An Analog Chemical Circuit with Parallel-Accessible Delay Line for Learning Temporal Tasks

Current synthetic chemical systems lack the ability to self-modify and learn to solve desired tasks. In this paper we introduce a new parallel model of a chemical delay line, which stores past concentrations over time with minimal latency. To enable temporal processing, we integrate the delay line with our previously proposed analog chemical perceptron. We show that we can successfully train our new memory-enabled chemical learner on four non-trivial temporal tasks: the linear moving weighted average, the moving maximum, and two variants of the Nonlinear AutoRegressive Moving Average (NARMA). Our implementation is based on chemical reaction networks and follows mass-action and Michaelis-Menten kinetics. We show that despite a simple design and limited resources, a single chemical perceptron extended with memory of variable size achieves 93-99% accuracy on the above tasks. Our results present an important step toward actual biochemical systems that can learn and adapt. Such systems have applications in biomedical diagnosis and smart drug delivery.