Thomas Louis

Thesis subject : Optimization and deployment of classical and bio-Inspired neural networks on spatializable embedded architectures. Develops quantization methods to minimize bit-depth in SNNs, analyzing trade offs in energy efficiency and performance between quantized SNNs and FNNs. Uses knowledge distillation and regularization to decrease spiking activity in SNNs, enhancing energy efficiency while maintaining accuracy. Introduces hybrid networks combining SNNs and FNNs, as well as multi timestep networks that process information at varied latencies, achieving reduced energy consumption without sacrificing performance. Validates approaches through comparative analysis on public datasets, assessing accuracy, energy use, and SNN activity. Develops QUALIA, a Framework to train, compress (quantization, knowledge distillation and regularization), deploy (on MCU and FPGA) and estimate energy consuption of FNNs and SNNs.

Part-time instructor on Machine Learning and Embedded AI. My courses covered fundamental concepts as well as practical applications, helping students understand and implement AI techniques in embedded systems.

Development of a C framework to convert models and execute inference in various formats (FLOAT32, INT16 with dynamic/static normalization). Pruning (Keras/TF) and deploying a Deep Learning object detection algorithm intended for space-oriented targets. Benchmarking of embedded targets and Deep Learning deployment frameworks.