Unsupervised learning in neuromorphic computing: self-organizing maps vs. spiking neural networks
During the last years, Deep Neural Networks (DNNs) have reached the highest performances in image classification. Nevertheless, such a success is mostly based on supervised and off-line learning: they require thus huge labeled datasets for learning, and once it is done, they cannot adapt to any change in the data from the environment. Moreover, DNNs are moving from the initial biological inspiration of Artificial Neural Networks, as it is unlikely that such an algorithm based on neuron-specific error signal (back-propagation) would be implemented in the brain. Instead, biological evidences lead rather to unsupervised learning methods.
As a PhD candidate in neuromorphic computing, I am confronting two approaches for unsupervised learning: the Kohonen-based Self-Organizing Maps (SOMs) and the STDP-based Spiking Neural Networks (SNNs). Our preliminary results showed that the SOMs offer a better compromise in terms of accuracy, dynamicity and scalability, thanks to a cellular neuromorphic architecture. However, SNNs have a lower hardware-cost thanks to the spike coding, and the neuromorphic computing leading companies (IBM, Intel, BrainChip, etc.) seem to follow this approach that has a closer relationship with the brain in terms of information coding.
I am aiming to discuss this question and the pros/cons arguments for both approaches, then try to have a constructive confrontation on both biological plausibility and quantitative performances for neuromorphic computing.
|Fri, 26.04.2019||20:30 - 21:30||Sala Panorama|