Brain Modules

Replicate brain modules' computations with Artificial Neural Networks

My approach to brain simulation

1. Identify a list of computational modules in the brain
2. For each module
   1. Understand what it computes: input -> output
   2. Get training data (synthetic or real-world)
   3. Train artificial neural networks to replicate its functionality
3. Combine modules

Why?

• Biological plausibility is a trap, simulating spikes and neurotransmitters does not help us understand how brain generates intelligence

• Analogy: considering transistor physics is irrelevant to understanding how a computer computes a + b -- they are on different isolated levels of abstractions, they do not depend on one another to work

• Current works in AI are mainly focused on solving daily-life tasks (text, image/video, game playing) -- I want to use these technologies to understand the brain

• It is well proven that ANNs can produce intelligence (LLMs, RL agents) -- making them qualified to model modules in the brain

List of implemented modules

1. Place Cells (incomplete, deprecated)
2. Brain's GPS System (Head Direction Cells, Grid Cells, Place Cells)

Minimalist RL

• Minimalist & Decoupled Reinforcement Learning

Gradient Free Optimizer

• Gradient-free optimization of neural network parameters

• Example: f(x) = x^2 (loss = log10(MSE(y_predict, x**2)))

• Genetic algorithm:

• AdamW (for comparison):