Imagine an AI whose brain is a 2D image. Every pixel carries values, memories, uncertainty, and rewards. That’s Neural Pixel AI System at DFAI Research.
Core concept
- Each agent has a neural texture: a pixel‐map where each pixel encodes Q-values, memory, confidence, inputs.
- Agents sense their world (hunger, obstacles, enemies) via input rows in their brain texture.
- They choose actions (move, wait, turn) based on reading their internal visual brain.
- After action, they update their brain texture per reward logic.
Evolution & reproduction
- When conditions are met (maturity, nourishment), agents reproduce.
- Offspring inherit brains via pixel‐wise averaging + mutation in structure or values.
- Genes include confidence scores, names, topology.
- Natural selection: weak agents die, strong ones persist.
Visualization & tools
- AI Observer Manager: monitor all active agents.
- Brain Viewer: inspect any agent’s neural texture, genealogy, stats.
- UI includes health bars, blinking eye effects, visual feedback.
Why it’s unique
- Behavior emerges purely from reward logic — no forced clamping or heuristics.
- Entirely visual, debuggable, evolvable brain structure.
- Combines reinforcement, mutation, neural memory in one system.