Bio-Inspired Techniques in a Fully Digital Approach for Lifelong Learning

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Bio-inspired Computing

Background:

• Biological organisms exhibit lifelong learning for environmental adaptation.
• Artificial neural networks (ANNs) lack the dynamic flexibility of biological systems.
• A new paradigm is needed to combine biological resilience with ANN accuracy.

Purpose of the Study:

• To present a digital implementation of a novel mixed supervised-unsupervised neural network.
• To enable lifelong learning capabilities in artificial systems.
• To bridge the gap between biological and artificial intelligence.

Main Methods:

• Utilized convolutional filters for feature extraction from MNIST and Fashion-MNIST datasets.
• Implemented a mixed supervised-unsupervised approach with transfer learning.
• Employed spike-timing dependent plasticity (STDP) for unsupervised learning and clustering of non-trained data.
• Demonstrated the network on a Xilinx Zynq-7000 System on Chip (SoC).

Main Results:

• The network successfully merged past and new information, demonstrating lifelong learning.
• Bio-inspired algorithms like neuronal redundancy and spike-frequency adaptation facilitated clustering of novel data.
• A user-friendly interface allowed testing with custom inputs and non-trained classes.
• Comparison with memristive devices highlighted the digital approach's capabilities.

Conclusions:

• The developed neural network offers a resilient and accurate lifelong learning system.
• Digital implementation on SoC provides a practical platform for bio-inspired AI.
• This approach advances the development of adaptable and continuously learning artificial intelligence.