Cortical Labs researchers separated signals from human neuronal cultures from recording artifacts in a digit recognition task
Cortical Labs researchers separated signals from human neuronal cultures from recording artifacts in a digit recognition task
On August 16, a team from Cortical Labs, Tohoku University, and the University of Melbourne posted a preprint describing experiments with human neurons grown on a chip. Handwritten digits from the standard MNIST dataset were encoded as sequences of electrical pulses, after which software identified each digit from the culture's response. Once the device signal was excluded, the best mean accuracy, 47%, came from cerebral cortical neuron cultures grown in 60 connected microwells.
The authors converted each digit into five time steps of electrical pulses and delivered them through 16 channels on the chip. After the final pulse, 59 electrodes recorded the cells' response, and the software used this recording to identify the digit. The result was averaged across five presentations of the same digit.
The first control was a dish containing saline solution. When frequency analysis of the signal began at the same time as stimulation, the software identified the digit correctly in 44.4% of cases, even though the dish contained no neurons. The classifier was distinguishing the electrical trace left by the pulse in the recording equipment. In experiments with neuronal cultures, this analysis produced a mean accuracy of 62.5%. For the primary evaluation, the authors therefore analyzed 400 milliseconds of recording beginning 10 milliseconds after the final pulse.
Under this protocol, mean accuracy across all variants was 39.3%, compared with 12.3% for the shuffled control. Modular cortical cultures produced the best mean result, 47%, while the highest result among the group of 22 cultures was 66.7%. Thin walls in this chip separate 60 groups of cells, while narrow channels connect each microwell to its neighbors.
The method used to split the data also changed the second result. When recordings were divided into adjacent temporal segments, individual estimates reached 98%. This protocol tests whether part of a single temporal response can be reconstructed. When the data were split by independent presentations of the digit, the maximum was 35%. This protocol instead tests whether a digit can be recognized from a separate recording.
After controlling for the artifact and using independent presentations, the modular architecture retained the best mean result. In the other two settings, the classifier obtained information about the digit either from the pulse trace or from adjacent segments of the same recording.