Moshiur Farazi

Moshiur Farazi, Sameera Ramasinghe, Mahbub Ahmed Turza et al.

Vision-Language Models (VLMs) struggle with compositional reasoning that requires understanding inter-object relationships. A natural remedy is to inject explicit scene graph triplets $\langle s, p, o \rangle$ from an off-the-shelf scene graph generator (SGG), but we show this backfires: discrete text labels collide with the continuous visual modality, degrading GQA accuracy from 60.38\% to 58.86\%. We propose \textbf{HyperVis}, which bypasses the SGG semantic bottleneck entirely. From $N$ class-agnostic region proposals, we compute a dense $O(N^2)$ visual relation tensor via spatially-biased cross-attention, project it onto a Lorentz hyperboloid, and enforce hierarchy through spatial physics, namely IoA-driven entailment cones and exterior-angle repulsion. We discover that HyperVis contributes in two complementary ways: (1) as a \emph{training-time regularizer}, the hyperbolic relational losses shape LoRA representations that improve generative VQA (GQA 61.03\% vs.\ 57.21\% for LoRA fine-tuning without relational losses, recovering and surpassing the baseline); and (2) as an \emph{inference-time relational encoder}, hyperbolic prefix tokens boost discriminative compositional scoring (SugarCrepe 79.94\%, $+$6.25pp over baseline). The learned curvature stabilises at $κ{=}4.0$, an order of magnitude above prior hyperbolic VLMs where $κ$ typically collapses toward zero, indicating that continuous visual features genuinely require the exponential volume of strongly curved space. A controlled Euclidean ablation confirms this decomposition: the relational pipeline regularises LoRA comparably in flat space (GQA 60.81\%), but the compositionality gain is specifically hyperbolic (SugarCrepe $+$4.58pp over Euclidean), with entailment loss ${\sim}6{\times}$ higher in Euclidean training. Codes are available at TBA.

Sameera Ramasinghe, Lachlan MacDonald, Moshiur Farazi et al.

Characterizing the remarkable generalization properties of over-parameterized neural networks remains an open problem. In this paper, we promote a shift of focus towards initialization rather than neural architecture or (stochastic) gradient descent to explain this implicit regularization. Through a Fourier lens, we derive a general result for the spectral bias of neural networks and show that the generalization of neural networks is heavily tied to their initialization. Further, we empirically solidify the developed theoretical insights using practical, deep networks. Finally, we make a case against the controversial flat-minima conjecture and show that Fourier analysis grants a more reliable framework for understanding the generalization of neural networks.

Yu Cui, Moshiur Farazi

Visual Relationship Detection (VRD) impels a computer vision model to 'see' beyond an individual object instance and 'understand' how different objects in a scene are related. The traditional way of VRD is first to detect objects in an image and then separately predict the relationship between the detected object instances. Such a disjoint approach is prone to predict redundant relationship tags (i.e., predicate) between the same object pair with similar semantic meaning, or incorrect ones that have a similar meaning to the ground truth but are semantically incorrect. To remedy this, we propose to jointly train a VRD model with visual object features and semantic relationship features. To this end, we propose VReBERT, a BERT-like transformer model for Visual Relationship Detection with a multi-stage training strategy to jointly process visual and semantic features. We show that our simple BERT-like model is able to outperform the state-of-the-art VRD models in predicate prediction. Furthermore, we show that by using the pre-trained VReBERT model, our model pushes the state-of-the-art zero-shot predicate prediction by a significant margin (+8.49 R@50 and +8.99 R@100).

Samiul Basir Bhuiyan, Md. Sazzad Hossain Adib, Mohammed Aman Bhuiyan et al.

