Research

PRISM introduces a data-free, task-agnostic debiasing framework for VLMs. It first leverages an LLM to generate bias-aware scene descriptions from simple class prompts, then learns a linear projection of the CLIP embedding space via a novel Latent-space Debiasing loss that enforces intra-class invariance and inter-class separability.

We introduced the novel problem of unsupervised federated domain generalization and proposed FedGaLA, a method that improves model generalization across unseen domains by aligning gradients at both the client and server levels. This work is grounded in a theoretical framework that links domain shift to gradient alignment. FedGaLA achieves state-of-the-art performance on several domain generalization benchmarks.

We propose FedSB, a method for federated domain generalization that improves model robustness across diverse domains using label smoothing to reduce local overconfidence and a balanced training mechanism to mitigate data heterogeneity.

This work presents a theoretical framework for addressing long-tailed recognition (LTR) through continual learning (CL), where models are trained sequentially on data subsets to balance performance across head (frequent) and tail (rare) classes. By proving bounds on model weight updates and demonstrating CL's effectiveness on benchmark datasets, the authors show that CL can significantly improve LTR performance, offering a unified approach that aligns both theoretical insights and practical results in machine learning.

This study introduces the first continual learning approach for pedestrian detection that can effectively address distribution shift, a common issue in prior works. We proposed modified Elastic Weight Consolidation for object detection networks, enabling the model to maintain its performance across different datasets and significantly improve the miss rate on CrowdHuman and CityPersons datasets by mitigating catastrophic forgetting.

A multi-scale crowd counting and localization platform is proposed in this work. This novel architecture alongside the multi-scale multi-task loss function has demonstrated promising performance.

We proposed the first Multi-scale gradient capsule GAN and utilized it for face super-resolution. This model outperformed state-of-the-art face super-resolution models.

In this work, the largest publicly available dataset for COVID-19 is collected and a powerful COVID-19 detection model based on CheXNet is proposed. This model can detect COVID19 accurately using meaningful features

One of the first attempts for high-scale super-resolution (8x) in biomedical domain. MSG-CapsGAN shows promising results in the medical domain as well.

A fully automated tooth extraction model is implemented using a genetic algorithm. A multi-feature extraction model with a capsule classifier is developed for caries detection.

We propose an AI-powered FPWFS method specifically for low-order mode estimation in infrared wavelengths. Our approach is trained on simulated data and validated on on-telescope data collected from the Keck I adaptive optic (K1AO) bench calibration source in K-band. This study paves the way for more compact, efficient, and high-performing AO systems for astronomical observations.

An efficient and scalable transistor-level implementation of two different recurrent neural networks is proposed using a memristor crossbar array.

An LSTM network is utilized for user activity prediction in 5G networks. The proposed model can accurately predict anomalies up to one hour in advance.

The characteristics of the eroded area in the pipe are extracted from the pressure response using a GRU network. This work represents the first erosion detection system leveraging deep learning.