Amortized Inference of Neuron Parameters on Analog Neuromorphic Hardware

2026-02-13 · Source: cs.NE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Neuromorphic Computing · Depth: Expert, extended

Summary

This work introduces a non-sequential simulation-based inference (SBI) algorithm to create an amortized neural density estimator (NDE) for approximating the posterior distribution of seven parameters of the adaptive exponential integrate-and-fire (AdEx) neuron model on the analog neuromorphic BrainScaleS-2 (BSS-2) substrate. The researchers constrained the extensive parameter space by training a binary classifier to identify parameter combinations that produce moderate spike counts (1-70 kHz), increasing relevant samples from 2% to 38%. They compared two NDEs: one using handcrafted summary statistics and another incorporating a summary network trained jointly with the NDE. The summary network approach yielded a more focused posterior distribution and generated posterior predictive traces that more accurately captured membrane potential dynamics, whereas handcrafted statistics showed deviations in exact dynamics despite matching included features. The results validate amortized SBI for parameterizing analog neuron circuits, despite observed biases and miscalibration in posteriors.

Key takeaway

For research scientists working with neuromorphic hardware or complex biological models, this study demonstrates that amortized simulation-based inference (SBI) can efficiently parameterize analog neuron circuits. You should prioritize integrating summary networks over handcrafted summary statistics when training neural density estimators (NDEs) to achieve more accurate membrane potential dynamics and focused posterior distributions, especially when dealing with high-dimensional observation data. This approach can significantly reduce the effort in identifying suitable model parameters.

Key insights

Amortized SBI effectively parameterizes analog neuron circuits on neuromorphic hardware, with summary networks outperforming handcrafted statistics.

Principles

• Constrain large parameter spaces to focus inference on regimes of interest.
• Neural density estimators can approximate complex posterior distributions.
• Summary networks improve posterior focus and dynamic accuracy over handcrafted features.

Method

A non-sequential SBI algorithm, BayesFlow, was used to train an amortized NDE. A binary classifier pre-filtered parameter space for moderate spike counts. Two NDEs were compared: one with handcrafted summary statistics and another with a jointly trained summary network.

In practice

• Use binary classifiers to pre-filter large parameter spaces for relevant behaviors.
• Consider summary networks for improved accuracy in membrane potential dynamics.
• Apply amortized SBI to parameterize stochastic analog neuron circuits.

Topics

• Amortized Inference
• Neuromorphic Hardware
• Simulation-Based Inference
• Neural Density Estimators
• AdEx Neuron Model

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Real-time processing of analog signals on accelerated neuromorphic hardware — Direct analog signal injection into neuromorphic hardware enables efficient, real-time, near-sensor processing.
• Modelling and Simulation of Neuromorphic Datasets for Anomaly Detection in Computer Vision — ANTShapes simulates neuromorphic vision datasets for anomaly detection, overcoming DVS data scarcity via configurable 3D scenes.
• One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators — Neural operators can universally approximate probabilistic conditioning, enabling amortized inference across diverse joint distributions.
• Otters++: A Time-to-first-spike Based Energy Efficient Optical Spiking Transformer — Otters++ leverages optoelectronic device decay for energy-efficient time-to-first-spike computation in spiking Transformers.
• Robust Parameter and State Estimation in Multiscale Neuronal Systems Using Physics-Informed Neural Networks — PINNs offer robust parameter and state estimation in multiscale neuronal systems, overcoming limitations of traditional methods.
• Convolutional Sparse Coding via the Locally Competitive Algorithm on Loihi 2 — Convolutional sparse coding via LCA is now implemented and benchmarked on Loihi 2, establishing a new benchmark for structured sparse inference on neuromorphic systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.NE updates on arXiv.org.