Unique Presentation Identifier:
41
Program Type
Graduate
Faculty Advisor
Dr. Tolga Ensari
Document Type
Poster
Location
Face-to-face
Start Date
9-4-2026 1:00 PM
End Date
4-2026 3:00 PM
Abstract
High dimensional temporal data processing, such as that required for neuroprosthetics and remote physiological monitoring presents significant challenges for real time deployment because transmitting and storing raw signals is computationally demanding and energy intensive. Effective data compression is essential to act as a "biological zip file," reducing transmission bandwidth while preserving the critical temporal features required for accurate signal reconstruction and analysis. This study proposes a deep Spiking Neural Network (SNN) Autoencoder designed for high-fidelity data compression by utilizing the event-driven firing behavior of Leaky Integrate-and-Fire (LIF) neurons, which ensures extreme computational efficiency compared to traditional models. The model is evaluated using the Spoken Heidelberg Digits (SHD) dataset, featuring a symmetrical encoder-decoder architecture where the network maps a 700 dimensional input into a highly compressed 64-neuron latent bottleneck through sequential hidden layers of 512 and 256 neurons. Each layer incorporates Layer Normalization for training stability and Dropout for robustness, while a strict 8% sparsity constraint is applied within the 64 neuron bottleneck to ensure only essential temporal information is retained. The decoder mirrors this structure to reconstruct the original 700 channel spike trains from the latent representation. The proposed SNN based autoencoder achieves a 10.9x structural compression ratio while maintaining a high spike train similarity of 91.34% during reconstruction. By incorporating learnable synaptic and membrane decay parameters, the network adaptively learns the temporal dynamics of the dataset, validating that SNN based autoencoders provide a viable, low power solution for compressing high dimensional temporal data in resource constrained neuromorphic hardware.
Recommended Citation
Bhandari, Shruti, "Deep Spiking Neural Network Autoencoders for Efficient Temporal Data Compression" (2026). ATU Scholars Symposium. 58.
https://orc.library.atu.edu/atu_rs/2026/2026/58
Deep Spiking Neural Network Autoencoders for Efficient Temporal Data Compression
Face-to-face
High dimensional temporal data processing, such as that required for neuroprosthetics and remote physiological monitoring presents significant challenges for real time deployment because transmitting and storing raw signals is computationally demanding and energy intensive. Effective data compression is essential to act as a "biological zip file," reducing transmission bandwidth while preserving the critical temporal features required for accurate signal reconstruction and analysis. This study proposes a deep Spiking Neural Network (SNN) Autoencoder designed for high-fidelity data compression by utilizing the event-driven firing behavior of Leaky Integrate-and-Fire (LIF) neurons, which ensures extreme computational efficiency compared to traditional models. The model is evaluated using the Spoken Heidelberg Digits (SHD) dataset, featuring a symmetrical encoder-decoder architecture where the network maps a 700 dimensional input into a highly compressed 64-neuron latent bottleneck through sequential hidden layers of 512 and 256 neurons. Each layer incorporates Layer Normalization for training stability and Dropout for robustness, while a strict 8% sparsity constraint is applied within the 64 neuron bottleneck to ensure only essential temporal information is retained. The decoder mirrors this structure to reconstruct the original 700 channel spike trains from the latent representation. The proposed SNN based autoencoder achieves a 10.9x structural compression ratio while maintaining a high spike train similarity of 91.34% during reconstruction. By incorporating learnable synaptic and membrane decay parameters, the network adaptively learns the temporal dynamics of the dataset, validating that SNN based autoencoders provide a viable, low power solution for compressing high dimensional temporal data in resource constrained neuromorphic hardware.