Duarte maps a narrow bridge between AI and neuroscience

Renato Duarte, an assistant researcher at the University of Coimbra's Center for Neuroscience and Cell Biology, ran a citation analysis across roughly 458 million publications in OpenAlex and found the connection between neuroscience and AI is 'real and gaining momentum' but 'narrow, tool-heavy and concentrated.' Writing in The Transmitter, he reports that the most recent of the bridge works tying the two fields together is from 2003.

The flow is lopsided. AI papers cite neuroscience more than the reverse, pulling from what Duarte calls a 'small, fixed canon of classic neuroscience papers.' Neuroscience's citations of AI have climbed since 2006 and accelerated from 2016 to 2019, but the shared bedrock has not refreshed.

About half the strongest interdisciplinary bridges are statistical and machine-learning tools, software packages like XGBoost and lme4. The other half are older neuroscience concepts: spike-timing-dependent plasticity work by Bi and Poo from 1998, Izhikevich's 2003 spiking neuron models, and studies of visual perception, attention and working memory. Named researchers whose work gets 'read across the borders' include Geoffrey Hinton, Yoshua Bengio, Karl Friston and Peter Dayan.

The bridge-builders are also concentrated. The top 100 institutions account for nearly 40% of the cross-field work, led by Alphabet and Microsoft, with MIT, UC Berkeley, Harvard and Stanford behind. Duarte's conclusion is flat: 'the astonishing recent achievements in engineering intelligence owe neuroscience surprisingly little in its actual machinery and operating principles.' Two of the researchers we follow had the piece in circulation the same week.