Evolution of an artificial visual cortex for image recognition

Evolutionary robotics has been successful in creating agents that successfully link perception with appropriate action. However, the visual fields utilized by such agents is usually extremely small compared to the retinas linked to the visual cortex of animals. Evolving a cortex that processes larger fields of view in a selective and robust manner is challenging because fitness landscapes that are sensitive to this level of details are difficult to design. Here, we decouple the perception from the action part of evolutionary robotics, and present a newway to evolve logic circuits to perform image recognition on the well-known MNIST data set, which comprises 60,000 training and 10,000 testing handwritten numerals. The logic circuits are encoded in a genome that is evolved using a fitness function based on the true positive and true negative classification rates of the numerals. Following evolution, individual circuits achieve in excess of 80% recognition accuracy on the testing data. By pooling highly evolved individual circuits from multiple evolutionary histories into a committee, testing accuracy is increased to 93.5%. This work demonstrates that evolving logic circuits to solve a classification task is feasible. We also found the evolved circuits to be much smaller in scale compared to other machine learning methods that are conventionally used on such problems. To our knowledge, this represents the first time that relatively small logic circuits have been evolved to reach this level of performance on the automated recognition of handwriting, and promises new approaches to the integration of evolutionary algorithms and intelligent systems.