arXiv subjects

Akshaya Rajesh

Publications and source records attributed to Akshaya Rajesh.

3 recordsLinked to original sources

arXiv2026

Autonomous Chaotic Time Series Prediction using Physical Neuromorphic Networks

Physical reservoir computing (PRC) with neuromorphic networks offers a promising approach to brain-inspired information processing, exploiting emergent nonlinear dynamics of physical neural networks as a computational resource. This study demonstrates fully autonomous closed-loop prediction of the Mackey--Glass (MG) chaotic time series using a simulated neuromorphic nanowire network as the physical reservoir. Two strategies are evaluated: the virtual node (VN) method, which expands the feature space by temporal multiplexing of reservoir states, and a non-VN approach that uses all physical node readouts directly without temporal multiplexing. Results are reported for two values of the MG time delay parameter, $\tau = 18$ and $\tau = 21$, the latter representing a more complex chaotic regime not previously evaluated for this class of physical reservoir. Over a short prediction horizon of $T = 100$ timesteps, the VN approach achieves autonomous prediction accuracies of $90.4$% and $89.7$% at $\tau = 18$ and $\tau = 21$, respectively, while the non-VN approach achieves $81.5$% and $76.2$%. Long-horizon analysis over $T = 500$ timesteps shows that both approaches reproduce the qualitative attractor structure and dominant spectral content of the true MG signal, with trajectories remaining bounded throughout. These results suggest that the intrinsic dynamics of neuromorphic nanowire networks are sufficient to support meaningful autonomous chaotic time series prediction without virtual node augmentation, and that performance may improve further as physical network sizes scale to the millions of nodes achievable in hardware. As this study uses simulated networks, extrapolation to physically fabricated large-scale arrays remains to be validated experimentally.

cond-mat.dis-nn↗

arXiv2025

Memristive Nanowire Network for Energy Efficient Audio Classification: Pre-Processing-Free Reservoir Computing with Reduced Latency

Efficient audio feature extraction is critical for low-latency, resource-constrained speech recognition. Conventional preprocessing techniques, such as Mel Spectrogram, Perceptual Linear Prediction (PLP), and Learnable Spectrogram, achieve high classification accuracy but require large feature sets and significant computation. The low-latency and power efficiency benefits of neuromorphic computing offer a strong potential for audio classification. Here, we introduce memristive nanowire networks as a neuromorphic hardware preprocessing layer for spoken-digit classification, a capability not previously demonstrated. Nanowire networks extract compact, informative features directly from raw audio, achieving a favorable trade-off between accuracy, dimensionality reduction from the original audio size (data compression) , and training time efficiency. Compared with state-of-the-art software techniques, nanowire features reach 98.95% accuracy with 66 times data compression (XGBoost) and 97.9% accuracy with 255 times compression (Random Forest) in sub-second training latency. Across multiple classifiers nanowire features consistently achieve more than 90% accuracy with more than 62.5 times compression, outperforming features extracted by conventional state-of-the-art techniques such as MFCC in efficiency without loss of performance. Moreover, nanowire features achieve 96.5% accuracy classifying multispeaker audios, outperforming all state-of-the-art feature accuracies while achieving the highest data compression and lowest training time. Nanowire network preprocessing also enhances linear separability of audio data, improving simple classifier performance and generalizing across speakers. These results demonstrate that memristive nanowire networks provide a novel, low-latency, and data-efficient feature extraction approach, enabling high-performance neuromorphic audio classification.

cs.SD↗

arXiv2025

Learn Like Feynman: Developing and Testing an AI-Driven Feynman Bot

The Feynman learning technique is an active learning strategy that helps learners simplify complex information through student-led teaching and discussion. In this paper, we present the development and usability testing of the Feynman Bot, which uses the Feynman technique to assist self-regulated learners who lack peer or instructor support. The Bot embodies the Feynman learning technique by encouraging learners to discuss their lecture material in a question-answer-driven discussion format. The Feynman Bot was developed using a large language model with Langchain in a Retrieval-Augmented-Generation framework to leverage the reasoning capability required to generate effective discussion-oriented questions. To test the Feynman bot, a controlled experiment was conducted over three days with fourteen participants. Formative and summative assessments were conducted, followed by a self-efficacy survey. We found that participants who used the Feynman Bot experienced higher learning gains than the Passive Learners' group. Moreover, Feynman Bot Learners' had a higher level of comfort with the subject after using the bot. We also found typing to be the preferred input modality method over speech, when interacting with the bot. The high learning gains and improved confidence with study material brought about by the Feynman Bot makes it a promising tool for self-regulated learners.

physics.ed-ph↗