Published in GlobalSouthML Workshop at International Conference on Machine Learning (ICML) 2026, Seoul, South Korea, 2026
We propose two efficient flow-based metrics, FLD and D-FLD, that provide accurate realness assessment of generated images with significantly fewer parameters and samples than FID, making them ideal for evaluating small image sets in diverse domains.
Published in GlobalSouthML Workshop at International Conference on Machine Learning (ICML) 2026, Seoul, South Korea, 2026
WaveMix is a new resource-efficient and scalable neural architecture for computer vision that achieves comparable or better accuracy than the state-of-the-art in multiple-computer vision tasks.
Published in GlobalSouthML Workshop at International Conference on Machine Learning (ICML) 2026, Seoul, South Korea, 2026
CXV is a new convolutional transformer hybrid architecture that outperforms other architectures in image classification, especially in scenarios with limited data and GPU resources.
Published in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops 2025, Honolulu, Hawaii, US, 2025
WavePaint is a new resource-efficient neural architecture for image inpainting that achieves comparable or better performance than the current state-of-the-art models without using adversarial training or diffusion.
Published in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops 2025, Honolulu, Hawaii, US, 2025
FLD+ is a new, more efficient and reliable image quality metric using normalizing flows that outperforms FID, especially with fewer images and on specialized domains.**
Published in Proceedings of the 2025 IEEE International Symposium on Biomedical Imaging (ISBI), Huston TX, USA, 2025
Our study introduces ResNet-L2 (RL2), a novel metric using ResNet features and a normalizing flow for evaluating generative models in histopathology, offering reliable assessments with fewer images and quicker assessments than traditional metrics, effectively handling diverse degradation types and diffusion processes.
Published in Transactions on Machine Learning Research (TMLR), 2025
Our study evaluates lightweight, pre-trained CNN backbones across diverse datasets, offering crucial insights for model selection in small dataset scenarios.
Published in Proceedings of the 2025 AAAI Conference on Artificial Intelligence (AAAI-25), Philadelphia, PA, USA, 2025
WaveMixSR-V2, an enhanced version of WaveMixSR, incorporating pixel shuffle and a multistage design, achieving state-of-the-art super-resolution performance on the BSD100 dataset with improved resource efficiency, lower latency, and higher throughput.
Published in Proceedings of the 2024 IEEE International Conference on Big Data (IEEE BigData), Washington, DC, USA, 2024
FLeNS combines Nesterov acceleration with adaptive Hessian sketching to achieve super-linear convergence in federated learning while significantly reducing communication overhead.
Published in Proceedings of the 27th International Conference on Pattern Recognition, Kolkata, India, 2024
The new technique, ATT, enhances unsupervised domain adaptation by introducing a novel transport loss that displaces classifier outputs to reduce class confusion and improve domain-invariant representations, leading to superior UDA results on benchmark datasets.
Published in Proceedings of the 35th British Machine Vision Conference (BMVC) 2024, Glasgow, UK, 2024
My work as an intern at Samsung Research where I developed a multi-label classification system for detecting facial action units in dogs and cats using a semi-supervised pseudo-labeling strategy using small loss trick and consistency regularization
Published in arxiv, 2024
EDSNet enhances the Direct-to-Summarize Network (DSNet) with resource-efficient token mixing mechanisms and alternative pooling strategies, significantly reducing computational costs while maintaining competitive performance in video summarization.
Published in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Waikaloa, HI, USA, 2024
WaveMixSR is a new neural network for image super-resolution that uses the WaveMix architecture, which is based on a 2D-discrete wavelet transform for spatial token-mixing, and achieves higher performance while requiring fewer resources and training data than transformer-based models.
Published in Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies, Rome, Italy, 2024
This study evaluates the robustness of different deep learning architectures for breast cancer histopathological image classification across magnification scales.
Published in Workshop on GRaphs in biomedicAl Image anaLysis, International Conference on Medical Image Computing and Computer Assisted Intervention, 2023
The paper introduces a novel approach using a heterogeneous graph neural network to capture spatial and hierarchical relations in histopathological images, achieving higher accuracy than transformer-based models on breast cancer datasets.
Published in Tiny Papers @ International Conference on Learning Representations (ICLR), 2023
The paper proposes a computationally-efficient WaveMix-based fully convolutional architecture for image inpainting that outperforms the current state-of-the-art models while using less parameters and lower training and evaluation times.
Published in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Waikoloa, HI, USA, 2022
We developed a resource-efficient architecture for vision by making three modifications on ViT to address their shortcomings, which significantly improved their performance and made them more accessible.
Published in Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, Punta Cana, Dominican Republic, 2021
A new dataset and model for computational humor are proposed, which can automatically rate the humor quotient of stand-up comedy clips.