Forum Numerica - Laurent PERRINET : Modelling memory to understand the brain

Forum Numerica - Laurent PERRINET : Modelling memory to understand the brain

Title history
  • Forum Numerica - Laurent PERRINET : Modelling memory to understand the brain — observed on 20261110-1852 UTC (YouTube)
🎙 Laurent Perrinet 👥 161 📅 November 10, 2026 ⏱ 55 min 👁 4 📄 original study 🧭 2026-11-10
Available in: English (current) Français

Keywords

spiking neural networkmemorypolychronizationLIF modeltransformer

Summary

Laurent Perrinet presents a research seminar on modelling memory using spiking neural networks (SNNs). He begins with the thermodynamic concept of entropy and the Maxwell demon as an analogy for living systems that maintain order through memory. He contrasts traditional recurrent architectures like LSTM and transformers with biologically plausible spiking neurons, emphasizing the importance of spike timing. The core proposal is the use of heterogeneous delays in recurrent spiking networks to form polychronous chains, enabling the detection and prediction of temporal patterns. He derives the leaky integrate-and-fire (LIF) neuron model and shows how surrogate gradient methods allow training via backpropagation. Initial results demonstrate the ability to learn temporal patterns and generate gait outputs. The talk bridges computational efficiency and biological realism, aiming to inspire more interpretable AI systems.

128 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides substantive value by presenting an original computational framework that leverages heterogeneous synaptic delays to create memory in spiking networks, extending Izhikevich’s polychronization concept to a trainable recurrent setting. The argumentation is solid, grounded in established neuroscience (precise spike timing) and AI (LSTM, transformers) literature. Perrinet clearly explains the limitations of current approaches (energy intensity of transformers) and how his model addresses these. The use of surrogate gradients and LIF neurons is technically sound, and the preliminary results are promising, though he acknowledges their early stage.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor through detailed mathematical formulation of the LIF model and references to seminal papers (e.g., Hochreiter & Schmidhuber on LSTM, Vaswani et al. on transformers, Izhikevich on polychronization). The single description link points to the seminar series page, which is a primary institutional source. The title is an accurate summary of the content, neither overstating nor understating the scope. No user comments were available to analyse.

173 words

Title / Content Match

The title accurately reflects the content, as the talk centers on computational modelling of memory via spiking neural networks to elucidate brain mechanisms.

Quality & Reliability

8/10

The talk presents a computationally rigorous approach rooted in published research, with clear technical details and references to foundational work. However, as a seminar presentation, some results are preliminary and not peer-reviewed.

Key Moments

Cited Sources

Contribution & Novelties

The original contribution lies in extending polychronous chain theory to a learnable recurrent spiking network with heterogeneous delays, enabling memory formation and pattern prediction in a biologically plausible framework. This bridges the gap between energy-hungry transformers and brain-inspired computing.

Pour aller plus loin :

75 words

Radar Profile

The profile shows high technical depth and moderate to high information quality, with slightly lower global reliability due to preliminary results. The balance between theoretical novelty and practical demonstration is strong, but the reliance on backpropagation and limited scale temper the ecological claims.

Reliability 7/10