Selected work

Representative Papers

Bridging the Gap in Bangla Healthcare: Machine Learning Based Disease Prediction Using a Symptoms-Disease Dataset

Feb 13, 2025European Conference on Cognitive Ergonomics

This study addresses the longstanding scarcity of localized disease prediction resources for Bengali-speaking populations, which has hindered equitable access to reliable health information. To bridge this gap, the authors present the first large-scale Bengali dataset comprising 85 diseases and 758 symptom–disease associations, which is publicly released. Leveraging this dataset, they design an ensemble learning model that integrates multiple machine learning algorithms with both soft and hard voting strategies to enable high-accuracy disease classification from Bengali symptom descriptions. Experimental results demonstrate that the proposed model achieves 98% accuracy on the newly curated dataset, significantly advancing the accessibility and equity of localized healthcare information for Bengali speakers.

1 citationsRead paper

Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields

Aug 08, 2026

This work addresses the challenges in tropical cyclone forecasting, where existing methods struggle to efficiently generate satellite imagery and multivariate atmospheric fields jointly, often suffering from high computational costs, multi-step sampling procedures, and physically inconsistent track predictions. To overcome these limitations, the authors propose a single-pass generative model that integrates a five-channel variational autoencoder for input compression, a conditional Rectified Flow UNet to model spatiotemporal dynamics, a factorized temporal attention module, and a DRaFT reward-based fine-tuning mechanism leveraging differentiable trajectory error. The approach achieves, for the first time, one-step joint generation of infrared satellite images and four atmospheric variables, attaining a PSNR of 16.35 dB and SSIM of 0.759 at +9-hour lead time, accelerating sampling by approximately 30×, reducing track error to 62.4 km (a 15% improvement over the baseline), with further 8–11% error reduction through reward fine-tuning.

0 citationsRead paper

Linear Multi-Timescale Retention as a Memory-Efficient Vision-Language Bridge

Aug 02, 2026

This work addresses the O(N²) memory bottleneck of Softmax-based attention in high-resolution vision-language models, which hinders global scene understanding and object constancy. Existing linear attention methods fail to preserve spatial sequence routing, leading to degraded performance. To overcome this, we propose the Linear Multi-timescale Memory Retention (LIA-MTR) module, which leverages ELU-based positive feature mappings, adaptive write-in gating, and a log-linear recurrent decay mechanism to enable lossless long-sequence routing and cross-modal fusion under strict O(N) complexity. LIA-MTR supports unbounded context extension and substantially mitigates the “lost-in-the-middle” phenomenon. Experiments demonstrate that LIA-MTR processes 262,144 visual tokens using only 11.2 GB of GPU memory—far exceeding the 16,384-token limit of standard multi-head attention—and achieves a 71.00% MME score, outperforming an MLP baseline (68.11%) with a 10-percentage-point gain in object constancy.

0 citationsRead paper

MixFrag: Fragility-Guided Mixed-Precision Post-Training Quantization for Vision Transformers

Jul 30, 2026

Existing post-training quantization methods for Vision Transformers employ uniform bit-widths across all components, disregarding their heterogeneous sensitivity to quantization and thereby incurring substantial accuracy degradation. This work proposes MixFrag, a novel framework that introduces, for the first time, a KL divergence–based metric to quantify the quantization fragility of each layer. By leveraging a small calibration set to assess output distribution shifts, MixFrag formulates mixed-precision bit allocation as a multiple-choice knapsack problem, enabling adaptive optimization under a given bit budget. The method achieves state-of-the-art performance on ImageNet-1K classification as well as COCO detection and segmentation tasks, yielding improvements of up to 9.6 AP over prior best approaches under MP3/MP3 settings.

0 citationsRead paper

Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge-Efficient Bengali ASR

Jul 10, 2026

This work addresses the challenges of word fragmentation and autoregressive collapse in lightweight autoregressive speech recognition models when applied to morphologically rich, non-Latin scripts such as Bengali, primarily caused by English-centric byte-level tokenizers. To overcome this without requiring full model re-pretraining, the authors propose a cross-script lexical transplantation method: replacing the decoder’s byte-level vocabulary with a Bengali-specific WordPiece vocabulary derived from BanglaBERT and rescaling the embedding matrix to better align with the target language’s linguistic structure. This approach enables the first efficient and reproducible tokenizer replacement for compact ASR architectures like Moonshine. Evaluated on the Lipi-Ghor dataset, the method achieves a word error rate of 21.54%, a real-time factor of 0.0053, an 85.8% reduction in autoregressive sequence length, and a substantial drop in word fertility from 9.16 to 1.30.

