How Brains, Computers, and Simple Little Grids Can Remember Things

Getting your Trinity Audio player ready…

A plain-English tour of cellular automata, neural networks, and the “attractor-basin” idea from Frank Schmidt’s article “Memory as an Attractor Basin—Why Brains and Artificial Networks Remember by Falling, Not Filing.”

1. The Big Picture

Most of us learned in school that memory works like a filing cabinet: you put a fact in a folder, stick it in a drawer, and pull it out later. Modern neuroscience and machine-learning research say that picture is wrong.

A better mental image is a landscape full of valleys. Each valley represents one memory. When your brain (or a trained computer network) starts to “think,” its activity is like a marble rolling around on the hillsides. If the hills are shaped just right, the marble automatically drops into the correct valley, and—boom—there’s the memory.

Frank Schmidt’s article calls each valley an “attractor basin.” You don’t “file” a memory; you fall into it.

2. Two Very Different Tools for Shaping Valleys

Tool	What it really is	Why it matters to memory
Neural network (the kind that powers voice assistants and image generators)	Millions of adjustable connections that can be strengthened or weakened during training	Training carves the slopes of the landscape so information naturally rolls toward the correct valley
Cellular automaton (CA)	A grid of simple cells that flip on or off according to a tiny rule like “turn on if exactly two neighbors are on”	Even with a super-simple rule, whole patterns—like growing crystals or mini tornadoes—appear on their own. These patterns are also valleys in a landscape

3. Cellular Automata: The Power of Tiny Local Rules

Think of a soccer stadium doing “the wave.” Each fan just follows one rule: stand up if the person on your left just stood. No one coordinates it, yet a moving wave races around the arena.

A cellular automaton works the same way, except the “fans” are little squares in a grid on your computer screen. At each tick of the clock every square updates itself based only on its neighbors. Out of that mind-numbingly simple routine come surprisingly rich patterns—some freeze, some oscillate, some wander like living creatures. These end states are ready-made valleys for storing information.

4. Neural Networks: Heavy Machinery for Detailed Sculpting

If a CA is a crowd doing the wave, a neural network is more like a giant sound mixer with thousands of knobs. During training, an algorithm turns each knob so the final mix (the network’s output) matches the target song (the right answer). All that knob-turning literally reshapes the memory landscape—steepening one valley, merging two others, or flattening a bump that caused confusion.

Neural nets are incredibly flexible, but flexibility has costs:

Energy-hungry—big matrix math on GPUs drinks electricity.
Hard to interpret—it’s tough to peek inside and see why it made a choice.
Can forget fast—tune it on new data and old valleys can get paved over.

5. Marrying the Two: A Memory System That’s Cheap

and

Smart

Imagine stacking the two ideas:

Bottom layer: a cellular automaton. It supplies cheap, built-in valleys because its rule always funnels activity into certain stable patterns.
Top layer: a small neural network. It fine-tunes the valleys, carving detailed niches without using much extra energy.

This combo—call it a CA-Neural Hybrid—behaves like a landscape that has solid bedrock (from the CA) plus a topsoil you can reshuffle endlessly (via the neural network). The bedrock prevents catastrophic landslides (massive forgetting). The soil lets you plant new memories quickly.

6. How Would Such a Hybrid Actually Work?

A simple recipe:

Encode the data. Turn today’s photo, sentence, or sensor reading into a starting pattern on the CA grid—a bit like sprinkling seeds on a chessboard.
Let the CA evolve for a few steps. Local rules churn and settle into a coarse pattern—the first valley.
Send the CA’s state into a small neural network. The network tweaks certain grid cells or adds an extra read-out layer, sharpening the valley’s shape so it points right at the correct answer.
Read the result. The final pattern (plus the neural tweaks) tells you “cat,” “heatwave tomorrow,” or whatever the task demands.

Early lab tests (in small research projects) hint that hybrids:

Remember more patterns per storage unit than classic models.
Keep working even after you randomly knock out a chunk of their cells (fault-tolerance).
Use less power because the simple CA update can be done with tiny logic gates, not giant GPU arrays.

7. Why This Resonates with Real Brains

Brains already mix “cheap” and “expensive” machinery:

Cortical columns—tiny repeating circuits—resemble CA cells handing local chores.
Thalamo-cortical loops act like neural gates that decide which signals get boosted.
Bioelectric fields in growing tissues guide regeneration, much as a CA’s rule guides patterns.

So the hybrid idea isn’t science fiction; it mirrors nature’s layered approach—basic physics at the bottom, sophisticated learning on top.

8. Open Questions (Put Simply)

How do we find the best CA rule? There are zillions. Computer searches or “evolution” algorithms help, but the hunt is hard.
Can gradients flow through hundreds of CA steps? Math that trains neural nets can fade to zero when passing through long chains of simple steps.
Will hybrids misbehave in surprising ways? Extra valleys can be good (more memory) or bad (strange glitches). Engineers must watch out.

9. Why You Should Care

If you build smartphones or drones, hybrids might give you on-device learning without draining the battery.
If you study biology, they offer a sandbox to test ideas about how tissues or brains store information.
If you worry about AI safety, clear, layered landscapes are easier to inspect than black-box mega-networks.

10. Bottom Line

Frank Schmidt’s core insight—memory is falling, not filing—invites us to focus on how we shape the hills and valleys of thought. Cellular automata give us sturdy, self-organizing hillsides for free. Neural networks supply the artistry to etch fine detail. Put them together and we could get memory systems that are:

Robust (hard to break),
Efficient (easy on power), and
Understandable (you can actually see the valleys).

That blend just might be the next step in building machines—and perhaps someday therapies—that remember as gracefully as we do.