Jethro's Braindump

A critique of pure learning and what artificial neural networks can learn from animal brains

The Genomic Bottleneck

The compression into the genome whatever innate processes are captured by evolution. This acts as a regularizing constraint on the rules for wiring up the brain.

In large and sparsely connected brains, most of the information in the genome has to be allocated to specify the non-zero elements of the connection matrix in the brain, rather than their precise values. Even if every nucleotide of the human genome is devoted to specifying brain connections, the information capacity would still be at least six orders of magnitude too small.

The implication of this is that the genome does not encode the connections directly, but rules in forming these connections. Evolution acts on the brain only indirectly through the genome.

What this means for ANNs

  1. There may be an outer-loop (evolution) that optimizes learning mechanisms, and an inner-loop that allows us to learn inductive biases quickly (i.e. Meta Learning)
  2. ANNs should attempt as much as possible to build on solutions to related problems (transfer learning)
  3. Wiring rules and topology should be studies as a target for optimization in artificial neural systems

Links to this note