Large vision-language models (VLMs) enable intuitive visual search using natural language queries. However, improving their performance often requires fine-tuning and scaling to larger model variants. In this work, we propose a mechanism inspired by traditional text-based search to improve retrieval performance at inference time: relevance feedback. While relevance feedback can serve as an alternative to fine-tuning, its model-agnostic design also enables use with fine-tuned VLMs. Specifically, we introduce and evaluate four feedback strategies for VLM-based retrieval. First, we revise classical pseudo-relevance feedback (PRF), which refines query embeddings based on top-ranked results. To address its limitations, we propose generative relevance feedback (GRF), which uses synthetic captions for query refinement. Furthermore, we introduce an attentive feedback summarizer (AFS), a custom transformer-based model that integrates multimodal fine-grained features from relevant items. Finally, we simulate explicit feedback using ground-truth captions as an upper-bound baseline. Experiments on Flickr30k and COCO with the VLM backbones show that GRF, AFS, and explicit feedback improve retrieval performance by 3–5% in MRR@5 for smaller VLMs, and 1–3% for larger ones, compared to retrieval with no feedback. Moreover, AFS, similarly to explicit feedback, mitigates query drift and is more robust than GRF in iterative, multi-turn retrieval settings. Our findings demonstrate that relevance feedback can consistently enhance retrieval across VLMs and open up opportunities for interactive and adaptive visual search.

In the absence of interaction history, image recommendations often depend on content-based approaches. Prompted by user queries in natural language, such systems rank items based on the similarity between textual and visual features. However, these approaches typically rely on static queries and do not offer alternative feedback mechanisms. In this paper, we present VisualReF: an interactive image retrieval prototype that introduces visual relevance feedback through fine-grained user annotations. Built on vision-language models (VLMs) for retrieval, our system allows users to label relevant and irrelevant regions in retrieved images. These regions are captioned using a generative vision-language model to refine the query vector. Our work bridges the gap between conventional static image retrieval and interactive, user-guided search by introducing visual relevance feedback. Finally, our prototype contributes to the field of visual recommendation by empowering researchers with practical tools for: (i) collecting region-level visual relevance signals from users, (ii) supporting integration of human feedback into interactive search pipelines, and (iii) explaining how the relevance feedback model perceives user input.

In face-to-face dialogues, the form-meaning relationship of co-speech gestures varies depending on contextual factors such as what the gestures refer to and the individual characteristics of speakers. These factors make co-speech gesture representation learning challenging. How can we learn meaningful gestures representations considering gestures’ variability and relationship with speech? This paper tackles this challenge by employing self-supervised contrastive learning techniques to learn gesture representations from skeletal and speech information. We propose an approach that includes both unimodal and multimodal pre-training to ground gesture representations in co-occurring speech. For training, we utilize a face-to-face dialogue dataset rich with representational iconic gestures. We conduct thorough intrinsic evaluations of the learned representations through comparison with human-annotated pairwise gesture similarity. Moreover, we perform a diagnostic probing analysis to assess the possibility of recovering interpretable gesture features from the learned representations. Our results show a significant positive correlation with human-annotated gesture similarity and reveal that the similarity between the learned representations is consistent with well-motivated patterns related to the dynamics of dialogue interaction. Moreover, our findings demonstrate that several features concerning the form of gestures can be recovered from the latent representations. Overall, this study shows that multimodal contrastive learning is a promising approach for learning gesture representations, which opens the door to using such representations in larger-scale gesture analysis studies.

Recent advancements in Deep and Self-Supervised Learning (SSL) have led to substantial improvements in Speech Emotion Recognition (SER) performance, reaching unprecedented levels. However, obtaining sufficient amounts of accurately labeled data for training or fine-tuning the models remains a costly and challenging task. In this paper, we propose a multiview SSL pre-training technique that can be applied to various representations of speech, including the ones generated by large speech models, to improve SER performance in scenarios where annotations are limited. Our experiments, based on wav2vec 2.0, spectral and paralinguistic features, demonstrate that the proposed framework boosts the SER performance, by up to 10% in Unweighted Average Recall, in settings with extremely sparse data annotations.

