|
But the point is that it does not do that. A neural network does not work like a computer. It does not have to predict. It is a parallel flow from input to output AT ONCE. There is no "processing" like in a CPU where it takes n amount of CPU cycles and then the result is sent on. And as I said, it uses a proxy - it does not try to predict anything, it uses the data it has at that moment and nothing else. Before you get mad at me, do take some neuroscience courses please. I'm an IT guy myself and it opened a completely new world for me. Arguing with someone who only sees one side is frustrating. And while I'm not good enough to be able to explain the neuroscience - maybe not at all, definitely not in a forum comment - I still know a little bit about the subject. "Prediction" and "Looking ahead" may be system outcomes, but it does not actually happen as part of the actual low-level process. Not for the low-level processes like catching a flying object, I'm not talking about conscious thought processes. When a moving object leads to input from different retinal ganglion cells - always in the form of action potential frequencies (so, an analog signal despite an action potential being all-or-nothing, just an aside) through temporal summation timing differences - which can be a function of the speed the object is moving in the real world - can lead to different subsequent processing neurons being activated, eventually leading to different motor neurons being activated or the same ones firing at different rates. So the computation takes place with the signal flowing as a "wave" across brain regions, but it all takes place at once. There is no "let's calculate where this is going to be in a second". This is implicit by connecting input directly to output through paths that change in subtle ways depending on said input. Yes, the end result (system outcome) is a "prediction", but not in the same way as a computer would do it. It just "happens", there is no actual effort to predict anything. There also is no representation of such a "prediction" anywhere else: It flows right into your movement, but as somebody else has already pointed out just because you manage to catch the ball doesn't mean you are any good at consciously being able to make actual predictions. By the way, the processing already starts in the retina, which consists of several layers of cells, and the ganglion cells that communicate with the visual cortex at the very back of the head (after being relayed through the geniculate nucleus of the thalamus in the middle of the head). They don't provide a signal like a camera CPU gets from an RGB chip which simply 1:1 sends pixel values. You have cells signaling movement from left to right, others from right to left, etc., coming from the retina. I think the main point is that the entire process in a neural network is completely different from how a computer operates. When we name outcomes we may be tricked into thinking it's similar, but when we look at how the output is generated it is a completely different world. That does matter, it has implications for how we think about the whole thing, what we think we can achieve, and how. If you did this in a computer, imagine not using any storage - not even CPU cache. All data must be processed at once, there are no buffers, not even on an "input pin". You have a stream of data and all you can do is decide where to move it next. It's a horrible analogy but the best I can do right now. Oh, and you don't have a system clock signal, the data is the clock signal. And you don't do any calculations either as a microchip performs them, instead you rely on analog processing: temporal and spacial distribution of the electrical signal matter. For example, if you send a lot of small signals, since they are all actually ions entering the cell (the dendrites of a neuron) it takes time to throw them out again, and if before the ion transporters manage to do that a new signal arrives with more and more of them the amount of ions increases, possibly until reaching threshold (for action potential firing). Same over space: On a dendrite there are many synapses over its length, connected to different neurons (their axons). The charges (ions) can equally build up over space, not just time. So length of wiring matters as well as the shape of the electrical signal - two things we don't want to see having any influence in our microchips. So computation in a chip and in a neural network is vastly different. Computation in the network happens "on the fly" simply by the movement of the signal through the network, encoded as the frequency of an all-or-nothing signal (action potential), but then every analog trick there is is used to decide if and when an action potential fires in connected cells. Actually "storing" values happens over a longer period by changing the connections: New synaptic connections form all the time and existing ones disappear, and existing synapses change ion channel and ion transport channel densities. That is way too slow to have an impact for any given computation, so it plays no roll for trying to catch the ball that's in the air right now. |
and watching this: https://www.youtube.com/watch?v=HeQfO4byFhg