Large language models (LLMs) have rapidly advanced in recent years, achieving remarkable performance across a wide range of natural language processing tasks. However, this progress has come at the cost of increasingly large model sizes, which pose significant challenges for deployment, scalability, and energy efficiency. To address these limitations, post-training pruning has emerged as a promising approach for reducing model size and inference latency without the need for retraining. Despite these advantages, many existing pruning methods result in substantial performance degradation or require computationally expensive fine-tuning. In this work, we introduce Z-Pruner, a novel post-training pruning method designed to induce sparsity in pretrained LLMs without any retraining. Unlike conventional approaches, Z-Pruner leverages both weight update magnitudes and activation patterns to identify and eliminate redundant parameters more effectively. Our method is model-agnostic, efficient, and easy to implement. We evaluate Z-Pruner using multiple widely-used LLM architectures, including LLaMA-2, LLaMA-3, and OPT, across a diverse set of standard language benchmarks. Experimental results demonstrate that Z-Pruner surpasses state-of-the-art pruning methods that require intensive weight updates. Specifically, Z-Pruner achieves the lowest perplexity scores and the highest overall average score for zero-shot accuracy. We have made the corresponding codes publicly available at https://github.com/sazzadadib/Z-Pruner.

Rafin Hassan, Zarin Tasnim Roshni, Rafiqul Bari et al.

Hyperspectral imaging (HSI) classification is a critical tool with widespread applications across diverse fields such as agriculture, environmental monitoring, medicine, and materials science. Due to the limited availability of high-quality training samples and the high dimensionality of spectral data, HSI classification models are prone to overfitting and often face challenges in balancing accuracy and computational complexity. Furthermore, most of HSI classification models are monomodal, where it solely relies on spectral-spatial data to learn decision boundaries in the high dimensional embedding space. To address this, we propose a general-purpose Semantic Spectral-Spatial Fusion Network (S3FN) that uses contextual, class specific textual descriptions to complement the training of an HSI classification model. Specifically, S3FN leverages LLMs to generate comprehensive textual descriptions for each class label that captures their unique characteristics and spectral behaviors. These descriptions are then embedded into a vector space using a pre-trained text encoder such as BERT or RoBERTa to extract meaningful label semantics which in turn leads to a better feature-label alignment for improved classification performance. To demonstrate the effectiveness of our approach, we evaluate our model on three diverse HSI benchmark datasets - Hyperspectral Wood, HyperspectralBlueberries, and DeepHS-Fruit and report significant performance boost. Our results highlight the synergy between textual semantics and spectral-spatial data, paving the way for further advancements in semantically augmented HSI classification models. Codes are be available in: https://github.com/milab-nsu/S3FN

Sameera Ramasinghe, Moshiur Farazi, Salman Khan et al.

Conditional GANs (cGAN), in their rudimentary form, suffer from critical drawbacks such as the lack of diversity in generated outputs and distortion between the latent and output manifolds. Although efforts have been made to improve results, they can suffer from unpleasant side-effects such as the topology mismatch between latent and output spaces. In contrast, we tackle this problem from a geometrical perspective and propose a novel training mechanism that increases both the diversity and the visual quality of a vanilla cGAN, by systematically encouraging a bi-lipschitz mapping between the latent and the output manifolds. We validate the efficacy of our solution on a baseline cGAN (i.e., Pix2Pix) which lacks diversity, and show that by only modifying its training mechanism (i.e., with our proposed Pix2Pix-Geo), one can achieve more diverse and realistic outputs on a broad set of image-to-image translation tasks. Codes are available at https://github.com/samgregoost/Rethinking-CGANs.

Hanif Rasyidi, Moshiur Farazi

This paper investigates various factors that influence the performance of end-to-end deep learning approaches for historical writer identification (HWI), a task that remains challenging due to the diversity of handwriting styles, document degradation, and the limited number of labelled samples per writer. These conditions often make accurate recognition difficult, even for human experts. Traditional HWI methods typically rely on handcrafted image processing and clustering techniques, which tend to perform well on small and carefully curated datasets. In contrast, end-to-end pipelines aim to automate the process by learning features directly from document images. However, our experiments show that many of these models struggle to generalise in more realistic, document-level settings, especially under zero-shot scenarios where writers in the test set are not present in the training data. We explore different combinations of pre-processing methods, backbone architectures, and post-processing strategies, including text segmentation, patch sampling, and feature aggregation. The results suggest that most configurations perform poorly due to weak capture of low-level visual features, inconsistent patch representations, and high sensitivity to content noise. Still, we identify one end-to-end setup that achieves results comparable to the top-performing system, despite using a simpler design. These findings point to key challenges in building robust end-to-end systems and offer insight into design choices that improve performance in historical document writer identification.

