Proteins in cell must change their connections (morph) to learn new tasks
This note is currently a stub.
References:
Bray, Wetware (2009), p. 121
What would be the corresponding training regimen for a cellp To learn a new task, a network of proteins has to change its connections. That is, it must alter binding strengths- change the rates of catalysis, modify the efficiencies of regulation- all in response to external influ-ences. How could this happen? This is indeed a rucial question and one that goes to the heart of the notion of wetware. The answer de-pends, I will argue, on the ability of the molecules of living organisms to change according to their surroundings. They change their chemical structure—I will use the term morph-so as to facilitate the survival of the organism. Familiar examples of this process occur in the brain, as you store new memories and learn new tasks, and in the immune sys-tem, when your body develops antibodies to resist the onslaught of a new infectious organism.
But the most far-reaching and important mechanism that cells use to train networks of proteins is evolution. The performance of a network depends on the processing capacities of its protein elements-the way they combine inputs, how much weight they give to particular connec-tions. These in turn depend on protein structure, specified by sequences of nucleotide bases in DNA. The seed corn of evolution is the continual variation in base sequences produced by mutation and rearrangements of the genetic material. From this long-term viewpoint, each learning trial corresponds to a single generation; each "test" is the ability of the organism containing the new network to survive and reproduce.
The notion that protein networks might evolve seems strange at first. Most people are familiar with the idea that individual genes can change with time. But I am now asking sets of genes to change at the same time, to work as a group in order to produce an improved collective perfor-mance. But in fact, similar strategies are widely employed for computer-based neural networks, and they can be highly effective at training.