Learning

• The system learns by changing the weights of the synapses.
• Hebbian learning. If one neuron causes another to fire frequently, its weight is increased.
• That's the way learning is really done.
• It's entirely local.
• The current popular algorithm is spike time dependent plasticity (STDP) (Bi and Poo). If the presynaptic neuron fires right before the post- synaptic neuron, the weight is increased; if this is reversed, the weight is decreased.
• We've also done stuff with the total weight leaving or entering a neuron. This compensatory learning rule keeps the totals in a target area.
• (You can also learn by adding new synapses, or removing them, or by adding or removing new neurons with synapses. This isn't done very often.)
• Synaptic growth and death
• Neural growth and death