Recent work with living human brain cells has demonstrated the potential of merging living tissue with a computer. A colony of living neurons, known as Brainoware, trained faster than artificial models, achieving almost the same results. Despite ethical considerations, living human brain cells may surpass current and future silicon-based neural networks in terms of performance and cost-effectiveness.
The Making of Brainoware
Using stem cells, scientists cultivated a brain organoid – a three-dimensional colony of cells that mimic the structure and connections of neurons in the brain. This is not the first experiment with human-derived living cells. In previous studies, a brain organoid was taught to play Pong successfully. The challenge in such studies is to transmit information to the “brain” and interpret it.
A group led by Professor Guo Fen from Indiana University in Bloomington, USA, proposed a simple solution – they grew the organoid on a high-density array of electrodes. These electrodes, essentially a computer interface, input data into the “brain” cells and read the subsequent activity. This approach effectively implemented a spike (pulse) neural network architecture as a reservoir one. Although scientists did not know what was happening within the array of neurons, the quasi-living model demonstrated a quick learning and calculation ability.
Training and Performance of Brainoware
Brainoware was trained over two days on a set of 240 audio recordings of eight Japanese men pronouncing vowel sounds. Following this, it was able to recognize a specific voice with 78% accuracy. The system was also capable of solving equations using Henot maps with approximately the same accuracy. This required an additional four days of training. Moreover, Brainoware solved differential equations with greater accuracy than an artificial neural network without a block of short-term memory elements.
The Future of Biocomputing Systems
The living artificial “brain” was not as accurate as artificial neural networks with a long chain of short-term memory elements. However, each of these networks underwent 50 stages of training. Brainoware achieved almost the same results in less than 10% of the training time spent on training artificial circuits.
The authors of the work, published in the journal Nature Electronics, envision that it may be decades before universal biocomputing systems are created. However, this research is likely to provide fundamental insights into the mechanisms of learning, neural development, and the cognitive consequences of neurodegenerative diseases.