HNTE - CH. 2

[warlokkz]

Introduction to Neural Networks

[warlokkz]

differences between biological and artificial neurons(neurodes) are mainly in processing signals and interactivity between the neurons which is what makes neural networks so powerful.

Controlling signal processing through amplitude instead of frequency lets us treat input/output signals as vectors, which simplifies modifying synaptic weights or throttling memory loops in recurrent NNs.

The objective of what you're trying to solve with a neural network dictates the style of learning that the network needs to adhere to - supervised for static i/o problems, and the more powerful type of learning, unsupervised learning which can adapt to any unstructured data. Although, having a diverse network of different types of evolving neurons can help the network to bootstrap past the first hurdles in a problem, and give it the plasticity it needs to adapt and solve different problems.

Unsupervised learning seeks to improve the plasticity of individual neurons for the betterment of the network, through different learning methods like specialization (neuromodulation), cooperation (hebbian & modulation), or resource/scarcity(competitive & SOM).

Choosing the specific methods/modules to use in the network is what an implementation of Neuroevolution is meant for.

[mudphone]

I think it's interesting an potentially important that the author mentions that biological systems utilize frequency encoded signals. Whereas most NNs use amplitude encoded signals. I wonder if we should take a look at "spiking neural networks" since they are a closer analog.

In order to learn more about these "spikes" I might take a look at the book Spikes: Exploring the Neural Code (Computational Neuroscience) recommended by Rich Hickey, in one of his talks.