MIT network wires itself up and learns without AI's usual training trick

Modern AI systems learn through a method called backpropagation, which sends error signals backward through a network to adjust its connections. Neuroscientists doubt the brain works that way.

Researchers at MIT's McGovern Institute report in PNAS a network that learns without a designer laying out its structure, according to MIT News. The team, which includes professors Tomaso Poggio and Mark Harnett, argues that this kind of self-organization is a basic test of whether an artificial network could resemble a biological one.

How it works

The network starts as a jumble of randomly connected, neuron-like units. Each unit follows the same simple rule, a version of the old idea that neurons that fire together wire together: when two connected units are active at once, their link grows stronger.

As it trained, the network formed small repeating patterns of connections that served as building blocks for working circuits. Those patterns reorganized in a way that carried out gradient descent, the error-reducing process at the heart of most AI, without backpropagation.

Results

• The network learned to recognize patterns and sort images.
• Its performance was at a similar level to backpropagation-trained networks, the team said.

Limits

The MIT article reports no accuracy figures, network sizes or datasets. The tests covered only image and pattern tasks. Co-first author Qianli Liao calls the work a high-level starting point, not proof that the brain learns this way. Whether real synapses form such patterns has not been shown.

The researchers say the work could eventually inform both brain science and machine learning.