Evolve an eye and a brain together, in simulation, and keep the designs that work.
Generate Propose a visual system from two primitives: artificial photoreceptors for the eye and artificial neurons for the brain.1, 2
Simulate Place the agent in a physics‑based world. It sees through its own eye and learns a behavior: navigate, track prey, tell food from poison.
Select Score each eye–brain pair on the task and on size, weight, power and cost. A search method3 proposes the next designs.
the fittest designs seed the next generation
1 In nature, eyes and brains co‑evolved over millions of years. Simulation runs the same loop in parallel, in days.
2 The design space is written as a grammar of parts and rules: a lens may sit in front of a sensor, never behind it.
3 Evolutionary search, reinforcement learning, gradient descent or a generative model.
For vision science: rerun evolution
patch
cup
pinhole
lens
Starting from one light‑sensitive cell, agents that navigate evolve compound‑type eyes; agents that tell food from poison evolve camera‑type eyes. Ask what nature never tried.
For engineering: invent new sensors
Self‑driving carcamera rig + perception network
Microscopeoptics + image reconstruction
Drone in a foresttime‑of‑flight sensor + flight policy
Each sensor is designed with the network that reads it.