A hybrid learning machine that combines the best features of neural networks and computers is described in a study published in Nature this week.
Conventional computers can process complex forms of data, but require manual programming to perform these tasks. Artificial neural networks have been developed to mimic brain-like learning that can identify patterns in data, but they lack the memory architectures needed for symbolic processing of structured data.
Alex Graves, Greg Wayne, Demis Hassabis and colleagues developed a so-called ‘differentiable neural computer’ (DNC), which comprises a neural network that can learn by example or through trial and error, and an external memory structure similar to random-access memory in a conventional computer. Thus, it can learn like a neural network but process complicated data like a computer.
The study shows the DNC can successfully understand graph structures like family trees or transport networks; for example, planning the best route on the London Underground without prior knowledge of this transport system or solving moving block puzzles with goals described in a symbolic language.
Epidemiology: Assessing non-pharmaceutical interventions in reducing COVID-19 spreadNature Human Behaviour
Epidemiology: A website to assess COVID-19 event risk in the US in real timeNature Human Behaviour