The 1000 Brain Theory of Intelligance

Authors proposed a method describing how brain might possibly work in perceiving the external environment
Authors make a claim that entire neocortex region of the brain is full or grid cells like neurons
Rather than learning one model brain tries to formulates 1000’s of models and make decisions based on deductive reasoning
Our brain is not a single brain it behaviors as if 1000’s of brain working simultaneously

Neocortex is the region in the cerebral cortex, which is related to perception and auditory signals
Grid cells are group of neurons whose simultaneous firing signifies tries to model two-dimensional external environment. Grid cells plays a major role in perceiving 3D space around us and also plays an important role in navigation
Recent studies based in fMRI analysis shows that neocortex is formed by the group of grid cells
In this paper the authors propose grid cell based technique to learn structure of objects and external environment

According to many theories causality and long-standing view in neocortex is the main reason for making decisions, but thousand brains theory is very different from them
According to thousand brain theory, the brain creates 1000s of models of an environment and decides based on deductive reasoning. Based on sensory signals are combined with the grid cell derived location, and long reange cortical connections work together and quickly identify objects
Authors have provided an example of coffee cup in the paper, which demonstrates the entire effect of the proposed theory

All the existing loss functions (in deep learning) optimizes variables for feature selection, but optimization for feature deduction might open up new possibilities, if it’s actually followed in brain
If we convert this problem to credit assignment problem, this proposed model can be formulated as multiagent reinforcement learning problem. Building a multi-agent framework where each and every agent is trained on different sensory feedback by collecting the experiences in one single buffer and have some attention mechanism to correlate and learn from experiences (something similar to DNC) and taking final decisions based on reward of each network would somewhat simulate the situation