In recent years, self-supervised learning (SSL) frameworks have been extensively applied to sensor-based Human Activity Recognition (HAR) in order to learn deep representations without data annotations. While SSL frameworks reach performance almost comparable to supervised models, studies on interpreting representations learnt by SSL models are limited. Nevertheless, modern explainability methods could help to unravel the differences between SSL and supervised representations: how they are being learnt, what properties of input data they preserve, and when SSL can be chosen over supervised training. In this paper, we aim to analyze deep representations of two recent SSL frameworks, namely SimCLR and VICReg. Specifically, the emphasis is made on (i) comparing the robustness of supervised and SSL models to corruptions in input data; (ii) explaining predictions of deep learning models using saliency maps and highlighting what input channels are mostly used for predicting various activities; (iii) exploring properties encoded in SSL and supervised representations using probing. Extensive experiments on two single-device datasets (MobiAct and UCI-HAR) have shown that self-supervised learning representations are significantly more robust to noise in unseen data compared to supervised models. In contrast, features learnt by the supervised approaches are more homogeneous across subjects and better encode the nature of activities.

Automated Human Activity Recognition has long been a problem of great interest in human-centered and ubiquitous computing. In the last years, a plethora of supervised learning algorithms based on deep neural networks has been suggested to address this problem using various modalities. While every modality has its own limitations, there is one common challenge. Namely, supervised learning requires vast amounts of annotated data which is practically hard to collect. In this paper, we benefit from the self-supervised learning paradigm (SSL) that is typically used to learn deep feature representations from unlabeled data. Moreover, we upgrade a contrastive SSL framework, namely SimCLR, widely used in various applications by introducing a temporal feature alignment procedure for Human Activity Recognition. Specifically, we propose integrating a dynamic time warping (DTW) algorithm in a latent space to force features to be aligned in a temporal dimension. Extensive experiments have been conducted for the unimodal scenario with inertial modality as well as in multimodal settings using inertial and skeleton data. According to the obtained results, the proposed approach has a great potential in learning robust feature representations compared to the recent SSL baselines, and clearly outperforms supervised models in semi-supervised learning. The code for this paper is available via the following link: https://github.com/bulatkh/csshar_tfa.

Human Activity Recognition is a field of research where input data can take many forms. Each of the possible input modalities describes human behaviour in a different way, and each has its own strengths and weaknesses. We explore the hypothesis that leveraging multiple modalities can lead to better recognition. Since manual annotation of input data is expensive and time-consuming, the emphasis is made on self-supervised methods which can learn useful feature representations without any ground truth labels. We extend a number of recent contrastive self-supervised approaches for the task of Human Activity Recognition, leveraging inertial and skeleton data. Furthermore, we propose a flexible, general-purpose framework for performing multimodal self-supervised learning, named Contrastive Multiview Coding with Cross-Modal Knowledge Mining (CMC-CMKM). This framework exploits modality-specific knowledge in order to mitigate the limitations of typical self-supervised frameworks. The extensive experiments on two widely-used datasets demonstrate that the suggested framework significantly outperforms contrastive unimodal and multimodal baselines on different scenarios, including fully-supervised fine-tuning, activity retrieval and semi-supervised learning. Furthermore, it shows performance competitive even compared to supervised methods.

Deep Learning models, applied to a sensor-based Human Activity Recognition task, usually require vast amounts of annotated time-series data to extract robust features. However, annotating signals coming from wearable sensors can be a tedious and, often, not so intuitive process, that requires specialized tools and predefined scenarios, making it an expensive and time-consuming task. This paper combines one of the most recent advances in Self-Supervised Leaning (SSL), namely a SimCLR framework, with a powerful transformer-based encoder to introduce a Contrastive Self-supervised learning approach to Sensor-based Human Activity Recognition (CSSHAR) that learns feature representations from unlabeled sensory data. Extensive experiments conducted on three widely used public datasets have shown that the proposed method outperforms recent SSL models. Moreover, CSSHAR is capable of extracting more robust features than the identical supervised transformer when transferring knowledge from one dataset to another as well as when very limited amounts of annotated data are available.

One of the most significant challenges in Human Activity Recognition using wearable devices is inter-class similarities and subject heterogeneity. These problems lead to the difficulties in constructing robust feature representations that might negatively affect the quality of recognition. This study, for the first time, applies deep triplet networks with various triplet loss functions and mining methods to the Human Activity Recognition task. Moreover, we introduce a novel method for constructing hard triplets by exploiting similarities between subjects performing the same activities using the concept of Hierarchical Triplet Loss. Our deep triplet models are based on the recent state-of-the-art LSTM networks with two attention mechanisms. The extensive experiments conducted in this paper identify important hyperparameters and settings for training deep metric learning models on widely-used open-source Human Activity Recognition datasets. The comparison of the proposed models against the recent benchmark models shows that deep metric learning approach has the potential to improve the quality of recognition. Specifically, at least one of the implemented triplet networks shows the state-of-the-art results for each dataset used in this study, namely PAMAP2, USC-HAD and MHEALTH. Another positive effect of applying deep triplet networks and especially the proposed sampling algorithm is that feature representations are less affected by inter-class similarities and subject heterogeneity issues.