Emotion-driven Neural Network


So, no, this is not intended as a blog of my sad self, or like that. This is also supposed to be a technical blog, where I'm putting some interesting creations of mine.

The first of them is one of my earliest and still of interest: Emotion-driven Neural Network. It started just from the notion, that differently modulated memories are remembered differently. This definetely can't be implemented on a simple case (or not-so-simple) of any invented neural network kind, so there may be the need of adding something else, external…


We think. We learn. We forget. We carry on. The machine, however, seems to be unable to guess, what will be important in the future and what can be ignored safely. Most neural network implementations have a point of saturation, when no learning is possible and neural network becomes overfitting to results, unable to carry out its task on anything other, than learning samples. There were attempts at implementing algorithms, overcoming this effect <insert relevant articles here>, but they proved to be not very helpful and had lots of limitations (do I even know this?? check and recheck, find relevant references).

I propose the idea of adding an additional, special layer on top of the whole neural network, that will dynamically change learning curves for selected zones. For convinience and because of its actual similarity in action, we will call it emotion layer.


We will return soon, with the second part of the draft spec, where we will look at the details of what an emotion is, how it is controlled, and how we use it on the network (probably part 3 already x) )

PS For those curious, why it's v0.2, not v0.1: There really exists version 0.1, pretty short and vague, written for prof. Tereshko of UWS, still available for archive here