"Learning" is different for Artificial Neural Networks (ANN). An...

"Learning" is different for Artificial Neural Networks (ANN).

An intro to ANN that isn't too complicated to start with
http://www.ai-junkie.com/ann/evolved/nnt1.html
A great text to learn about ANN is"Fundamentals of Neural Networks" by Laurene Fausett

Roughly speaking:
In an Artificial Neural Network (ANN) multiple neurons in the first layer may be connected to multiple neurons in the next layer, with values being passed from each neuron in the first layer to each of the neurons in the next layer of neurons. If there are more layers then this is repeated to the next layer, and the next . . .

At each receiving neuron these incoming values are multiplied by weights to increase or decrease each Input value. All the Input values received by a neuron are then added up at that receiving neuron:
- if their total is greater than the Threshold/Trigger value of that receiving neuron, then that receiving neuron triggers and sends a value out from itself which may be passed to the next layer of neurons, or may be a final output value
- if the sum of the Input values is less than the threshold/trigger then no value is sent on from the receiving neuron to the next layer, or to final output

IBM TrueNorth does not modify weights, instead all that is done off-chip as corelets, which are then loaded into TrueNorth. This means a Neural Network on TrueNorth does not have the option to adapt and improve as time goes by

The BRN design does have the option to do on-chip adaptation

In Computing, Hardware and Software are interchangeable. Programs can be written as Software, or implemented as Hardware, ie RSA encryption software vs RSA encryption addon cards
- Hardware implementations can run faster
- Software implementations can be modified faster

So, this also speeds up the development of BRN ANNs as TrueNorth ANNs are only designed
using software, where BRN can also do it in hardware