Publicly available simulation framework for Sensory Neural Networks based on resistive sensing elements
Overview
The NEURALSENS software framework supports the modelling and design of Sensory Neural Networks (SeNNs) for low-level in-sensor data processing.
The framework explores how the electrical and sensitivity characteristics of sensing elements can be incorporated directly into the network model as physically defined parameters. This approach provides a bridge between conventional neural-network processing and hardware-oriented Sensory Neural Networks.
Software concept
The simulation framework is implemented in Python and combines conventional neural-network modelling with sensor-dependent response functions.
The current implementation includes:
- a conventional fully connected neural network used as a reference model,
- sensor-response functions represented by polynomial characteristics,
- sensor-dependent network weights,
- simulation of temperature-dependent sensor behaviour,
- training and error evaluation,
- comparison between conventional ANN processing and sensor-dependent SeNN processing,
- visualization of predicted outputs, errors and training loss.
The software therefore provides a computational environment for investigating how physical sensor characteristics can contribute directly to the implementation of neural-network functionality.
Relation to the NEURALSENS project
Within NEURALSENS, the SeNN simulation methodology is developed as part of KPB2.
The project methodology starts from experimentally measured sensor characteristics and uses these data to define and optimize the response of sensing nodes in a crossbar-array architecture. The same framework is subsequently intended to be extended from temperature sensing towards hydrogen sensing elements.
The long-term objective is to generalize the methodology for crossbar-compatible resistive sensing elements and enable rapid design of new in-sensor computing architectures.
Software repository
The current NEURALSENS SeNN simulation framework is publicly available on GitHub:
GitHub repository:
https://github.com/jaro221/NEURALSENS_PNN
The repository contains the Python implementation developed within the NEURALSENS project.
Software implementation: Jaromír Klarák, Institute of Informatics, Slovak Academy of Sciences (II-SAS).
Project funding
Development of the software framework was supported by the European Union – NextGenerationEU through the Recovery and Resilience Plan for Slovakia under project No. 09I05-03-V02-00058 (NEURALSENS).
Related project results
The software framework is complemented by experimental work on geometry-dependent TiO₂ gas-sensor responses and temporal feature extraction for subsequent neural-network integration.