Artificial intelligence research has focused on simulating the structure of the human brain. These solutions are limited by the difficulty of imitating the complex functioning of neurons and their connections. In response to this observation, American scientists designed an electronic system directly involving brain cells, so the principleorganoid intelligence.
Brain organoids: mini-brains
It’s in the magazine Natural electronics that American scientists have published a study that shows that it is possible to use brain cells as specialized electronic components in artificial intelligence.
It’s not a brain-to-computer connection like v Nut (film released in 1999, directed by the Wachowski sisters) and still less to obtain it from human beings as in Psycho-Pass (a Japanese dystopian animated series, released in 2012, in which the police are controlled by a computer that turns out to be a collection of interconnected brains collected from inmates on death row). Instead, researchers turned to a less cruel solution, which the scientific world calls organoids. These are small clusters of cells obtained from stem cells, which are then “programmed” to obtain the desired cell type. In their study, the researchers used brain organoids, clusters of cortical cells, which they then connected to conventional electronics to send an input signal and obtain an output signal.
These organoids allow them to have a system that functions like a mini-brain. They provide neurons with input signals by external electrical stimulation (transmitted by a set of electrodes). The organelle’s interneuronal connections process the signal and provide a visible output signal through the electrical activity of certain neurons. The organoid “learns” without supervision, much like the human brain. Connections between neurons weaken or increase, and relatively rarely connections are made. This development takes place in the human brain and is to some extent mimicked by neural networks commonly used in artificial intelligence.
Therefore, one of the main problems is to interpret the electrical output signal from the organoid. This signal is not directly explicit and therefore not directly interpretable. To do this, scientists provided data to regression programs, the simplest models machine learning the ability to perform model classification or modification. It was thanks to these examples that scientists could experiment with their system on specific tasks.
Read alsoCONVERSATION. “The question of moral status arises for human embryoids”
Effective voice recognition
To demonstrate the applicability of this new technology, the researchers applied their system to two specific tasks: speech recognition and the prediction of chaotic nonlinear equations (equations with quasi-random behavior, difficult to predict). Speech recognition was tested on five organoids during three training phases. During these learning phases, the organoids go from being between 55% and 70% accurate to over 75% accurate. Thus, their system requires a low number of training sessions to obtain relatively convincing results, which is a clear advantage over traditional artificial intelligence.
This learning speed is all the more visible in the case of predicting chaotic nonlinear equations, for which the researchers also used a more classical artificial neural network algorithm to have a point of comparison. Organoids only need four learning cycles to reach 80% accuracy. For artificial neural network algorithms, it’s a completely different story: it will take 50 training phases and the addition of “long short-term memory” (the ability to keep information from distant phases in memory). in the past, but without preserving all states of the neural network).
Read alsoTowards organisms on a chip
This study is promising and shows that organoid intelligence is a possible way to improve artificial intelligence. The researchers specify that their system faces several limitations, starting with the absence of vascularization of the organoids, which leads to premature death of the cells in the cluster. Another current limitation of cerebral organoids is the lack of consistency in their formation: with similar parameters during their development, two organoids will have very different shapes.