Profile cover photo
Profile photo
Korrelan AI
21 followers -
Seeking the essence of intelligence
Seeking the essence of intelligence

21 followers
About
Posts

Post has shared content
I’ve been busy updating my software suite, profiling, optimizing, etc, as well as working out the algorithms required for neurogenesis in my AGI. For this vid there are approx 19k neuron groups/ voxels and 142k axon group/ connections, the simulation is running on a single thread just to slow it down so we can see what’s happening.

Up to this point I’ve been generating the initial neuron columns/ voxel positions based on the Fibonacci sequence. This gives a nice even distribution over the models volume, and provides the initial skeletal framework/ layout required for generating the long range tracts/ axons required for passing the GTP facets between cortex areas.

The algorithms for neurogenesis are designed to move/ generate new neurons in to areas that require a higher resolution due to pattern complexity. A new neuron will migrate to its desired position, wait and listen to its neighbours, attach its dendrites/ synapse to the correct locations and then join in with/ enhance the local activity. This has greatly enhanced the learning abilities of the foetal model.

The vid shows how the foetal model, once trained on a specific type/ frequency of stimulus can now learn/ forget completely new/ novel variations with just one exposure.

At the end of the vid you clearly see the higher density of neurons on the stimulated right hemisphere.

Post has attachment
I’ve been busy updating my software suite, profiling, optimizing, etc, as well as working out the algorithms required for neurogenesis in my AGI. For this vid there are approx 19k neuron groups/ voxels and 142k axon group/ connections, the simulation is running on a single thread just to slow it down so we can see what’s happening.

Up to this point I’ve been generating the initial neuron columns/ voxel positions based on the Fibonacci sequence. This gives a nice even distribution over the models volume, and provides the initial skeletal framework/ layout required for generating the long range tracts/ axons required for passing the GTP facets between cortex areas.

The algorithms for neurogenesis are designed to move/ generate new neurons in to areas that require a higher resolution due to pattern complexity. A new neuron will migrate to its desired position, wait and listen to its neighbours, attach its dendrites/ synapse to the correct locations and then join in with/ enhance the local activity. This has greatly enhanced the learning abilities of the foetal model.

The vid shows how the foetal model, once trained on a specific type/ frequency of stimulus can now learn/ forget completely new/ novel variations with just one exposure.

At the end of the vid you clearly see the higher density of neurons on the stimulated right hemisphere.
Add a comment...

Post has attachment
The ideal shape/ volume I’ve found so far for the cortex connectome model is spherical but the four lobes per hemisphere, the separate hemispheres themselves and even the gyri/ sulci are still required, they all have a functional purpose within the model.

This is an experiment to roughly map my models connectome to the same area/ shape as the human cortex. Besides the procrastination element to this experiment, the idea is that it should aid understanding of the schema to the layman… and I think it looks cool lol.

Each voxel = one functional column (50 ish neurons, 1000 ish synapse) per 10.5K voxels.

Although even this fetal stage can learn hundreds of millions patterns I’m sticking to the usual 40 distinct colours for clarity. 40 patterns learnt * 250 facets = 10k pattern facets learned.

As usual the voxel colours match the bar graph lower left. Each colour represents one injected/ learned full pattern. As the patterns are injected into the model (top right) the height of the bars represent confidence in recognising that pattern. Ideally just one bar should rise for each pattern. If more than one bar rises then a similarity in the patterns exists.
Add a comment...

Post has shared content
Upgraded my bots cameras to HD, with 3D gimbals for easier stereo alignment.

Post has attachment
Add a comment...

Post has attachment
Q & A #1
Q & A #1
korrelan.blogspot.com
Add a comment...

Post has attachment
Attention Theory
Attention
Attention
korrelan.blogspot.com
Add a comment...

Post has attachment
Old Video - Training a multi layer network to recognise simple line orientations.
Add a comment...

Post has attachment
Machine Learning, Artificial Intelligence, Neuroscience & Robotics

Personal thoughts on the Sandman.
Add a comment...

Post has attachment
Project Update
Project Update
korrelan.blogspot.com
Add a comment...
Wait while more posts are being loaded