Selected work

Representative Papers

Demodulation of chaotic signals using convolutional neural network

Jul 18, 2026

This study addresses the challenge of efficiently demodulating chaotic bifurcation parameter shift keying signals under low signal-to-noise ratio (SNR) conditions. To this end, the authors propose an end-to-end demodulation method based on convolutional neural networks that directly recovers binary information from chaotic time series generated by the logistic map, without relying on conventional synchronization mechanisms. This work represents the first application of deep learning to the demodulation of such chaotic signals and demonstrates a notable ability to recognize chaotic patterns not encountered during training. Evaluated in additive white Gaussian noise, the method achieves a bit error rate of 0.0819 even at an SNR as low as −13 dB (corresponding to a normalized SNR of +20 dB) with a 1.34% deviation in the bifurcation parameter, thereby confirming its robustness and generalization capability.

0 citationsRead paper

Just on Time: Token-Level Early Stopping for Diffusion Language Models

Feb 11, 2026

Diffusion language models suffer from inefficient inference, as most tokens converge well before the final denoising step. This work proposes a training-free, token-wise early stopping mechanism that dynamically determines whether each token position has stabilized and can be frozen ahead of schedule, based on lightweight signals such as model prediction confidence and local contextual consistency. For the first time, this approach enables adaptive, per-token early termination without any fine-tuning. Evaluated across diverse benchmarks—including mathematical reasoning, general question answering, and scientific comprehension—it substantially reduces the number of diffusion steps required while preserving generation quality, achieving state-of-the-art inference efficiency.

0 citationsRead paper

Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusion

Jan 06, 2026arXiv.org

This work addresses the problem of concept erasure in text-to-image diffusion models by proposing an efficient reinforcement learning–based removal method. By formulating the denoising process as a sequential decision-making problem, the approach introduces a timestep-aware critic and a CLIP-based noise-conditioned reward signal to enable policy gradient updates of the reverse diffusion kernel. This design significantly improves credit assignment accuracy and training stability. Empirical results demonstrate that the method achieves forgetting performance on multiple concepts that is either superior or comparable to strong existing baselines, while effectively preserving the model’s overall generation quality and prompt fidelity.

0 citationsRead paper

Application of deep learning approaches for medieval historical documents transcription

Dec 21, 2025

To address the significant degradation in OCR performance on medieval Latin manuscripts (9th–11th centuries), this paper introduces the first end-to-end deep learning framework explicitly designed for paleographic characteristics. Methodologically, it systematically models glyphic variation, ink fading, and unstructured layout—enabling morphology-aware data augmentation and a novel word-level image embedding recognition paradigm. The architecture integrates CNNs and RNNs, augmented by self-supervised pretraining and word-image matching. Evaluated on authentic manuscript datasets, the framework achieves an F1 score of 0.89 and reduces average string edit distance by 42% over general-purpose OCR baselines. To foster reproducibility and scalability, the source code and dataset are publicly released—establishing a robust, extensible foundation for historical document digitization.

0 citationsRead paper

Machine learning based animal emotion classification using audio signals

Mar 23, 2025

This work addresses the challenge of acoustic modeling under small-sample conditions in animal emotion recognition, proposing the first end-to-end audio-based emotion classification framework tailored to individual dogs. To overcome data scarcity, the method fuses Mel-spectrogram representations with time-domain statistical acoustic features and establishes a comparative multi-model framework comprising SVM, random forests, and shallow neural networks. Evaluated on a single-dog audio dataset, it achieves over 70% overall emotion classification accuracy. The study breaks through a key technical bottleneck in quantifying animal emotions from limited acoustic data, empirically validating the feasibility of fine-grained canine emotion discrimination using audio signals alone. It introduces a novel paradigm for interpretable, individualized animal affective computing—advancing both the understanding of animal intent in human–machine interaction and the development of robust, dog-specific emotional inference tools.

