance in this era of continuous integration. In this paper, we propose a novel “discovery by example” approach to autonomously search for and identify system changes. Our approach learns characteristic features of system changes automatically, without requiring any explicit rule definitions or specific knowledge of the underlying software or systems. In this approach, given a system change, our method searches a repository that contains previous stored system changes and returns those that are similar to it. We further explore the use of various forms of “fingerprints” to represent system changes efficiently and faithfully in a compact manner. We propose and evaluate two types of fingerprints: the “basename fingerprint” and the “1-D histogram fingerprint”. We show that both fingerprints exhibit different efficiency and accuracy trade-offs, and they can be effectively employed in different use cases. We evaluate the performance of our approach with both techniques and further present an application of it in system real-time streaming monitoring.

Today's US power markets offer new opportunities for the energy consumers to reduce their energy costs by first promising an average consumption rate for the next hour and then by following a regulation signal broadcast by the independent system operators (ISOs), who need to match supply and demand in real time in presence of volatile and intermittent renewable energy generation. This paper leverages the power regulation capabilities of the servers so as to enable the data centers to participate in these emerging power markets. As the data center energy consumption continues to grow, proposed participation in the power markets has the promise to achieve significant monetary savings. The paper first solves a data center regulation service (RS) optimization problem to determine the optimal average power consumption and regulation quantity that minimize the energy cost. We then propose a dynamic server power capping technique to modulate the real-time power consumption in response to ISO requests while maintaining the desired quality-of-service (QoS). Experiments on a real-life server demonstrate that our technique can reduce the energy cost by 29% on average compared to using a fixed power cap.