[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]
AGIHound @TrueAIHound on x 1692 followers
Created: 2025-07-22 23:44:27 UTC
Neuroscience bits from my research.
Pattern recognition and representation in the brain.
In a deep neural net, representations are stored in multiple layers as synaptic weights. A synapse is normally a connection between X neurons located in separate layers. An upper layer contains compositions of features from lower layers. During learning, weights are adjusted to optimize an objective function that controls an output node.
There is no function optimization in the brain. Why? It's because function optimization is the opposite of generalization. There is no hierarchical architecture either since object representation is compositional.
Cortical learning consists of identifying unique characteristics of an object. These are memory traces (recorded spiking events) that are unique to a particular object or a class of objects. The question is: how does the brain determine which traces are unique to a particular object?
My theory is that memory traces that survive the continual battle to control the sensory space are unique to specific patterns. Non-unique traces are automatically eliminated due to conflicts. No weight adjustments or backprop signals are needed. A trace either survives or it doesn't. If a future sensed pattern reactivates a unique memory trace, this causes the strength of its percept to increase. The strongest percept wins. Learning is XXX% driven by the external world.
Competitive pattern learning/recognition is fast and powerful. I anticipate that this approach will be several orders of magnitude more data-efficient than deep learning. At least, that's the theory.
I won't know if the theory is correct until I write code to test a computer model of it. It's more complicated than it sounds. I'm still struggling with my monochromatic retina model. It's a low-res model with only X edge orientations. Even generating random microsaccades is a pain in the butt. And I can't use vibe coding because the LLMs have no idea what I'm talking about. 😁
XXX engagements
