In 2022, Australian researchers announced that they had played the game 'PONG' using nerve cells. PONG is a game in which players move a paddle up and down to hit a ball, located on the left side of the screen. The researchers cultivated 800,000 brain cells from humans and mice on electrodes while sending the position information of the ball as electrical signals, prompting the cells to move the paddle in response.
The researchers rewarded the brain cells with regular signals when the paddle hit the ball and punished them with irregular signals when it missed the ball. This method is the reinforcement learning technique used by Google DeepMind when developing the artificial intelligence (AI) AlphaGo, which defeated Go champion Lee Sedol. Instead of explaining a specific behavior to a puppy repeatedly, this training method rewards the dog with praise or treats when it performs the desired behavior. The brain cells adapted to the PONG game so well that they could rally with the ball correctly within 5 minutes of repeated reinforcement learning.
Research attempting to overcome the limitations of AI using bio intelligence derived from nerve cells is emerging. At the 'AI for Good Summit' held last month in Geneva, Switzerland, David Gracias, a professor of chemical and biomolecular engineering at Johns Hopkins University, presented 'Organoid Intelligence', a biological computer system that acts like AI using brain cells akin to those that played the PONG game. This event, which began in 2017 under the United Nations (UN) International Telecommunication Union (ITU), gathers the world's leading experts to discuss how AI can contribute to the public good.
◇Power-hungry starfish AI solved with nerve cells
Organoids is a neologism created by adding the suffix 'oid', which means resembling an organ. These are stem cells cultivated into three-dimensional structures similar to organs and are referred to as mini-organs. Professor Gracias made a bold statement that brain organoids can be utilized not only as tools for brain research or drug development but also as AI systems.
Thomas Hartung, a professor at Johns Hopkins University, first proposed the concept of 'Organoid Intelligence', a bio-computer system that combines brain organoids grown from nerve cells with brain-computer interfaces (BCI) in 2023. BCI is a technology that translates brain waves into electrical signals for communication with computers. The Johns Hopkins research team proposed that the signals from mini-brains, or organoids, could handle computational tasks similar to AI.
AI originally mimicked the brain. AI can learn large amounts of information and independently acquire methods without pre-programmed inputs. This is due to the machine learning technique known as deep learning. Deep learning emulates the visual information processing that occurs in the human brain. While nerve cells process specific information individually, they interconnect and share information in a complex manner to interpret images.
The problem is that AI, which mimics the brain, consumes too much energy. Training AI with large-scale data and keeping it operational demands an immense amount of electricity. For instance, it takes 1,300 megawatt-hours (MWh) of power to train generative AI models such as OpenAI's GPT-3. This is equivalent to the annual electricity consumption of 130 households in the U.S. AI's energy demand is expected to double within five years, accounting for 3% of global electricity consumption.
The Johns Hopkins research team stated that biochips combined with organoids could reduce power consumption by a factor of 1 million to 1 billion. The human brain, composed of 100 billion nerve cells and 100 trillion consolidations, consumes about 20W of energy, which is merely one-fifth of a light bulb's energy consumption.
◇Controlling mini robotic cars with organoids
The Johns Hopkins research team developed a brainwave (EEG) measurement device using brain organoids. Existing EEG devices are in the form of helmets. Weighing 1.4 kg, it detects electrical signals flowing on the brain's surface using 25 electrodes. A high-density EEG device can have up to 256 electrodes.
The organoid EEG is much denser. The brain organoid, with a diameter of 300 μm (micrometers; 1 μm is one-millionth of a meter), has 25 electrodes connected to it. This density is a million to a hundred million times higher than existing devices.
The researchers trained the brain organoids using reinforcement learning like AI. They applied electrical stimuli to specific areas and administered dopamine hormones that trigger feelings of happiness in the brain's reward center when the desired electrical signals were produced. Over time, the organoids learned to associate specific stimuli with outcomes.
The brain organoids that underwent reinforcement learning operated mini robotic cars. This demonstrated the potential for application in autonomous vehicles. Organoids are being used to assess drug efficacy instead of experimental animals. The researchers noted that the principles behind brain cognitive impairments, which AI systems can't test, may be understood through organoid intelligence.
Corporations are also developing organoid intelligence. Australia's Cortical Lab developed a brain cell system that played the PONG game. This year, they began selling a system capable of keeping brain cells alive for up to six months.
Switzerland's FinalSpark has built an organoid intelligence system that can be experimented on remotely for $1,000 a month. Last year, the company unveiled a system consisting of four brain organoids, each 0.5 mm in diameter and composed of 10,000 nerve cells, connected to 8 electrodes. This system allows sending electrical signals to the organoids and measuring signals from the nerve cells while training them to perform desired tasks.
◇Maintaining organoids, limitations in signal input and output
The organoid intelligence system is still in the proof of concept stage and not ready for practical applications. There are many challenges to overcome. Above all, sustaining the life of organoids made up of living cells is crucial. Even with oxygen and nutrients supplied, they cannot permeate to the center of the organoids. The Johns Hopkins research team has developed a system that actively injects oxygen, nutrients, and growth factors while removing waste.
There is also the limitation that organoids are still in an immature stage. This means that brain cells are at a developmental stage similar to that of a fetus. They are not yet completely organized like adult brain cells, so they cannot perform higher cognitive functions. Additionally, the operating methods of organoids vary, and standardization has yet to be achieved.
Koo Bon-kyung, Director General of the Institute for Basic Science, noted that "the bio circuits of brain organoids could offer greater energy and computational efficiency than the electronic circuits of silicon chips" and added, "Researchers in organoid intelligence seem to believe that organoids could confer efficiency to AI, similar to how existing computers are connected to quantum computers to create hybrid systems."
In reality, existing computers calculate sequentially, whereas nerve cells are interconnected and can perform parallel operations simultaneously. Director General Koo stated, "Especially, the lightweight nature of the organoid system may have significance for mobile AI operations" and added, "Currently, it is still challenging to correctly implement signal input and output in brain organoids, indicating that we are at a stage where we see potential and take on challenges."
References
AI for Good Summit 2025, https://aiforgood.itu.int/event/biochips-for-future-ai-computers/
Frontiers in Science (2023), DOI: https://doi.org/10.3389/fsci.2023.1017235
Neuron (2022), DOI: https://doi.org/10.1016/j.neuron.2022.09.001