In this paper, we introduce FROSTER, an effective framework for open-vocabulary action recognition. The CLIP model has achieved remarkable success in a range of image-based tasks, benefiting from its strong generalization capability stemming from pretaining on massive image-text pairs. However, applying CLIP directly to the open-vocabulary action recognition task is challenging due to the absence of temporal information in CLIP's pretraining. Further, fine-tuning CLIP on action recognition datasets may lead to overfitting and hinder its generalizability, resulting in unsatisfactory results when dealing with unseen actions. To address these issues, FROSTER employs a residual feature distillation approach to ensure that CLIP retains its generalization capability while effectively adapting to the action recognition task. Specifically, the residual feature distillation treats the frozen CLIP model as a teacher to maintain the generalizability exhibited by the original CLIP and supervises the feature learning for the extraction of video-specific features to bridge the gap between images and videos. Meanwhile, it uses a residual sub-network for feature distillation to reach a balance between the two distinct objectives of learning generalizable and video-specific features. We extensively evaluate FROSTER on open-vocabulary action recognition benchmarks under both base-to-novel and cross-dataset settings. FROSTER consistently achieves state-of-the-art performance on all datasets across the board. Project page: https://visual-ai.github.io/froster.
In recent times, significant advancements have been made in the field of large language models (LLMs), represented by GPT series models. To optimize task execution, users often engage in multi-round conversations with GPT models hosted in cloud environments. These multi-round conversations, potentially replete with private information, require transmission and storage within the cloud. However, this operational paradigm introduces additional attack surfaces. In this paper, we first introduce a specific Conversation Reconstruction Attack targeting GPT models. Our introduced Conversation Reconstruction Attack is composed of two steps: hijacking a session and reconstructing the conversations. Subsequently, we offer an exhaustive evaluation of the privacy risks inherent in conversations when GPT models are subjected to the proposed attack. However, GPT-4 demonstrates certain robustness to the proposed attacks. We then introduce two advanced attacks aimed at better reconstructing previous conversations, specifically the UNR attack and the PBU attack. Our experimental findings indicate that the PBU attack yields substantial performance across all models, achieving semantic similarity scores exceeding 0.60, while the UNR attack is effective solely on GPT-3.5. Our results reveal the concern about privacy risks associated with conversations involving GPT models and aim to draw the community's attention to prevent the potential misuse of these models' remarkable capabilities. We will responsibly disclose our findings to the suppliers of related large language models.
Omnidirectional and 360{\deg} images are becoming widespread in industry and in consumer society, causing omnidirectional computer vision to gain attention. Their wide field of view allows the gathering of a great amount of information about the environment from only an image. However, the distortion of these images requires the development of specific algorithms for their treatment and interpretation. Moreover, a high number of images is essential for the correct training of computer vision algorithms based on learning. In this paper, we present a tool for generating datasets of omnidirectional images with semantic and depth information. These images are synthesized from a set of captures that are acquired in a realistic virtual environment for Unreal Engine 4 through an interface plugin. We gather a variety of well-known projection models such as equirectangular and cylindrical panoramas, different fish-eye lenses, catadioptric systems, and empiric models. Furthermore, we include in our tool photorealistic non-central-projection systems as non-central panoramas and non-central catadioptric systems. As far as we know, this is the first reported tool for generating photorealistic non-central images in the literature. Moreover, since the omnidirectional images are made virtually, we provide pixel-wise information about semantics and depth as well as perfect knowledge of the calibration parameters of the cameras. This allows the creation of ground-truth information with pixel precision for training learning algorithms and testing 3D vision approaches. To validate the proposed tool, different computer vision algorithms are tested as line extractions from dioptric and catadioptric central images, 3D Layout recovery and SLAM using equirectangular panoramas, and 3D reconstruction from non-central panoramas.
Activity detection is an important task in the next generation grant-free multiple access. While there are a number of existing algorithms designed for this purpose, they mostly require precise information about the network, such as large-scale fading coefficients, small-scale fading channel statistics, noise variance at the access points, and user activity probability. Acquiring these information would take a significant overhead and their estimated values might not be accurate. This problem is even more severe in cell-free networks as there are many of these parameters to be acquired. Therefore, this paper sets out to investigate the activity detection problem without the above-mentioned information. In order to handle so many unknown parameters, this paper employs the Bayesian approach, where the unknown variables are endowed with prior distributions which effectively act as regularizations. Together with the likelihood function, a maximum a posteriori (MAP) estimator and a variational inference algorithm are derived. Extensive simulations demonstrate that the proposed methods, even without the knowledge of these system parameters, perform better than existing state-of-the-art methods, such as covariance-based and approximate message passing methods.
3D human pose estimation captures the human joint points in three-dimensional space while keeping the depth information and physical structure. That is essential for applications that require precise pose information, such as human-computer interaction, scene understanding, and rehabilitation training. Due to the challenges in data collection, mainstream datasets of 3D human pose estimation are primarily composed of multi-view video data collected in laboratory environments, which contains rich spatial-temporal correlation information besides the image frame content. Given the remarkable self-attention mechanism of transformers, capable of capturing the spatial-temporal correlation from multi-view video datasets, we propose a multi-stage framework for 3D sequence-to-sequence (seq2seq) human pose detection. Firstly, the spatial module represents the human pose feature by intra-image content, while the frame-image relation module extracts temporal relationships and 3D spatial positional relationship features between the multi-perspective images. Secondly, the self-attention mechanism is adopted to eliminate the interference from non-human body parts and reduce computing resources. Our method is evaluated on Human3.6M, a popular 3D human pose detection dataset. Experimental results demonstrate that our approach achieves state-of-the-art performance on this dataset.
