Deploying Vision Transformer (ViT) models on edge devices poses significant challenges due to the high bandwidth, energy demands, and latency associated with transmitting large weight parameter sets to the sensing unit, along with limited on-chip memory resources, which are often insufficient for storing these parameters. To address these constraints, we present a software-hardware co-design framework that incorporates a novel in-sensor Compressed Weight Retrieval mechanism within an intelligent vision sensor. This framework offers two key contributions. First, we propose an innovative hardware-friendly weight compression algorithm that substantially reduces bandwidth and power consumption by optimizing on-chip memory usage for storing weight parameters. Second, we leverage the exceptional efficiency of Silicon Photonic (SiPh) devices and design a novel in-sensor accelerator called INSPIRE for the first time to perform in-sensor retrieval of the compressed weights and parallel fine-grained convolution operations next to the pixel array, enabling low-power adaptable ViT inference on resource-constrained edge platforms. Our extensive simulation results show that INSPIRE can remarkably reduce the memory footprint of ViT results with favorable accuracy. Besides, INSPIRE significantly reduces the bandwidth and power requirements associated with storing weight parameters in on-chip memory. INSPIRE achieves up to 245.4 Kilo FPS/W and reduces the data transfer energy by a factor of ∼11× on average compared with 4-bit quantized ViTs.

Semi-Supervised Federated Learning (SSFL) is gaining popularity over conventional Federated Learning in many real-world applications. Due to the practical limitation of limited labeled data on the client side, SSFL considers that participating clients train with unlabeled data, and only the central server has the necessary resources to access limited labeled data, making it an ideal fit for real-world applications (e.g., healthcare). However, traditional SSFL assumes that the data distributions in the training phase and testing phase are the same. In practice, however, domain shifts frequently occur, making it essential for SSFL to incorporate generalization capabilities and enhance their practicality. The core challenge is improving model generalization to new, unseen domains while the client participate in SSFL. However, the decentralized setup of SSFL and unsupervised client training necessitates innovation to achieve improved generalization across domains. To achieve this, we propose a novel framework called the Unified Alignment Protocol (UAP), which consists of an alternating two-stage training process. The first stage involves training the server model to learn and align the features with a parametric distribution, which is subsequently communicated to clients without additional communication overhead. The second stage proposes a novel training algorithm that utilizes the server feature distribution to align client features accordingly. Our extensive experiments on standard domain generalization benchmark datasets across multiple model architectures reveal that proposed UAP successfully achieves SOTA generalization performance in SSFL setting. Our code is available at https://github.com/ML-SecurityResearch-LAB/UAP.

Vision Transformers (ViTs) excel in tackling complex vision tasks, yet their substantial size poses significant challenges for applications on resource-constrained edge devices. The increased size of these models leads to higher overhead (e.g., energy, latency) when transmitting model weights between the edge device and the server. Hence, ViTs are not ideal for edge devices where the entire model may not fit on the device. Current model compression techniques often achieve high compression ratios at the expense of performance degradation, particularly for ViTs. To overcome the limitations of existing works, we rethink model compression strategy for ViTs from first principle approach and develop an orthogonal strategy called DeepCompress-ViT. The objective of the DeepCompress-ViT is to encode the model weights to a highly compressed encoded representation using a novel training method, denoted as Unified Compression Training (UCT). Proposed UCT is accompanied by a decoding mechanism during inference, which helps to gain any loss of accuracy due to high compression ratio. We further optimize this decoding step by reordering the decoding operation using associative property of matrix multiplication, ensuring that the compressed weights can be decoded during inference without incurring any computational overhead. Our extensive experiments across multiple ViT models on modern edge devices show that DeepCompress-ViT can successfully compress ViTs at high compression ratios (> 14×). DeepCompress-ViT enables the entire model to be stored on edge device, resulting in unprecedented reductions in energy consumption (> 1470×) and latency (> 68×) for edge ViT inference. Our code is available at https://github.com/ML-Security-Research-LAB/DeepCompress-ViT.

In text-to-image (T2I) generation, achieving fine-grained control over attributes - such as age or smile - remains challenging, even with detailed text prompts. Slider-based methods offer a solution for precise control of image attributes. Existing approaches typically train individual adapter for each attribute separately, overlooking the entanglement among multiple attributes. As a result, interference occurs among different attributes, preventing precise control of multiple attributes together. To address this challenge, we aim to disentangle multiple attributes in slider-based generation to enbale more reliable and independent attribute manipulation. Our approach, CompSlider, can generate a conditional prior for the T2I foundation model to control multiple attributes simultaneously. Furthermore, we introduce novel disentanglement and structure losses to compose multiple attribute changes while maintaining structural consistency within the image. Since CompSlider operates in the latent space of the conditional prior and does not require retraining the foundation model, it reduces the computational burden for both training and inference. We evaluate our approach on a variety of image attributes and highlight its generality by extending to video generation.

