Lab-Grown Brain Cells Help A Computer Identify Braille By Touch

Living Human Brain Cells Are Being Tested as Computer Parts

A system built around lab-grown human brain cells sorted Braille letters correctly 83% of the time, but only when three tiny clusters of those cells worked together. A single cluster averaged 61%. That beats random guessing, which would land on the right letter about 4% of the time across 26 options, yet it still missed roughly four of every ten letters.

Researchers at the University of Bristol in the United Kingdom built the setup and described it in Frontiers in Artificial Intelligence. A robotic arm slid a touch sensor across raised Braille letters, and the readings were converted into electrical pulses sent to the cells through a remote cloud platform. Cells did not recognize the letters on their own. A standard computer program studied how the cells fired in response, and that program made the final call on each letter.

Scientists are testing brain organoids, clumps of human brain cells grown in a lab, as possible computer parts because the human brain handles huge workloads on roughly 20 watts, about what a small light bulb draws. Each organoid is less than a millimeter across and holds about 10,000 neurons, according to the university. Silicon chips pushed toward brain-level ability are projected to need far more energy, so researchers want to know whether living tissue can carry useful information at all.

Brain Organoids Received Touch as Electrical Pulses

Organoids in the study came from a commercial supplier called FinalSpark, whose cloud platform lets scientists send pulses to living tissue and record its responses from anywhere. Every organoid sits on eight electrodes, tiny metal contacts that deliver the pulses and pick up the cells’ electrical activity. The tissue can stay active for up to about 100 days.

Early tests showed that more pulses drew more activity and that weak pulses rarely drew any, so the team built its touch-to-electricity translation around those rules. For the Braille test, a camera-based touch sensor called Evetac, mounted on a robotic arm, slid across a 3D-printed board holding all 26 English Braille letters, each a distinct pattern of raised dots, so every answer could be scored right or wrong. The arm pressed at five different depths and made ten passes per depth for each letter, producing 1,300 recordings. Each recording was split into eight regions to match the eight electrodes. Four features of the touch in each region, including how intense it was and when it peaked, set the number, length, timing, and strength of the pulses delivered.

Cell responses were counted at each electrode over about half a second. A support vector machine, a common computer program that learns to match patterns with labels, then guessed the letter.

Three Brain Organoids Beat One in Braille Accuracy

Individual organoids ranged from roughly 57% to 67% accuracy. Pooling the responses of three organoids on the same electrode board produced the 83%, a gain of 22 percentage points over the one-organoid average.

Why the trio worked better is still a hypothesis. Researchers suspect each organoid responds a little differently to the same pulses, so three together hand the program more distinct clues, much like three witnesses who describe the same event in slightly different ways. Testing that idea would require measurements the study did not include.

Brain Organoid Teams Held Up Better Against Messy Data

Researchers then deliberately corrupted the recorded cell responses, after stimulation, in half of the readings from each organoid’s eight electrodes. The corruptions included random static, blanked-out entries, and some readings tripled to mimic faulty equipment. A single organoid lost as much as 26.5% of its starting accuracy, depending on the type of corruption. The three-organoid team’s worst loss was 18.1%.

A further check skipped the living tissue altogether. The same program was fed the electrical pulse patterns directly. On clean data, that shortcut scored 82%, essentially tied with the 83% from three organoids. Under corruption, the three-organoid team came out ahead in all four types, including 83% versus 55% with random noise.

A single organoid trailed the shortcut in three of four corruption types, dropping to 42% with missing data against the shortcut’s 59%. The authors take the results as evidence that the tissue added something useful once organoids were combined.

Energy Savings From Brain Organoid Computing Remain Unmeasured

First author Tianyi Liu said in a university news release that the team “did not demonstrate learning or long-term changes in the organoids.” The setup was open-loop, meaning the cells received no feedback on their answers and had no chance to improve. Power use was never measured, so the energy appeal behind organoid computing remains a motivation rather than a result.

According to Techlicious, Liu said the team next wants to try textures and edges, then give organoids feedback to see whether they can adapt. Only a handful of organoids were tested, so this is a first look, not a verdict. For now, the 83% belongs to cells plus software. As a first step, it is credible. As proof of brain-powered computing, it is not there yet.

---

Disclaimer: This article is based on peer-reviewed scientific research but is intended for informational purposes only and is not intended to serve as medical advice, diagnosis, or treatment. Always consult a licensed physician or qualified healthcare provider with any questions regarding a medical condition or before beginning a new health or fitness regimen.

---