Reproduction of AdEx dynamics on neuromorphic hardware through data embedding and simulation-based inference

TL;DR

This work tackles automated parameter calibration for a complex AdEx neuron model emulated on neuromorphic hardware, where manual tuning is impractical. It introduces a pipeline that first uses a convolutional autoencoder to extract a compact feature representation from membrane traces and then applies simulation-based inference (SNPE) to approximate the posterior $p(\boldsymbol{\theta}|\boldsymbol{x}^*)$ over model parameters. The approach, demonstrated on BrainScaleS-2, shows that the encoder distills relevant dynamics and that SNPE identifies parameter regions consistent with the target trace, even in the presence of hardware-induced temporal noise, with posteriors centered near $V_r$ and $g_{\tau_w}$. This methodology offers a promising path toward rapid, automated emulation and calibration of neural dynamics on neuromorphic hardware, with potential extensions to biological neurons and broader complex-system calibration.

Abstract

The development of mechanistic models of physical systems is essential for understanding their behavior and formulating predictions that can be validated experimentally. Calibration of these models, especially for complex systems, requires automated optimization methods due to the impracticality of manual parameter tuning. In this study, we use an autoencoder to automatically extract relevant features from the membrane trace of a complex neuron model emulated on the BrainScaleS-2 neuromorphic system, and subsequently leverage sequential neural posterior estimation (SNPE), a simulation-based inference algorithm, to approximate the posterior distribution of neuron parameters. Our results demonstrate that the autoencoder is able to extract essential features from the observed membrane traces, with which the SNPE algorithm is able to find an approximation of the posterior distribution. This suggests that the combination of an autoencoder with the SNPE algorithm is a promising optimization method for complex systems.

Figures (5)

• Figure 1: Random samples drawn from the dataset. -- For visualization, only the first 300 are displayed; the traces are recorded for 1. Due to the finite sampling frequency of the and the interpolation of the recorded traces, the potentials at spike time are not identical. When the reset potential $V_\text{r}$ is high, the membrane voltage remains at high levels.
• Figure 2: Training of the autoencoder -- Mean test and validation loss during training as well as one standard deviation of the validation loss. Both losses decrease continuously over the course of the training.
• Figure 3: Reconstructions of the autoencoder -- Randomly drawn voltage traces from the test set and their reconstructions by the trained autoencoder. Traces are only displayed for 300 to aid visual comparison. The reconstructions (red) follow the original traces (black) closely on a long time scale. However, unlike the original traces, they show some high frequency fluctuations.
• Figure 4: Samples drawn from the approximated posterior -- One- and two-dimensional marginals of 500.0 samples drawn from the approximated posterior. The vertical and horizontal lines represent the parameterization of the target trace. Note, uniform priors from 01022 were chosen for all parameters, i.e. the posterior distribution is restricted to a much smaller region of the parameter space.
• Figure 5: Example traces for different experiment trials and parameterizations drawn from the approximated posterior -- Black traces represent the chosen target observation. On the left side, the experiment is repeated several times with the same parameterization. Due to temporal fluctuations in the analog components of bss2, the traces do not align exactly. However, the overall behavior matches between all traces. On the right side, we drew four parameterizations form the approximated posterior, \ref{['fig:sbi']}, and emulated the neuron behavior.