An Intrinsically Knowledge-Transferring Developmental Spiking Neural Network for Tactile Classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Gradient descent with backpropagation (BP) is common for training spiking neural networks (SNNs).
• BP-based SNN training faces challenges like manual architecture tuning, catastrophic forgetting, and high computational cost.
• Existing methods struggle with efficient continual learning and adapting to new information.

Purpose of the Study:

• To introduce brain-mimetic developmental spiking neural networks (BDNNs) that emulate biological neural development.
• To address limitations of BP-based SNN training, including manual tuning, forgetting, and computational demands.
• To evaluate BDNNs in a neuromorphic tactile system for object classification during grasping.

Main Methods:

• Developed BDNNs inspired by postnatal neural circuit development.
• Implemented dynamic neuron recruitment in response to input data.
• Evaluated BDNNs on a tactile object recognition task using a neuromorphic system.
• Compared BDNN performance against standard BP methods and continual learning algorithms.

Main Results:

• BDNNs demonstrated dynamic growth by recruiting neurons incrementally, enhancing classification accuracy without manual tuning.
• The growth process autonomously adapted to data complexity, showing robust knowledge transfer for learning new objects.
• BDNNs achieved performance comparable to BP methods but learned 100-1000 times faster.
• BDNNs outperformed existing continual learning algorithms in both performance and speed.

Conclusions:

• BDNNs offer a promising, self-adapting approach for continual learning in artificial neural networks.
• The dynamic growth and knowledge transfer capabilities make BDNNs suitable for real-time edge computing.
• BDNNs present a significant advancement over traditional SNN training methods, particularly for evolving datasets.