0 citationsRead paper

Recent publications

Latest Papers

Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields

Aug 08, 2026

This work addresses the challenges in tropical cyclone forecasting, where existing methods struggle to efficiently generate satellite imagery and multivariate atmospheric fields jointly, often suffering from high computational costs, multi-step sampling procedures, and physically inconsistent track predictions. To overcome these limitations, the authors propose a single-pass generative model that integrates a five-channel variational autoencoder for input compression, a conditional Rectified Flow UNet to model spatiotemporal dynamics, a factorized temporal attention module, and a DRaFT reward-based fine-tuning mechanism leveraging differentiable trajectory error. The approach achieves, for the first time, one-step joint generation of infrared satellite images and four atmospheric variables, attaining a PSNR of 16.35 dB and SSIM of 0.759 at +9-hour lead time, accelerating sampling by approximately 30×, reducing track error to 62.4 km (a 15% improvement over the baseline), with further 8–11% error reduction through reward fine-tuning.

0 citationsRead paper

Linear Multi-Timescale Retention as a Memory-Efficient Vision-Language Bridge

Aug 02, 2026

This work addresses the O(N²) memory bottleneck of Softmax-based attention in high-resolution vision-language models, which hinders global scene understanding and object constancy. Existing linear attention methods fail to preserve spatial sequence routing, leading to degraded performance. To overcome this, we propose the Linear Multi-timescale Memory Retention (LIA-MTR) module, which leverages ELU-based positive feature mappings, adaptive write-in gating, and a log-linear recurrent decay mechanism to enable lossless long-sequence routing and cross-modal fusion under strict O(N) complexity. LIA-MTR supports unbounded context extension and substantially mitigates the “lost-in-the-middle” phenomenon. Experiments demonstrate that LIA-MTR processes 262,144 visual tokens using only 11.2 GB of GPU memory—far exceeding the 16,384-token limit of standard multi-head attention—and achieves a 71.00% MME score, outperforming an MLP baseline (68.11%) with a 10-percentage-point gain in object constancy.

0 citationsRead paper

MixFrag: Fragility-Guided Mixed-Precision Post-Training Quantization for Vision Transformers

Jul 30, 2026

Existing post-training quantization methods for Vision Transformers employ uniform bit-widths across all components, disregarding their heterogeneous sensitivity to quantization and thereby incurring substantial accuracy degradation. This work proposes MixFrag, a novel framework that introduces, for the first time, a KL divergence–based metric to quantify the quantization fragility of each layer. By leveraging a small calibration set to assess output distribution shifts, MixFrag formulates mixed-precision bit allocation as a multiple-choice knapsack problem, enabling adaptive optimization under a given bit budget. The method achieves state-of-the-art performance on ImageNet-1K classification as well as COCO detection and segmentation tasks, yielding improvements of up to 9.6 AP over prior best approaches under MP3/MP3 settings.

0 citationsRead paper

Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge-Efficient Bengali ASR

Jul 10, 2026

This work addresses the challenges of word fragmentation and autoregressive collapse in lightweight autoregressive speech recognition models when applied to morphologically rich, non-Latin scripts such as Bengali, primarily caused by English-centric byte-level tokenizers. To overcome this without requiring full model re-pretraining, the authors propose a cross-script lexical transplantation method: replacing the decoder’s byte-level vocabulary with a Bengali-specific WordPiece vocabulary derived from BanglaBERT and rescaling the embedding matrix to better align with the target language’s linguistic structure. This approach enables the first efficient and reproducible tokenizer replacement for compact ASR architectures like Moonshine. Evaluated on the Lipi-Ghor dataset, the method achieves a word error rate of 21.54%, a real-time factor of 0.0053, an 85.8% reduction in autoregressive sequence length, and a substantial drop in word fertility from 9.16 to 1.30.

0 citationsRead paper

Closing the Null Space: Guidance-Aware Quantization for Classifier-Free Diffusion

Jul 09, 2026

Existing post-training quantization methods overlook the coupled structure between the conditional and unconditional branches in Classifier-Free Guidance (CFG), leading to unconditional branch drift and degraded generation quality. This work is the first to identify and formally characterize this “branch drift trap” and proposes Guidance-Aware Mixed Precision (GAMP), a novel quantization strategy that calibrates models using the guidance-predicted output as the target. By evaluating per-layer activation sensitivity through guidance output degradation, GAMP employs a greedy knapsack-based approach to allocate mixed precision bit-widths. This mechanism inherently prevents branch drift, significantly enhancing generation fidelity while maintaining low-bit deployment efficiency. Moreover, it circumvents spurious optima arising from conventional calibration metrics, thereby aligning theoretical compression gains with actual inference performance.

0 citationsRead paper