In this paper, we propose VidLA, an approach for video-language alignment at scale. There are two major limitations of previous video-language alignment approaches. First, they do not capture both short-range and long-range temporal dependencies and typically employ complex hierarchical deep network architectures that are hard to integrate with existing pretrained image-text foundation models. To effectively address this limitation, we instead keep the network architecture simple and use a set of data tokens that operate at different temporal resolutions in a hierarchical manner, accounting for the temporally hierarchical nature of videos. By employing a simple two-tower architecture, we are able to initialize our video-language model with pretrained image-text foundation models, thereby boosting the final performance. Second, existing video-language alignment works struggle due to the lack of semantically aligned large-scale training data. To overcome it, we leverage recent LLMs to curate the largest video-language dataset to date with better visual grounding. Furthermore, unlike existing video-text datasets which only contain short clips, our dataset is enriched with video clips of varying durations to aid our temporally hierarchical data to-kens in extracting better representations at varying temporal scales. Overall, empirical results show that our proposed approach surpasses state-of-the-art methods on Multiple retrieval benchmarks, especially on longer videos, and performs competitively on classification benchmarks.

Self-supervised learning has unlocked the potential of scaling up pretraining to billions of images, since annotation is unnecessary. But are we making the best use of data? How more economical can we be? In this work, we attempt to answer this question by making two contributions. First, we investigate first-person videos and introduce a ``Walking Tours'' dataset. These videos are high-resolution, hours-long, captured in a single uninterrupted take, depicting a large number of objects and actions with natural scene transitions. They are unlabeled and uncurated, thus realistic for self-supervision and comparable with human learning. Second, we introduce a novel self-supervised image pretraining method tailored for learning from continuous videos. Existing methods typically adapt image-based pretraining approaches to incorporate more frames. Instead, we advocate a ``tracking to learn to recognize'' approach. Our method called DoRA, leads to attention maps that **D**isc**O**ver and t**RA**ck objects over time in an end-to-end manner, using transformer cross-attention. We derive multiple views from the tracks and use them in a classical self-supervised distillation loss. Using our novel approach, a single Walking Tours video remarkably becomes a strong competitor to ImageNet for several image and video downstream tasks.

Semi-Supervised Learning can be more beneficial for the video domain compared to images because of its higher an-notation cost and dimensionality. Besides, any video understanding task requires reasoning over both spatial and temporal dimensions. In order to learn both the static and motion related features for the semi-supervised action recognition task, existing methods rely on hard in-put inductive biases like using two-modalities (RGB and Optical-flow) or two-stream of different playback rates. Instead of utilizing unlabeled videos through diverse in-put streams, we rely on self-supervised video represen-tations, particularly, we utilize temporally-invariant and temporally-distinctive representations. We observe that these representations complement each other depending on the nature of the action. Based on this observation, we propose a student-teacher semi-supervised learning frame-work, TimeBalance, where we distill the knowledge from a temporally-invariant and a temporally-distinctive teacher. Depending on the nature of the unlabeled video, we dy-namically combine the knowledge of these two teach-ers based on a novel temporal similarity-based reweighting scheme. Our method achieves state-of-the-art performance on three action recognition benchmarks: UCF101, HMDB51, and Kinetics400. Code: https://github.com/DAVEISHAN/TimeBalance.

Weakly-supervised Temporal Action Localization (WTAL) attempts to localize the actions in untrimmed videos using only video-level supervision. Most recent works approach WTAL from a localization-by-classification perspective where these methods try to classify each video frame followed by a manually-designed post-processing pipeline to aggregate these per-frame action predictions into action snippets. Due to this perspective, the model lacks any explicit understanding of action boundaries and tends to focus only on the most discriminative parts of the video resulting in incomplete action localization. To address this, we present PivoTAL, Prior-driven Supervision for Weakly-supervised Temporal Action Localization, to approach WTAL from a localization-by-localization perspective by learning to localize the action snippets directly. To this end, PivoTAL leverages the underlying spatio-temporal regularities in videos in the form of action-specific scene prior, action snippet generation prior, and learnable Gaussian prior to supervise the localization-based training. PivoTAL shows significant improvement (of at least 3% avg mAP) over all existing methods on the benchmark datasets, THUMOS-14 and ActivitNet-v1.3.

