Searching for minimal neural networks in fourier space

The principle of minimum description length suggests looking for the simplest network that works well on the training examples, where simplicity is measured by network description size based on a reasonable programming language for encoding networks. Previous work used an assembler-like universal network encoding language (NEL) and Speed Prior-based search (related to Levin's Universal Search) to quickly find low-complexity nets with excellent generalization performance. Here we define a more natural and often more practical NEL whose instructions are frequency domain coefficients. Frequency coefficients may get encoded by few bits, hence huge weight matrices may just be low-complexity superpositions of patterns computed by programs with few elementary instructions. On various benchmarks this weight matrix encoding greatly accelerates the search. The scheme was tested on pole-balancing, long-term dependency T-maze, and ball throwing. Some of the solutions turn out to be unexpectedly simple as they are computable by fairly short bit strings.