Three lab-grown human forebrain organoids helped a machine-learning system identify Braille letters more accurately than a software-only baseline in offline tests that artificially corrupted the classifier inputs. On clean samples, the organoid system and the direct decoder were nearly tied: 83% and 82% accuracy, respectively.
A robot sent touch patterns to living neural tissue
A University of Bristol team mounted an optical tactile sensor on a robotic arm and traced a board containing the 26 English Braille letters. They recorded each letter at five probe depths and repeated each depth 10 times, producing 1,300 sensor traces.
Software compressed each trace into eight spatial regions and translated touch intensity, timing and duration into electrical pulse patterns. Those pulses went to forebrain organoids sitting on a multielectrode array, a chip that delivers stimulation and records the tissue’s electrical activity. A support-vector machine (SVM), a conventional machine-learning classifier, then chose the letter from the recorded spike counts. The organoids supplied part of the signal processing; the SVM made the classification.
The advantage appeared when classifier inputs were corrupted
For the three-organoid system, the researchers combined eight recorded values from each organoid into one 24-value input to the SVM. In cross-validation across the 1,300 recorded responses, this system reached 83% accuracy on the clean condition. The same classifier applied directly to a 32-value description of the electrical stimulation reached 82%.
The researchers then added four kinds of artificial corruption to the classifier inputs. The organoid system scored 83% under Gaussian noise, 82% under uniform noise, 72% when selected values were removed, and 68% when selected values were turned into outliers. The direct decoder scored 55%, 67%, 59% and 65% in those same tests. These were offline stress tests on encoded features, not trials in which a sensor or robot encountered real-world damage.
The experiment did not show organoids learning Braille
The result is a test of a fixed pipeline: a sensor records a letter, software maps it to electrical stimulation, an organoid responds, and an SVM assigns a label. The researchers say they did not demonstrate learning or long-term changes in the organoids. The task also covered a fixed board of 26 letters, not continuous reading or an operating device.
Sources
- Original research paper in Frontiers in Artificial Intelligence
- University of Bristol research summary
Image: Figure 2 by Tianyi Liu, Hemma Philamore and Benjamin Ward-Cherrier, reproduced under CC BY 4.0; original paper.