Instability of artificial brain can actually improve learning ability of artificial neural networks

The research team led by Chung Hong-sik, a professor at UNIST’s Department of New Materials Engineering, and researchers at Tsinghua University in China, proposed a new learning method that improves the learning ability of artificial neural networks by using instability of artificial neural network chips. Artificial neural network chips are future technologies that make neurons and synapses in the brain with semiconductor chips. Artificial neural network chips, like the human brain, can perform both computational and memory operations while using less energy. However, artificial neural network chips, which are physically integrated with many devices, often show errors. Since the existing artificial neural network learning method is a learning method assuming a perfect artificial neural network chip with no error, the actual learning ability is low.

The research team developed a phase-change memory-based memristor artificial neural network learning method based on the fact that the actual human brain does not need near-perfect behavior. This method reflects ‘resistance drift phenomenon’ (increased electrical resistance) of phase change materials in memory semiconductors. Since information update patterns are recorded in the form of increased electrical resistance in memristors that act as synapses during the learning process, synapses learn additionally the changing patterns and the associations between the data they learn.

The research team showed that the learning method developed through an experiment that classifies handwriting composed of numbers 0 to 9 has a learning ability improvement effect of about 3%. In particular, in the case of the number 8, which is difficult to classify with handwriting, the accuracy has been improved significantly. Learning ability has been improved thanks to the synaptic update patterns that change with the degree of difficulty in handwriting classification.

The researchers expect that the study will serve as a catalyst for artificial intelligence-based convergence research that connects two important topics in the field of artificial intelligence: the development of artificial neural network chips and the implementation of brain neoclassical function through artificial neural networks. The study was published in the online edition of the international journal Nature Communications on January 12, 2021.

Source: Lim, DH., Wu, S., Zhao, R. et al. Spontaneous sparse learning for PCM-based memristor neural networks. Nat Commun 12, 319 (2021).

Categories: Tech&Innovation