Source-free domain adaptation (SFDA) is a popular unsupervised domain adaptation method where a pre-trained model from a source domain is adapted to a target domain without accessing any source data. Despite rich results in this area, existing literature overlooks the security challenges of the unsupervised SFDA setting in presence of a malicious source domain owner. This work investigates the effect of a source adversary which may inject a hidden malicious behavior (Backdoor/Trojan) during source training and potentially transfer it to the target domain even after benign training by the victim (target domain owner). Our investigation of the current SFDA setting reveals that because of the unique challenges present in SFDA (e.g., no source data, target label), defending against backdoor attack using existing defenses become practically ineffective in protecting the target model. To address this, we propose a novel target domain protection scheme called secure source-free domain adaptation (SSDA). SSDA adopts a single-shot model compression of a pre-trained source model and a novel knowledge transfer scheme with a spectral-norm-based loss penalty for target training. The proposed static compression and the dynamic training loss penalty are designed to suppress the malicious channels responsive to the backdoor during the adaptation stage. At the same time, the knowledge transfer from an uncompressed auxiliary model helps to recover the benign test accuracy. Our extensive evaluation on multiple dataset and domain tasks against recent backdoor attacks reveal that the proposed SSDA can successfully defend against strong backdoor attacks with little to no degradation in test accuracy compared to the vulnerable baseline SFDA methods. Our code is available at https://github.com/ML-Security-Research-LAB/SSDA.

Few-shot video action recognition is an effective approach to recognizing new categories with only a few labeled examples, thereby reducing the challenges associated with collecting and annotating large-scale video datasets. Existing methods in video action recognition rely on large labeled datasets from the same domain. However, this setup is not realistic as novel categories may come from different data domains that may have different spatial and temporal characteristics. This dissimilarity between the source and target domains can pose a significant challenge, rendering traditional few-shot action recognition techniques ineffective. To address this issue, in this work, we propose a novel cross-domain few-shot video action recognition method that leverages self-supervised learning and curriculum learning to balance the information from the source and target domains. To be particular, our method employs a masked autoencoder-based self-supervised training objective to learn from both source and target data in a self-supervised manner. Then a progressive curriculum balances learning the discriminative information from the source dataset with the generic information learned from the target domain. Initially, our curriculum utilizes supervised learning to learn class discriminative features from the source data. As the training progresses, we transition to learning target-domain-specific features. We propose a progressive curriculum to encourage the emergence of rich features in the target domain based on class discriminative supervised features in the source domain. We evaluate our method on several challenging benchmark datasets and demonstrate that our approach outperforms existing cross-domain few-shot learning techniques. Our code is available at https://github.com/Sarinda251/CDFSL-V

Self-supervised learning on large-scale multi-modal datasets allows learning semantically meaningful embeddings in a joint multi-modal representation space without relying on human annotations. These joint embeddings enable zero-shot cross-modal tasks like retrieval and classification. However, these methods often struggle to generalize well on out-of-domain data as they ignore the semantic structure present in modality-specific embeddings. In this context, we propose a novel Semantic-Structure-Preserving Consistency approach to improve generalizability by preserving the modality-specific relationships in the joint embedding space. To capture modality-specific semantic relationships between samples, we propose to learn multiple anchors and represent the multifaceted relationship between samples with respect to their relationship with these anchors. To assign multiple anchors to each sample, we propose a novel Multi-Assignment Sinkhorn-Knopp algorithm. Our experimentation demonstrates that our proposed approach learns semantically meaningful anchors in a self-supervised manner. Furthermore, our evaluation on MSR-VTT and YouCook2 datasets demonstrates that our proposed multi-anchor assignment based solution achieves state-of-the-art performance and generalizes to both in-and out-of-domain datasets. Code: https://github.com/Swetha5/Multi_Sinkhorn_Knopp

