Emulating Neural Assemblies by Molecular Assemblies: Molecular Computing Simulation of Digit Recognition

Je-Hwan Ryu, Hyo-Sun Chun, Christina Baek, Ji-Hoon Lee, and Byoung-Tak Zhang

Abstract

Cognition is an emergent phenomenon of the brain through the massive interactions of nerve cells. Thus, it is highly likely that studying the massive interactions of molecules sheds light on the computational principles of the neural assemblies. Here we study the collective behavior of neural assemblies by emulating it with molecular assemblies using DNA computing technology. We use the handwritten digit recognition problem as a case and perform in silico simulations. We use the DNA hypernetwork structure as a pattern recognition model based on a molecular assembly model of the neural assemblies by interpreting a collection of DNA hyperedges (molecular assembly) as a neural assembly. The DNA hyperedges represent the combinations of pixels of the digit image data and encoded as DNA molecules. During the learning phase, the hyperedges are created, selected, and amplified to change the molecular assemblies in the DNA hypernetwork. In the pattern recognition phase, a given handwriting image is converted to a collection of random hyperedges which are then matched to the hyperedges of the DNA hypernetwork. The degree of matching determines the recognition label of the new image. Our simulation results of incremental learning on a real-life training dataset (MNIST data of size of order 10^4 or higher) demonstrate the statistical tendency of monotonic improvement of recognition accuracy as new training data are observed online. The analysis of the learned hypernetwork structure reveals interesting “building blocks” (molecular assemblies) which can be re-interpreted as “neural assemblies” potentially important for solving the pattern recognition problem.