The benefit of a neural net and reinforcement learning is that genes don't have to encode the connections. They can instead encode the reward function(s), and some of this training can occur in the womb, this is probably why babies 'kick' while gestating.
You could imagine a hierarchy that goes something like this:
- Reward neurons that successfully transmit signal from brain to muscle fiber
- Reward neuron groups that get co-located muscle fibers to fire together.
Now the next levels can operate on muscles as a group (calf/hamstring) rather than individual fibers:
- Reward neuron groups that sense balance
- Reward neuron groups that move a muscle in response to balance output.
Believing that genes have to operate at the level of individual neuron connections to shape behavior just doesn't make sense.
You could imagine a hierarchy that goes something like this:
- Reward neurons that successfully transmit signal from brain to muscle fiber
- Reward neuron groups that get co-located muscle fibers to fire together.
Now the next levels can operate on muscles as a group (calf/hamstring) rather than individual fibers:
- Reward neuron groups that sense balance
- Reward neuron groups that move a muscle in response to balance output.
Believing that genes have to operate at the level of individual neuron connections to shape behavior just doesn't make sense.