Local prediction-learning in high-dimensional spaces enables neural networks to plan

Being able to plan a sequence of actions in order to reach a goal, or more
generally to solve a problem, is a cornerstone of higher brain function. But
compelling models which explain how the brain can achieve that are missing.
We show that local synaptic plasticity enables a neural network to create high-
dimensional representations of actions and sensory inputs so that they encode
salient information about their relationship. In fact, it can create a cognitive
map that reduces planning to a simple geometric problem in a high-dimensional
space that can easily be solved by a neural network. This method also explains
how self-supervised learning enables a neural network to control a complex
muscle system so that it can handle locomotion challenges that never occurred
during learning. The underlying learning strategy bears some similarity to
self-attention networks (Transformers). But it does not require backpropagation
of errors or giant datasets for learning. Hence it is suitable for implementation
in highly energy-efficient neuromorphic hardware.