Supervised deep learning methods require a large repository of annotated data; hence, label noise is inevitable. Training with such noisy data negatively impacts the generalization performance of deep neural networks. To combat label noise, recent state-of-the-art methods employ some sort of sample selection mechanism to select a possibly clean subset of data. Next, an off-the-shelf semi-supervised learning method is used for training where rejected samples are treated as unlabeled data. Our comprehensive analysis shows that current selection methods disproportionately select samples from easy (fast learnable) classes while rejecting those from relatively harder ones. This creates class imbalance in the selected clean set and in turn, deteriorates performance under high label noise. In this work, we propose UNICON, a simple yet effective sample selection method which is robust to high label noise. To address the disproportionate selection of easy and hard samples, we introduce a Jensen-Shannon divergence based uniform selection mechanism which does not require any probabilistic modeling and hyperparameter tuning. We complement our selection method with contrastive learning to further combat the memorization of noisy labels. Extensive experimentation on multiple benchmark datasets demonstrates the effectiveness of UNICON; we obtain an 11.4% improvement over the current state-of-the-art on CIFAR100 dataset with a 90% noise rate. Our code is publicly available.11https://github.com/nazmul-karim170/UNICON-Noisy-Label

Contrastive learning has nearly closed the gap between supervised and self-supervised learning of image representations, and has also been explored for videos. However, prior work on contrastive learning for video data has not explored the effect of explicitly encouraging the features to be distinct across the temporal dimension. We develop a new temporal contrastive learning framework consisting of two novel losses to improve upon existing contrastive self-supervised video representation learning methods. The local–local temporal contrastive loss adds the task of discriminating between non-overlapping clips from the same video, whereas the global–local temporal contrastive aims to discriminate between timesteps of the feature map of an input clip in order to increase the temporal diversity of the learned features. Our proposed temporal contrastive learning framework achieves significant improvement over the state-of-the-art results in various downstream video understanding tasks such as action recognition, limited-label action classification, and nearest-neighbor video retrieval on multiple video datasets and backbones. We also demonstrate significant improvement in fine-grained action classification for visually similar classes. With the commonly used 3D ResNet-18 architecture with UCF101 pretraining, we achieve 82.4% (+5.1% increase over the previous best) top-1 accuracy on UCF101 and 52.9% (+5.4% increase) on HMDB51 action classification, and 56.2% (+11.7% increase) Top-1 Recall on UCF101 nearest neighbor video retrieval. Code released at https://github.com/DAVEISHAN/TCLR.

In many real-world problems, collecting a large number of labeled samples is infeasible. Few-shot learning (FSL) is the dominant approach to address this issue, where the objective is to quickly adapt to novel categories in presence of a limited number of samples. FSL tasks have been predominantly solved by leveraging the ideas from gradient-based meta-learning and metric learning approaches. However, recent works have demonstrated the significance of powerful feature representations with a simple embedding network that can outperform existing sophisticated FSL algorithms. In this work, we build on this insight and propose a novel training mechanism that simultaneously enforces equivariance and invariance to a general set of geometric transformations. Equivariance or invariance has been employed standalone in the previous works; however, to the best of our knowledge, they have not been used jointly. Simultaneous optimization for both of these contrasting objectives allows the model to jointly learn features that are not only independent of the input transformation but also the features that encode the structure of geometric transformations. These complementary sets of features help generalize well to novel classes with only a few data samples. We achieve additional improvements by incorporating a novel self-supervised distillation objective. Our extensive experimentation shows that even without knowledge distillation our proposed method can outperform current state-of-the-art FSL methods on five popular benchmark datasets.

Activity detection in security videos is a difficult problem due to multiple factors such as large field of view, presence of multiple activities, varying scales and viewpoints, and its untrimmed nature. The existing research in activity detection is mainly focused on datasets, such as UCF-101, JHMDB, THUMOS, and AVA, which partially address these issues. The requirement of processing security videos in real-time makes this even more challenging. In this work, we propose Gabriella, a real-time online system to perform activity detection on untrimmed security videos. The proposed method consists of three stages: tubelet extraction, activity classification, and online tubelet merging. For tubelet extraction, we propose a localization network which takes a video clip as input and spatio-temporally detects potential foreground regions at multiple scales to generate action tubelets. We propose a novel Patch-Dice loss to handle large variations in actor size. Our online processing of videos at a clip level drastically reduces the computation time in detecting activities. The detected tubelets are assigned activity class scores by the classification network and merged together using our proposed Tubelet-Merge Action-Split (TMAS) algorithm to form the final action detections. The TMAS algorithm efficiently connects the tubelets in an online fashion to generate action detections which are robust against varying length activities. We perform our experiments on the VIRAT and MEVA (Multiview Extended Video with Activities) datasets and demonstrate the effectiveness of the proposed approach in terms of speed (~100 fps) and performance with state-of-the-art results. More details about this work are available on our project webpage11https://www.crcv.ucf.edu/research/projects/gabriella-an-online-system-for-real-time-activity-detection-in-untrimmed-security-videos.