MIT neural network assembles its own learning machinery

MIT researchers have demonstrated a neural network that builds its own learning machinery from simple rules governing how connections change. Starting with layers of randomly connected neuron-like units, the system organized itself to recognize patterns and classify images, performing at a similar level to networks trained using back-propagation.

The network uses rules thought to help reshape biological brain circuits, including strengthening connections between neurons activated together. These interactions produce self-assembling motifs that act as building blocks for functional circuits. During learning, the motifs reorganize through an optimization strategy similar to gradient descent, which gradually adjusts connections in response to errors.

Researchers at MIT’s McGovern Institute for Brain Research and K. Lisa Yang Integrative Computational Neuroscience Center propose self-organization as a key measure of whether an AI model is biologically plausible. Conventional networks depend on learning algorithms designed by engineers, while the brain must organize its own connections and the machinery that adjusts them.

The findings do not establish that the brain learns through the same mechanism. Instead, they offer a starting point for investigating biological learning. Understanding these processes could support research into brain-computer interfaces, aging and disease, while potentially helping engineers develop more capable or human-like machine learning systems.