The Kirat community is an indigenous group of eastern Nepal with a rich linguistic heritage, including Kirat Rai languages and their writing systems. Many of these languages are increasingly under-documented and are experiencing rapid decline in intergenerational transmission, making them vulnerable to long-term loss. This highlights the need for computational approaches to support their preservation and accessibility. In this work, we focus on the digital preservation of Kirat Rai Lipi through AI-based character recognition. We introduce the first curated and annotated dataset of Kirat Rai Lipi characters for low-resource script modeling. To establish a baseline, we fine-tune and evaluate multiple pretrained deep learning vision models, including efficient transformer-based architectures optimized for low-latency inference on edge devices. Experimental results demonstrate strong performance in character classification, achieving competitive accuracy and F1-scores under limited data conditions. To the best of our knowledge, this is the first study to develop a dedicated Kirat Rai Lipi dataset and evaluate pretrained models for its recognition, enabling future real-time digitization and multimodal language processing systems. Our code is available at: GitHub
Short ECG recordings are common in research datasets, but many clinical and physiological phenomena, such as autonomic modulation, circadian HRV changes, and transient arrhythmias, require long continuous recordings. Existing generative models either synthesize short beats/segments or fail to preserve physiologically meaningful beat-to- beat variability over long durations.We propose an architecture PhysioDyn with (a) a physiology-constrained GAN that separates beat morphology from inter-beat interval (IBI) dynamics and (b) IDM models inter-beat intervals (RR series) as a continuous-time latent process (GRU-Neural ODEs) and (c) Composer compose beat morphologies spaced by IBIs and (d) a multi-scale discriminator suite. This enables the synthesis of longer-range ECG that preserve clinically and statistically significant HRV properties.
This manuscript is available upon request. Contact Author
While ECG signals are harder to forge than fingerprints or facial features, privacy, spoofing, and identification in noisy environments remain concerns. We present an attention-based (self and cross attention) multimodal that leverages the MobileNetV4 and BiGRU as encoders. We simulate noisy or wearable acquisition conditions by segmenting ECG signals into random non-overlapping windows of (1-4) seconds. Multiple noise sources, including Gaussian noise, muscle artifacts, powerline interference, base- line wander, electrode noise, and motion artifacts, were randomly injected at varying signal-to-noise ratio (SNR) levels ranging from -6 dB to 24 dB for each subject and signal segment. Using ECG-ID with a synthetic(GAN) dataset under adversarial training, our method achieves an accuracy of 96.93% with an EER of 3.70%.
This manuscript is available upon request. Contact Author
Hybrid quantum-classical neural networks have emerged as a promising approach for leveraging quantum computing in machine learning while overcoming current hardware limitations. This paper presents a Deep Quantum Neural Network (DQNN) for handwritten digit classification on the MNIST dataset using a hybrid architecture that combines classical feature extraction with a Variational Quantum Circuit (VQC). The proposed model first compresses high-dimensional image features through a classical neural network before encoding them into an 8-qubit quantum circuit based on angle embedding and strongly entangling layers. The measured quantum expectation values are then used for final classification. To provide a comprehensive evaluation, the study will investigate model reproducibility through multi-seed experiments, analyze performance under various datasets (MNIST and low-resource indigenous scripts of Nepal) with varying training dataset sizes, and include a lightweight interpretability analysis to examine latent feature representations and quantum measurement outputs. The proposed framework aims to offer an efficient and reproducible benchmark for hybrid quantum machine learning while improving the transparency and practical understanding of quantum-enhanced classification models.
This manuscript is available upon request. Contact Author