0 citationsRead paper

Recent publications

Latest Papers

Demodulation of chaotic signals using convolutional neural network

Jul 18, 2026

This study addresses the challenge of efficiently demodulating chaotic bifurcation parameter shift keying signals under low signal-to-noise ratio (SNR) conditions. To this end, the authors propose an end-to-end demodulation method based on convolutional neural networks that directly recovers binary information from chaotic time series generated by the logistic map, without relying on conventional synchronization mechanisms. This work represents the first application of deep learning to the demodulation of such chaotic signals and demonstrates a notable ability to recognize chaotic patterns not encountered during training. Evaluated in additive white Gaussian noise, the method achieves a bit error rate of 0.0819 even at an SNR as low as −13 dB (corresponding to a normalized SNR of +20 dB) with a 1.34% deviation in the bifurcation parameter, thereby confirming its robustness and generalization capability.

0 citationsRead paper

Just on Time: Token-Level Early Stopping for Diffusion Language Models

Feb 11, 2026

Diffusion language models suffer from inefficient inference, as most tokens converge well before the final denoising step. This work proposes a training-free, token-wise early stopping mechanism that dynamically determines whether each token position has stabilized and can be frozen ahead of schedule, based on lightweight signals such as model prediction confidence and local contextual consistency. For the first time, this approach enables adaptive, per-token early termination without any fine-tuning. Evaluated across diverse benchmarks—including mathematical reasoning, general question answering, and scientific comprehension—it substantially reduces the number of diffusion steps required while preserving generation quality, achieving state-of-the-art inference efficiency.

0 citationsRead paper

Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusion

Jan 06, 2026arXiv.org

This work addresses the problem of concept erasure in text-to-image diffusion models by proposing an efficient reinforcement learning–based removal method. By formulating the denoising process as a sequential decision-making problem, the approach introduces a timestep-aware critic and a CLIP-based noise-conditioned reward signal to enable policy gradient updates of the reverse diffusion kernel. This design significantly improves credit assignment accuracy and training stability. Empirical results demonstrate that the method achieves forgetting performance on multiple concepts that is either superior or comparable to strong existing baselines, while effectively preserving the model’s overall generation quality and prompt fidelity.

0 citationsRead paper

Application of deep learning approaches for medieval historical documents transcription

Dec 21, 2025

To address the significant degradation in OCR performance on medieval Latin manuscripts (9th–11th centuries), this paper introduces the first end-to-end deep learning framework explicitly designed for paleographic characteristics. Methodologically, it systematically models glyphic variation, ink fading, and unstructured layout—enabling morphology-aware data augmentation and a novel word-level image embedding recognition paradigm. The architecture integrates CNNs and RNNs, augmented by self-supervised pretraining and word-image matching. Evaluated on authentic manuscript datasets, the framework achieves an F1 score of 0.89 and reduces average string edit distance by 42% over general-purpose OCR baselines. To foster reproducibility and scalability, the source code and dataset are publicly released—establishing a robust, extensible foundation for historical document digitization.

0 citationsRead paper

Machine learning based animal emotion classification using audio signals

Mar 23, 2025

This work addresses the challenge of acoustic modeling under small-sample conditions in animal emotion recognition, proposing the first end-to-end audio-based emotion classification framework tailored to individual dogs. To overcome data scarcity, the method fuses Mel-spectrogram representations with time-domain statistical acoustic features and establishes a comparative multi-model framework comprising SVM, random forests, and shallow neural networks. Evaluated on a single-dog audio dataset, it achieves over 70% overall emotion classification accuracy. The study breaks through a key technical bottleneck in quantifying animal emotions from limited acoustic data, empirically validating the feasibility of fine-grained canine emotion discrimination using audio signals alone. It introduces a novel paradigm for interpretable, individualized animal affective computing—advancing both the understanding of animal intent in human–machine interaction and the development of robust, dog-specific emotional inference tools.

0 citationsRead paper