[Korrelan]

If you reverse engineer the human ocular system (ignoring focal distance and depth of field, etc) this is an approximate procedural/ algorithmic representation of the (colour) retinal schema projected onto V1.

Multiple interpolated (receptive field ranges) resolutions are recognised at the same time through a polar centric schema, a low resolution periphery graded to a high resolution fovea. So for any high resolution information there is also a low resolution spatial element of the surrounding overall area/ scene. The schema of the retina provides a top down influence for free, so… if the rough outline is cat shaped and fovea detects feline’s eyes… it’s probably a cat.

The small image (bottom left of left window) represents the whole of the circular/ retina, spatially arranged into a Cartesian format which is roughly analogous to the V1 spatial map. Rotational invariance simply moves the overall image detail left or right, scale invariance moves the image detail up/ down.

Eye movements and saccades are learned/ made relative to this polar map, so for any unrecognised object/ patch in the periphery the fovea can instantly traverse to focus on that region.

For textures or repeating patterns like fur, the center high res fovea recognises the texture, and all other joined/ un-bordered retinal locations with the same/ similar colour/ contrast attributes are assumed to be the same material, show by the colour fill.

https://youtu.be/m6Pw7qugpKE

If this retinal modality/ output is then translated into neural code, even foetal models can instantly recognise simple objects after one exposure.

https://youtu.be/uZYO6lCPQYs

The extra spikes for some objects on the pattern lock are caused by extreme similarities.  The system has noticed (from its point of view, not ours) similarities in composition, colour, etc.  Prior to being shown these objects the system had learned to ‘see’ from scratch, it learned to differentiate and categorise the visual sensory stream, its then able to recognise new objects from memory.  The spikes are an accumulation of episodic memories firing that encode/ remember the salient qualities of each object.



 

[Korrelan]

Closing post on project thread… My AI-Dream… has become an AI-Reality…

Thanks all for following my project, AI-Dreams has been my ‘lounge’, a place where I could relax, reflect and contemplate away from the usual harshness of the internet, both the forum and the good natured members have provided the environment I required, and helped me to bring my project to fruition... in the coming years… keep an eye out for the ‘blue K’.

Thanks and cheers.

 

https://www.youtube.com/user/korrelan

End

[Korrelan]

OK... Thanks for the hard work keeping the forum going... I guess I'll keep posting

This is the equivalent of learning 80 'words' 5000 characters long... in 3 minutes.

No big deal for a modern PC but... this very small section of my models neocortex can recognise any of the complete 80 * 5000 instantly, or any length sub-section of the 5k pattern, at any location within the 5k wide constant stream within 0.002 sec. It's limited to 80 patterns for testing (7K neurons and 30K synapse).

It shows my optimal dendrite branching modality so far, growing from scratch... AI savant mode lol.

https://youtu.be/oRLhtwvMWUE

 

[Korrelan]

Its been brought to my attention that some peers find my videos confusing and are not understanding/ reading the ‘Pattern Lock’ graphs correctly.  So I’ve done some redesigning, hopefully making them clearer. The new graph can be seen in the video below, and the following paragraphs are the description/ explanation I intend to use…

The ‘Pattern Lock’ graph shows both the input and the output/ best guess.  There are 80 colour coded inputs/ patterns (left to right) and the small white square moving along the scale indicates the current number of the input being injected into the model.  The moving green rectangle shows the models output/ best guess for each input pattern and should ideally match/ line-up with the input.  The height of the peak shows the confidence in the pattern match.

Each input represents 5000 human faces/ parameters.

https://youtu.be/UcUY0qlwGhc

Is this clear? Any suggestions?

 

[Hopefully Something]

The general idea makes sense. One question, how many unique pieces are in a column of 5000 pieces so that the net can recognize any length subsection?

[Korrelan]

To show the graph peak clearly, for this demo (above) I've tried to make all the patterns as unique as possible, so no face set is duplicated, though it is finding some similarities, shown by the low response/ confidence peaks.

The model can learn to recognise any length pattern/ sub pattern embedded in any length stream.

Normally pattern similarities in pattern groups show up as extra peaks, like in this object recognition demo, the extra peaks show its finding a commonality with another object, could be the general shape/ outline, colour, etc.  Each object is tuned to a specific location along the graph.

https://youtu.be/uZYO6lCPQYs

This brings up a good point actually... the green rectangle is just showing the strongest pattern response, on some of my vids the graph is showing features not single/ specific recognition... this needs more thought.