The use of living brains to perform computing tasks

In the cornerstone of science fiction literature, DUNE Frank Herbert’s (1965), humans ruled out the use of thinking machines due to a cruel war between artificial intelligences and humans, some 10,000 years before the events narrated in Denis Villeneuve’s film.

Instead of computers, in DUNE We are introduced to “mentats”, human beings capable of performing a large number of calculations that would put any modern supercomputer to shame. In parallel in the real world, during the early years of the space race, NASA used human calculators to verify the aerospace data used in the trajectory analyzes of the missions that sought to put man on the Moon. Unlike what happened in DUNE, these human calculators were replaced by the IBM 7090 computers.

The use of living brains to perform the computing tasks usually carried out by a CPU seems like something out of the horror and science fiction movies of the 70s, but it is exactly what recent research by Dr. Brett Kagan proposes. , principal investigator at Cortical Labs, an Australian company specialized in the design of interfaces between neurons and electronic boards.

The results of this study were published in the journal Neuron and the article presents the ““BrainDish”a neural network capable of learning to perform specific tasks.

He BrainDish is a culture of mouse neuronal cells that are located on an HD-MEA (High Density MicroElectrode Array) chip, a well-type plate that has a high density of electrodes capable of sending and receiving electrical information from the neuronal network. These electrodes are coated in platinum, an inert metal under physiological conditions, which allows these neurons to grow on the electrodes, form connections between them and retain their ability to generate action potentials spontaneously. This type of culture is known as a brain organoid.

The objective of this project was to demonstrate whether this neural network was capable of solving tasks “intelligently” by taking advantage of the electrical language shared by neurons. There are two conditions for this behavior to be considered intelligent. The first thing is that the system has to be able to perceive its external environment, to which it must then react in a particular way according to the sensory information it receives. This is analogous to when we walk down the street without colliding with everything in front of us. During the walk we make a conscious decision not to walk into a wall or a moving car, for example, based on what we can see, smell or hear.

To observe if there is “intelligence” in the decisions of the neural network, the researchers evaluated its performance when playing the PONG video game. The video game is based on an extremely simple version of table tennis where two vertical bars simulate being the paddles and a pixel that bounces between them simulates being the ball. The simplicity of this system allowed the company DeepBrain to develop the first “Deep Learning” algorithms used in the training of artificial intelligence.

So that the neural network can interact with the game, the surface of the HD-MEA chip was separated into different sections. A sensory section would tell the network where the ball is and two motor sections would drive the movement of the paddles. In order to differentiate the movement of the paddles, the motor area of ​​the crop was separated into two halves, one upper and one lower. When the activity of the upper hemisphere was greater than that of the lower one, the paddle would move upwards and the same in the opposite case to move it downwards. If the movement of the paddle results in the interception of the ball an electrical stimulus (75 mV per 100 ms) would be applied to all electrodes. In the opposite case, a different electrical stimulus (150 mV for 400 ms) would be applied randomly between the electrodes of the sensory section.

It is important to differentiate whether the behavior of the neural network is learning or if it is a series of lucky random events. To verify this, the principle of free energy for perception and action was used. This principle is based on the ability of living beings to plan actions that allow them to survive in their environment. In practical terms, it is a statistical model that allows us to differentiate whether the behavior of the BrainDish minimizes the free energy used to perform the action, in this way we can determine if the neural network learns to perform the task effectively or if it was a stroke of luck.

Based on this free energy principle, a series of tests were designed to demonstrate that this network of neurons is capable of performing a task and improving its performance as it repeats it. With this objective, the performance of the network was evaluated by modifying the feedback stimuli it receives when hitting or not hitting the ball with the paddles and the length of the game session. The first modification was a silent feedback test where the network did not receive an electrical stimulus when intercepting the ball, but the time in which this feedback would occur was maintained. The second modification was a no-feedback test, where both electrical and timing signals were suppressed and play continued even though the net failed to connect with the ball. During these tests, the number of times the net intercepted the ball was evaluated in sessions of between one and twenty minutes. As a control, a network of cells without electrical activity and the electrical noise of plates without cells, which is produced by unwanted interference in the electromagnetic environment, were used.

In all conditions, feedback, whether with electrical stimuli or silent, proved to be necessary for the network to perform the task successfully, accumulating a greater number of hits on the ball as the session time increased. while in the series of sessions where there was no feedback there was no difference with the performance of the control network. The results of these gaming sessions suggest that the network learns when it has positive feedback, while it learns nothing when it does not. This seems obvious to us, but it is very significant when evaluating the ability of these neural networks to perform specific tasks.

Currently, when talking about the development of artificial intelligences, it is impossible not to think about language models such as ChatGPT or Bard. These models are based on “neural networks” trained with a large amount of data and emulate how neurons used by living beings share information with each other. These models are capable of generating text responses from a query or prompt. Despite its versatility to generate everything from supermarket lists to software codes, ChatGPT generates only text and that is fine, because that is what it was designed for, but the method it uses, Machine Learning, can be applied to other things, such as generating original images from text inputs such as DALL-E or Stable Diffussion. In the end, these artificial intelligences are different tools to develop different tasks and are only limited by the computing capacity they have available or the restrictions imposed on them by the team of researchers that is developing it.

Although it is in its early stages of development, the possible applications of brain organoids, unlike synthetic artificial intelligences, do not have such well-defined limits, mainly because we do not yet know what the potential of this mass of cells is for learn, make intelligent decisions or if they are capable of developing a language. How complex should a brain organoid be before being considered a conscious living being? Should these “artificial intelligences” have some kind of rights? Are organoids capable of suffering? These types of questions are the same ones we ask ourselves today when using animals for the study of neuroscience or the development of new drugs, where we seek to reduce animal suffering in the development of new technologies.

There is an interactive version of this experiment at https://spikestream.corticallabs.com/.

Source: https://www.sciencedirect.com/science/article/pii/S0896627322008066?via%3Dihub

*This article arises from the agreement with the Interdisciplinary Neuroscience Center of the University of Valparaíso.