Sadia Abdulhalim, Muaz Albaghdadi, Moshiur Farazi

Traditional sentiment analysis has long been a unimodal task, relying solely on text. This approach overlooks non-verbal cues such as vocal tone and prosody that are essential for capturing true emotional intent. We introduce Dynamic Attention Fusion (DAF), a lightweight framework that combines frozen text embeddings from a pretrained language model with acoustic features from a speech encoder, using an adaptive attention mechanism to weight each modality per utterance. Without any finetuning of the underlying encoders, our proposed DAF model consistently outperforms both static fusion and unimodal baselines on a large multimodal benchmark. We report notable gains in F1-score and reductions in prediction error and perform a variety of ablation studies that support our hypothesis that the dynamic weighting strategy is crucial for modeling emotionally complex inputs. By effectively integrating verbal and non-verbal information, our approach offers a more robust foundation for sentiment prediction and carries broader impact for affective computing applications -- from emotion recognition and mental health assessment to more natural human computer interaction.

Najla Zuhir, Amna Mohammad Salim, Parvathy Premkumar et al.

We present an advanced approach to mobile app review analysis aimed at addressing limitations inherent in traditional star-rating systems. Star ratings, although intuitive and popular among users, often fail to capture the nuanced feedback present in detailed review texts. Traditional NLP techniques -- such as lexicon-based methods and classical machine learning classifiers -- struggle to interpret contextual nuances, domain-specific terminology, and subtle linguistic features like sarcasm. To overcome these limitations, we propose a modular framework leveraging large language models (LLMs) enhanced by structured prompting techniques. Our method quantifies discrepancies between numerical ratings and textual sentiment, extracts detailed, feature-level insights, and supports interactive exploration of reviews through retrieval-augmented conversational question answering (RAG-QA). Comprehensive experiments conducted on three diverse datasets (AWARE, Google Play, and Spotify) demonstrate that our LLM-driven approach significantly surpasses baseline methods, yielding improved accuracy, robustness, and actionable insights in challenging and context-rich review scenarios.

Ibrahim Altan, Abdulla Bachir, Yousuf Parbhulkar et al.

Phishing emails pose a persistent and increasingly sophisticated threat, undermining email security through deceptive tactics designed to exploit both semantic and structural vulnerabilities. Traditional detection methods, often based on isolated analysis of email content or embedded URLs, fail to comprehensively address these evolving attacks. In this paper, we propose a dual-path phishing detection framework that integrates transformer-based natural language processing (NLP) with classical machine learning to jointly analyze email text and embedded URLs. Our approach leverages the complementary strengths of semantic analysis using fine-tuned transformer architectures (e.g., DistilBERT) and structural link analysis via character-level TF-IDF vectorization paired with classical classifiers (e.g., Random Forest). Empirical evaluation on representative email and URL datasets demonstrates that this combined approach significantly improves detection accuracy. Specifically, the DistilBERT model achieves a near-optimal balance between accuracy and computational efficiency for textual phishing detection, while Random Forest notably outperforms other classical classifiers in identifying malicious URLs. The modular design allows flexibility for standalone deployment or ensemble integration, facilitating real-world adoption. Collectively, our results highlight the efficacy and practical value of this dual-path approach, establishing a scalable, accurate, and interpretable solution capable of enhancing email security against contemporary phishing threats.

