Energy-based analog neural network framework.

Area of Science:

• Neuromorphic Engineering
• Machine Learning
• Nanotechnology

Background:

• Neuromorphic systems show disruptive potential but are limited by slow, laborious model development.
• Existing methods for building, training, and evaluating mixed-signal neural models are complex and time-consuming.

Purpose of the Study:

• Introduce EBANA, an open-source framework for building and validating analog neural networks (ANNs).
• Provide a unified, modular, and extensible infrastructure for ANN research, similar to conventional machine learning pipelines.

Main Methods:

• EBANA offers a Python interface with Keras-like syntax, abstracting analog simulation complexities.
• The framework includes common building blocks and supports easy integration of new electrical and technological models.
• Demonstrates capabilities using Energy-Based Models (EBMs) with Equilibrium Propagation (EP) training.

Main Results:

• Experiments utilized 3 datasets with up to 60,000 entries.
• Network topologies with over 1,000 electrical nodes were explored.
• EBANA's native parallelization enabled efficient benchmarking of complex ANNs.

Conclusions:

• EBANA facilitates experimentation with diverse ANN topologies and design space trade-offs.
• The framework accelerates research in neuromorphic computing and advanced machine learning applications.
• EBANA empowers researchers to explore complex analog neural network designs efficiently.