Peter's second patent relies on the storage of the "training model" to conventional computers (copy of the abstract for patent application US 13/461,800 below).
What you do not realize is that in order to arrive at this weighted values the chip needs to go through logic.
Imagine playing ping pong. Brainchip can play ping pong. It will hit the ball as soon as it receives the ball on its paddle. But after the hit, how does it know whether it was a good hit or a bad hit. There needs to be some mechanism that can provide feedback to the system telling brainchip that it was a good hit.
What if you can plug in the API to tell the chip that it was a good hit or a bad hit ? And thats why it needs the API interface and possibly a BDK.
Thats where the weighted value come in to place as they become the learned resource.
Training models are scalable. Which means if you can perform the task of getting the brainchip to draw a line in 3 seconds it is reasonably assumed that it can continue to draw a line at a similar rate. Given enough learned resource it will exponentially continue to continue drawing straight lines as a function of available paper and pencil and with appropriate decay of time.
After one hour , it would reduce drawing the same line at 2 seconds and after 3 hours down to one second and so on and so forth...
Walls are just a physical representation of a physical element on a race track as is the car. Introducing hooks and turns would not matter to the basic function of running the car through the track because the learning mechanism is independent of the physical barriers.
Inputs (sensors) react to physical objects and hardware takes action based on inputs received. Whether it is a race track of complexity 3 or 100 it really does not matter as long as the chip is capable of storing enough learned resources which depends on its ability to scale and flow information both ways.
No because the inputs will receive no input for walls (physical object) and hence will continue as normal without waiting for a learned resource.
So will a brainchip car, only faster
It is tested with a race car algorithm that a global company uses today and takes 15 minute to complete.
Sensors whether good or bad do not have a influence on how quickly the task is performed. A sensor only passes a 1/0 signal for an instance and the speed at which this travels to the processor is dependent upon the resources available.
Think about it this way.
If you ask a human what is 5 x 4 . Most people will say 20 in a second.
Now what id you go find a calculator , press the switches and then get the results - takes about a minute.
Thats the difference between brainchip and software AI.
Hope that makes sense.
(20min delay)