Ce Wang, Moshiur Farazi, Nick Barnes

General purpose semantic segmentation relies on a backbone CNN network to extract discriminative features that help classify each image pixel into a 'seen' object class (ie., the object classes available during training) or a background class. Zero-shot semantic segmentation is a challenging task that requires a computer vision model to identify image pixels belonging to an object class which it has never seen before. Equipping a general purpose semantic segmentation model to separate image pixels of 'unseen' classes from the background remains an open challenge. Some recent models have approached this problem by fine-tuning the final pixel classification layer of a semantic segmentation model for a Zero-Shot setting, but struggle to learn discriminative features due to the lack of supervision. We propose a recursive training scheme to supervise the retraining of a semantic segmentation model for a zero-shot setting using a pseudo-feature representation. To this end, we propose a Zero-Shot Maximum Mean Discrepancy (ZS-MMD) loss that weighs high confidence outputs of the pixel classification layer as a pseudo-feature representation, and feeds it back to the generator. By closing-the-loop on the generator end, we provide supervision during retraining that in turn helps the model learn a more discriminative feature representation for 'unseen' classes. We show that using our recursive training and ZS-MMD loss, our proposed model achieves state-of-the-art performance on the Pascal-VOC 2012 dataset and Pascal-Context dataset.

Zahidul Islam, Mohammad Rukonuzzaman, Raiyan Ahmed et al.

Automatically detecting violence from surveillance footage is a subset of activity recognition that deserves special attention because of its wide applicability in unmanned security monitoring systems, internet video filtration, etc. In this work, we propose an efficient two-stream deep learning architecture leveraging Separable Convolutional LSTM (SepConvLSTM) and pre-trained MobileNet where one stream takes in background suppressed frames as inputs and other stream processes difference of adjacent frames. We employed simple and fast input pre-processing techniques that highlight the moving objects in the frames by suppressing non-moving backgrounds and capture the motion in-between frames. As violent actions are mostly characterized by body movements these inputs help produce discriminative features. SepConvLSTM is constructed by replacing convolution operation at each gate of ConvLSTM with a depthwise separable convolution that enables producing robust long-range Spatio-temporal features while using substantially fewer parameters. We experimented with three fusion methods to combine the output feature maps of the two streams. Evaluation of the proposed methods was done on three standard public datasets. Our model outperforms the accuracy on the larger and more challenging RWF-2000 dataset by more than a 2% margin while matching state-of-the-art results on the smaller datasets. Our experiments lead us to conclude, the proposed models are superior in terms of both computational efficiency and detection accuracy.

Shafkat Farabi, Hasibul Himel, Fakhruddin Gazzali et al.

Action quality assessment (AQA) aims at automatically judging human action based on a video of the said action and assigning a performance score to it. The majority of works in the existing literature on AQA divide RGB videos into short clips, transform these clips to higher-level representations using Convolutional 3D (C3D) networks, and aggregate them through averaging. These higher-level representations are used to perform AQA. We find that the current clip level feature aggregation technique of averaging is insufficient to capture the relative importance of clip level features. In this work, we propose a learning-based weighted-averaging technique. Using this technique, better performance can be obtained without sacrificing too much computational resources. We call this technique Weight-Decider(WD). We also experiment with ResNets for learning better representations for action quality assessment. We assess the effects of the depth and input clip size of the convolutional neural network on the quality of action score predictions. We achieve a new state-of-the-art Spearman's rank correlation of 0.9315 (an increase of 0.45%) on the MTL-AQA dataset using a 34 layer (2+1)D ResNet with the capability of processing 32 frame clips, with WD aggregation.

Moshiur Farazi, Salman Khan, Nick Barnes

Humans explain inter-object relationships with semantic labels that demonstrate a high-level understanding required to perform complex Vision-Language tasks such as Visual Question Answering (VQA). However, existing VQA models represent relationships as a combination of object-level visual features which constrain a model to express interactions between objects in a single domain, while the model is trying to solve a multi-modal task. In this paper, we propose a general purpose semantic relationship parser which generates a semantic feature vector for each subject-predicate-object triplet in an image, and a Mutual and Self Attention (MSA) mechanism that learns to identify relationship triplets that are important to answer the given question. To motivate the significance of semantic relationships, we show an oracle setting with ground-truth relationship triplets, where our model achieves a ~25% accuracy gain over the closest state-of-the-art model on the challenging GQA dataset. Further, with our semantic parser, we show that our model outperforms other comparable approaches on VQA and GQA datasets.