Sunday 02 February 2025
A team of researchers has made significant progress in developing a method for inferring the parameters of neural models on neuromorphic hardware, which mimics the behavior of biological neurons. The approach uses a combination of autoencoders and simulation-based inference to compress high-dimensional data and estimate the posterior distribution of model parameters.
The researchers used the BrainScaleS-2 system, a mixed-signal neuromorphic chip that can accelerate neural simulations by orders of magnitude compared to traditional computing architectures. They generated a dataset of membrane voltage recordings from an adaptive exponential integrate-and-fire (AdEx) neuron model, which is a common type of neural model used in neuroscience.
The team then trained an autoencoder on the dataset to compress the high-dimensional data into a lower-dimensional representation, which can be used for inference. The autoencoder was designed using convolutional layers and batch normalization, and it achieved impressive reconstruction accuracy.
Next, the researchers used the encoder part of the autoencoder to reduce the dimensionality of the data and fed it into a neural density estimator (NDE) using the sequential neural posterior estimation (SNPE) algorithm. The NDE is a type of probabilistic model that can be trained to estimate the posterior distribution of model parameters.
The results show that the method was able to accurately infer the parameters of the AdEx neuron model, including the adaptation current and the spike-triggered adaptation. The inferred posterior distributions were also found to be consistent with the true values used to generate the dataset.
This work has significant implications for neuroscience research, as it enables the simulation-based inference of neural models on neuromorphic hardware. This can accelerate the development of new algorithms and models for neural simulations, which can in turn help us better understand the behavior of biological neurons.
The approach also has potential applications in other fields, such as robotics and artificial intelligence, where fast and efficient neural simulations are crucial for real-time decision-making. Overall, this work demonstrates the power of combining autoencoders with simulation-based inference to accelerate the development of new algorithms and models for neural simulations.
Cite this article: “Inferring Neural Model Parameters on Neuromorphic Hardware Using Autoencoders and Simulation-Based Inference”, The Science Archive, 2025.
Neural Networks, Neuromorphic Hardware, Autoencoders, Simulation-Based Inference, Brainscales-2, Adex Neuron Model, Neural Density Estimator, Sequential Neural Posterior Estimation, Neuroscience Research, Robotics, Artificial Intelligence
