Artificial intelligence has reached a stage where neural networks can recognize images, process language, and even write code. Like brain circuits, they reorganize by adjusting connection strengths during learning. A paper published on Oct 9 in PNAS by researchers at MIT’s McGovern Institute and the K. Lisa Yang ICoN Center argues that a key metric for the biological plausibility of an artificial network is its ability to self‑organize without an external designer. The authors gave the network only very simple update rules—e.g., “neurons that fire together wire together,” meaning simultaneous activation strengthens the synapse—and let a collection of randomly connected units reshape itself while being trained on data. The self‑organizing network uses an optimization strategy akin to gradient descent to reduce error and eventually learns to classify images, achieving performance comparable to conventional back‑propagation networks. While back‑propagation is unlikely to be the brain’s actual learning algorithm, discovering analogous synaptic motifs could bridge artificial and natural learning mechanisms. Self‑organization is therefore treated as the first high‑level framework for assessing biological plausibility, paving the way for deeper investigations into brain‑like learning rules. Beyond neuroscience, these insights may help engineers build machine‑learning systems that operate more efficiently or in a more human‑like manner.
Review