Three human organoids helped identify Braille letters with 83% accuracy in an experiment using a fixed encoding scheme
Three human organoids helped identify Braille letters with 83% accuracy in an experiment using a fixed encoding scheme
On 16 October, researchers at the University of Bristol published a paper on an interface between a tactile sensor and human forebrain organoids. A robot scanned raised Braille dots, software converted the sensor signal into electrical pulses, and an external classifier identified the letter from spikes in the tissue’s activity. Using responses from three organoids, the classifier correctly identified 83% of the 26 letters. Under four types of artificially introduced disruption, this approach outperformed direct decoding of the input pulses.
An organoid is a three-dimensional piece of neural tissue grown from human cells. It was placed on an eight-electrode array that delivers pulses and records spikes in activity. Because the tactile sensor produces a much more detailed signal, the authors divided its output into eight zones. For each zone, the software used the event count, contact duration and timing, and change in signal strength to set the pulse parameters for one electrode.
The team first measured separately how pulse count, strength, duration, and delay affected spiking activity. These measurements allowed them to select ranges that produced distinguishable tissue responses. The robot then moved the sensor across 26 Braille letters at five contact depths, collecting 1 300 recordings.
The classifier counted spikes between 10 and 510 milliseconds after stimulation. Responses from one organoid yielded about 57% accuracy. Combining responses from three organoids into 24 values, eight from each, increased accuracy to 83%. This result reflects the entire processing chain: the organoids received an already prepared encoding of the letter.
A direct classifier given the 32 electrical stimulation parameters achieved 82% accuracy on clean data. The authors then introduced Gaussian noise, uniform noise, missing values, or outliers into selected channels. Under these four types of distortion, the direct classifier achieved 55%, 67%, 59%, and 65% accuracy, respectively; responses from the three organoids yielded 83%, 82%, 72%, and 68%.
In the experiment with a mouse organoid and a virtual pendulum, the tissue’s response changed the next stimulus, providing feedback that allowed researchers to test for learning. Here, information flowed in one direction: the sensor determined the pulse, the organoid responded, and the classifier identified the letter. For their next experiment, the authors envision a closed-loop system in which the organoid’s response changes the next stimulus, allowing them to measure how its responses change over time.