Zacharelos, Efstratios (2023) Integrated Spike Detection using Approximate Methods. [Tesi di dottorato]

Preview

Text Zacharelos_Efstratios_35.pdf Download (5MB) | Preview

Collection description

Research on Brain-Machine Interfaces has progressed at a notable speed since the realization that devices directly controlled by ensembles of cortical neurons are viable. Recording the voltage fluctuations resulting from neural oscillations helps us determine changes in brain activity that might be useful in diagnosing brain disorders, like epilepsy. These readings reflect the summation of the synchronous activity of millions of neurons that have similar spatial orientation. Transmitting massive data wirelessly for external post-processing poses unrealistic bandwidth and power requirements. Spike detection algorithms can mitigate the problem by processing the raw signal on-line and transmitting only the relevant information off-brain. During the last years, machine learning approaches have been considered for spike detection and classification. The innate error factor in machine learning and the vast number of required multiplications in a NN, make introducing approximate computing to this application, a viable option. In this work, two main categories of approximate binary multipliers are considered: multipliers using approximate compressors and multipliers using recursive architectures with smaller approximate modules. In this way, a significant reduction in silicon area and power dissipation can be achieved. A dense network has been developed to detect and categorize spikes in simulated brain activity, achieving an accuracy of 98%. The model’s architecture has been described in HDL-Verilog. Using 12 bits for signal representation, the accuracy remains at high levels, 97%. By employing approximate methods, a reduction in silicon area by 34% can be achieved, for an accuracy drop of only 1.4%.

Downloads

Actions (login required)

View Item