PhD Thesis Defense - Jonathan Courtois – Université Côte d’Azur · CNRS · LEAT : Sparsity in SNNs

PhD Thesis Defense - Jonathan Courtois – Université Côte d’Azur · CNRS · LEAT : Sparsity in SNNs

by Jonathan Courtois

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQNNeural networks and fuzzy systems
🎙 Jonathan Courtois 👥 154 📅 December 16, 2025 ⏱ 53 min 👁 2K 📄 original study 🧭 2026-08-15
Available in: English (current) Français

Keywords

spiking neural networksevent-based visionneuromorphic computingenergy efficiencyFPGA

Summary

This PhD defense presents a comprehensive methodology for designing, evaluating, and deploying spiking neural networks (SNNs) for embedded event-based vision, with a focus on space applications. The work introduces a hardware-agnostic energy efficiency metric that reveals SNNs can achieve 6-8x energy reduction compared to formal neural networks (FNNs) at comparable accuracy, with memory access dominating energy consumption. The thesis extends the QUALIA framework for training and deploying SNNs on various platforms, integrates the SPLEAT FPGA-based neuromorphic accelerator, and develops QUALIABENCH for automated benchmarking. Advanced applications include object detection on the Gen1 dataset, achieving 46% faster inference and 3x energy efficiency on FPGA compared to CPU, and satellite pose estimation using direct and indirect methods. The work demonstrates the feasibility of SNNs for complex tasks in constrained environments, highlighting the importance of detailed sparsity analysis and the potential for further optimization.

140 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation provides valuable insights into the practical energy efficiency of SNNs, challenging earlier optimistic estimates. The argumentation is solid, supported by quantitative results from multiple datasets and real hardware deployments. The speaker systematically builds from metrics to tools to applications, demonstrating a clear logical progression. The discussion of limitations, such as the variability in pose estimation results and the computational bottlenecks, adds credibility. The work contributes original tools and benchmarks that are open-source, enhancing its value to the research community.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high, with references to relevant literature and published papers. The speaker cites specific studies (e.g., Dufour et al., 2022) and acknowledges limitations in the energy metric. The sources are credible, including peer-reviewed conferences and journals. The title accurately reflects the content, focusing on sparsity in SNNs. The presentation includes a clear structure with chapters, aiding comprehension. The defense includes a formal oath, indicating adherence to academic standards.

167 words

Title / Content Match

The title accurately reflects the content, focusing on sparsity in spiking neural networks for embedded event-vision processing.

Quality & Reliability

8/10

The defense presents original research with clear methodology, quantitative results, and references to published work. The speaker is a PhD candidate, and the content is peer-reviewed through the thesis process. However, as a presentation, some details are summarized and not fully verifiable.

Chapters

Cited Sources

  • QUALIA framework — Open-source framework for designing, training, and deploying neural networks, extended in this thesis.
  • SPLEAT accelerator — Configurable neuromorphic accelerator on FPGA, used for deployment.
  • Scenic dataset — Event-based dataset for spacecraft pose estimation, used in the thesis.

Concurring Sources

  • Dufour et al. (2022) — Independent study finding similar energy efficiency results for SNNs, supporting the thesis's conclusions.

Contribution & Novelties

The thesis provides a novel hardware-agnostic energy metric for SNNs, a complete toolchain for deployment, and demonstrates the first embedded event-based object detection on neuromorphic hardware. It also explores satellite pose estimation with SNNs, a relatively unexplored area.

Pour aller plus loin :

71 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-rounded and rigorous presentation. The strongest aspects are the quantity and quality of information, as well as technical depth, reflecting the comprehensive nature of the thesis work.

Reliability 8/10