It’s time to pay at the cashier, so you take your wallet out of your pocket, slide out a credit card and slot the card into a card reader. These motions are simple for most people, but in-hand object manipulation – such as rotating a credit card with your fingers – requires a large-scale brain network, making it extraordinarily difficult for artificial hands. Enter the Human Brain Project, a ten-year supercomputing project aimed at advancing brain-related research – and which now, in its final phase, has taken up the study and supercomputer simulation of in-hand manipulation processes.
To learn more about in-hand manipulation, the researchers employed a biologically constrained recurrent convolutional neural network (RCNN) through EBRAINS (the Human Brain Project’s digital research infrastructure for neuroscience), using an RCNN for its ability to understand sequences of states – a major boon for robotics work.
Using the Human Brain Project’s Multilevel Human Brain Atlas, the researchers fed the constraints into the RCNN and trained the model on the 21.2 Linpack petaflops Piz Daint supercomputer at the Swiss National Supercomputing Center (CSCS). “The RCNN requires hours of simulated experience to train,” said Mario Senden, one of the lead scientists on the project and an assistant professor of vision and cognitive neuroscience at Maastricht University, in an interview with the Human Brain Project’s Evan Hancock. “The EBRAINS high-performance computing infrastructure allows us to speed up the process immensely and train the network in just a few minutes.”
The trained network, tasked with running simulations of a hand, successfully touched its fingers to its thumb after a couple of thousand learning runs.
Now, the researchers are aiming higher, trying to get the network to manipulate objects within that simulated hand. “The unique combination of interoperable services on EBRAINS will allow us to break new ground in cognitive computational neuroscience,” said Paweł Świeboda, CEO of EBRAINS and director general of the Human Brain Project.
Once the neural network- and HPC-powered simulations display sufficient aptitude, the plan is to begin incorporating a third element of the Human Brain Project: robotics work, and specifically the Shadowhand robot that is the subject of the current simulations.
“The workflow … incorporates a range of complementary EBRAINS services, including the multilevel human brain atlas, high-performance computing infrastructure, and, eventually, the Neurorobotics Platform,” said Katrin Amunts, the scientific director of the Human Brain Project. “EBRAINS is the only research infrastructure in Europe that offers this powerful combination of services.”
To learn more, read the coverage from the Human Brain Project’s Evan Hancock here.