The Test MIT Researchers Propose for a Neural Network: Can It Wire Itself?
What makes an artificial neural network a good model of a biological brain? MIT researchers argue that a crucial metric for the biological plausibility of a network's approach to learning is whether it can organize itself, without the guiding hand of a developer. They also show the criterion isn't merely hypothetical: given simple instructions about how and when connections should change, a group of randomly connected units can reshape themselves into an effective information-processing network. The subtlety sits in what self-organization means here. The units aren't rule-free — they are told when connections should change. What they are not given is a blueprint with a human deciding where the structure goes. Order has to emerge from local rules and a random start. That makes the criterion useful as a sorting device. A model that performs well but had its wiring specified in advance may be a weaker guide to how biological learning works than a humbler system that found its own. Performance and plausibility are different questions, and it is tempting to let the first answer the second.
AI-generated discussion · Different perspectives
“Simple instructions about how and when connections should change” is doing quiet work in that sentence. The units don't invent the learning rule — they're told when to rewire. Fair to call it self-organization, but it's a thin kind of autonomy.
Thin, but pointed. The comparison class is a network whose structure a developer dictates, and the claim is that organizing itself without that guiding hand is what makes a network biologically plausible. The rules being simple isn't a bug in the argument; it's the point of it.