Abstract
Organisms perceive their environment and respond. The origin of perception-response traits presents a puzzle. Perception provides no value without response. Response requires perception. Recent advances in machine learning may provide a solution. A randomly connected network creates a reservoir of perceptive information about the recent history of environmental states. In each time step, a relatively small number of inputs drives the dynamics of the relatively large network. Over time, the internal network states retain memory of past inputs. To achieve a functional response to past states or to predict future states, a system must learn only how to match states of the reservoir to the target response. Basic regression can often achieve that goal. In the same way, a random biochemical or neural network of an organism can provide an initial perceptive basis. With a solution for one side of the two-step perception-response challenge, evolving an adaptive response may not be so difficult. Three broader themes emerge. First, organisms may often achieve precise traits from sloppy components. Second, evolutionary puzzles often follow the same outlines as the challenges of machine learning. In each case, the basic problem is how to learn, either by artificial computational methods or by natural selection. Third, the network connectivity of complex reservoirs may vary between individuals, requiring each individual to learn its appropriate response, a sort of primitive critical learning period.
Publisher
Cold Spring Harbor Laboratory
Reference21 articles.
1. Real-Time Computing Without Stable States: A New Framework for Neural Computation Based on Perturbations
2. Goodfellow, I. , Bengio, Y. & Courville, A. 2016. Deep Learning. Cambridge, MA: MIT Press.
3. Echo state network
4. Next generation reservoir computing;Nature Communications,2021
5. Hands-on reservoir computing: a tutorial for practical implementation;Neuromorphic Computing and Engineering,2022