YouTube faces a global crisis with the dissemination of false information and hate speech. To counter these issues, YouTube has implemented strict rules against uploading content that includes false information or promotes hate speech. While numerous studies have been conducted to reduce offensive English-language content, there's a significant lack of research on Sinhala content. This study aims to address the aforementioned gap by proposing a solution to minimize the spread of violence and misinformation in Sinhala YouTube videos. The approach involves developing a rating system that assesses whether a video contains false information by comparing the title and description with the audio content and evaluating whether the video includes hate speech. The methodology encompasses several steps, including audio extraction using the Pytube library, audio transcription via the fine-tuned Whisper model, hate speech detection employing the distilroberta-base model and a text classification LSTM model, and text summarization through the fine-tuned BART-Large- XSUM model. Notably, the Whisper model achieved a 48.99\% word error rate, while the distilroberta-base model demonstrated an F1 score of 0.856 and a recall value of 0.861 in comparison to the LSTM model, which exhibited signs of overfitting.
Large Language Models (LLMs) have emerged as a pivotal force in language technology. Their robust reasoning capabilities and expansive knowledge repositories have enabled exceptional zero-shot generalization abilities across various facets of the natural language processing field, including information retrieval (IR). In this paper, we conduct an in-depth investigation into the utility of documents generated by LLMs for IR. We introduce a simple yet effective framework, Multi-Text Generation Integration (MuGI), to augment existing IR methodologies. Specifically, we prompt LLMs to generate multiple pseudo references and integrate with query for retrieval. The training-free MuGI model eclipses existing query expansion strategies, setting a new standard in sparse retrieval. It outstrips supervised counterparts like ANCE and DPR, achieving a notable over 18% enhancement in BM25 on the TREC DL dataset and a 7.5% increase on BEIR. Through MuGI, we have forged a rapid and high-fidelity re-ranking pipeline. This allows a relatively small 110M parameter retriever to surpass the performance of larger 3B models in in-domain evaluations, while also bridging the gap in out-of-distribution situations. We release our code and all generated references at https://github.com/lezhang7/Retrieval_MuGI.
Neuron labeling is an approach to visualize the behaviour and respond of a certain neuron to a certain pattern that activates the neuron. Neuron labeling extract information about the features captured by certain neurons in a deep neural network, one of which uses the encoder-decoder image captioning approach. The encoder used can be a pretrained CNN-based model and the decoder is an RNN-based model for text generation. Previous work, namely MILAN (Mutual Information-guided Linguistic Annotation of Neuron), has tried to visualize the neuron behaviour using modified Show, Attend, and Tell (SAT) model in the encoder, and LSTM added with Bahdanau attention in the decoder. MILAN can show great result on short sequence neuron captioning, but it does not show great result on long sequence neuron captioning, so in this work, we would like to improve the performance of MILAN even more by utilizing different kind of attention mechanism and additionally adding several attention result into one, in order to combine all the advantages from several attention mechanism. Using our compound dataset, we obtained higher BLEU and F1-Score on our proposed model, achieving 17.742 and 0.4811 respectively. At some point where the model converges at the peak, our model obtained BLEU of 21.2262 and BERTScore F1-Score of 0.4870.
We introduce the concept of multiple temporal perspectives, a novel approach applicable to Recurrent Neural Network (RNN) architectures for enhancing their understanding of sequential data. This method involves maintaining diverse temporal views of previously encountered text, significantly enriching the language models' capacity to interpret context. To show the efficacy of this approach, we incorporate it into the Receptance Weighted Key Value (RWKV) architecture, addressing its inherent challenge of retaining all historical information within a single hidden state. Notably, this improvement is achieved with a minimal increase in the number of parameters --even as little as $0.04\%$ of the original number of parameters. Further, the additional parameters necessary for the multiple temporal perspectives are fine-tuned with minimal computational overhead, avoiding the need for a full pre-training. The resulting model maintains linear computational complexity during prompt inference, ensuring consistent efficiency across various sequence lengths. The empirical results and ablation studies included in our research validate the effectiveness of our approach, showcasing improved performance across multiple benchmarks. The code, model weights and datasets are open-sourced at: https://github.com/RazvanDu/TemporalRNNs.
In this paper, we find that ubiquitous time series (TS) forecasting models are prone to severe overfitting. To cope with this problem, we embrace a de-redundancy approach to progressively reinstate the intrinsic values of TS for future intervals. Specifically, we renovate the vanilla Transformer by reorienting the information aggregation mechanism from addition to subtraction. Then, we incorporate an auxiliary output branch into each block of the original model to construct a highway leading to the ultimate prediction. The output of subsequent modules in this branch will subtract the previously learned results, enabling the model to learn the residuals of the supervision signal, layer by layer. This designing facilitates the learning-driven implicit progressive decomposition of the input and output streams, empowering the model with heightened versatility, interpretability, and resilience against overfitting. Since all aggregations in the model are minus signs, which is called Minusformer. Extensive experiments demonstrate the proposed method outperform existing state-of-the-art methods, yielding an average performance improvement of 11.9% across various datasets.