A new field of research dubbed “organoid intelligence” is emerging as scientists look to build computers from lumps of brain cells grown in a petri dish.
These organoids are cultured from stem cells harvested from skin samples, and are tiny lumps of brain cells containing a jumble of neurons. Researchers believe they can store information, and can be trained to learn simple tasks like a computer that operates much more efficiently than artificial neural networks.
AI can now outperform humans at multiple tasks, but humans are still much better at processing information and learning new things, academics who coined the term “organoid intelligence” argued in a paper in Frontiers in Science.
DeepMind’s AlphaGo system beat the world’s best player at a competition in 2016, but required intensive training on samples generated from hundreds of thousands of games. The neural network was trained for weeks on 50 GPUs slurping approximately 40 billion joules, roughly the same amount of energy to sustain the metabolism of an adult for a decade, they said. Humans are over a million times better at handling data efficiency than AI algorithms.
“The brain is still unmatched by modern computers,” Thomas Hartung, co-author of the study and a professor of environmental health sciences at the Johns Hopkins University (JHU), said in a statement.
“Frontier … is a $600 million, 6,800-square-feet installation,” Hartung said, referring to the Oak Ridge National Lab’s supercomputer. “Only in June of last year, it exceeded for the first time the computational capacity of a single human brain – but using a million times more energy.”
Researchers believe some of the brain’s computational abilities and power could be replicated in organoids. Unlike traditional computers, these brain cells can’t be programmed using software. Instead, their electronic signals have to be manipulated somehow and used to control an output device.
The fascinating ability of the brain to make sense of millions of incoming neurons firing with information is very different to the logic of a computer program, which is executing a flow of zeros and ones
“What we humans learn is how to react to certain inputs,” Lena Smirnova, first author of the paper and an assistant professor at JHU, explained to The Register.
“The fascinating ability of the brain to make sense of millions of incoming neurons firing with information is very different to the logic of a computer program, which is executing a flow of zeros and ones.”
Smirnova believes organoids can be manipulated through electrode arrays where input signals can influence output signals. These jumbled mass of cells could, in theory, be trained to carry out simple tasks.
“The electrical signals exchanged between brain cells can be recorded by electrodes in contact with the organoid. An impulse or better a specific pattern of such signals can be translated into an action, for example of a connected robot or an action in a virtual game environment,” Smirnova said.
Researchers have already demonstrated that these brain cells can learn how to play Pong, but transforming them into biocomputers is a formidable challenge. Each brain organoid is about one three-millionth the size of a human brain; it’s unknown how much memory they can store; and it’s not clear how we will decode its electrical signals.
Advancing organoid intelligence will involve developing new machine learning and statistical algorithms, and scaling these biological structures to larger sizes – from 50,000 cells to 10 million. Organoids also require oxygen and nutrients to survive; their function and abilities will depend on their architectures too.
Still, researchers believe it’s worth pursuing even though brain cells in a dish are not comparable and could never replace their silicon-based counterparts. Instead, these biocomputers can be used as an interface with other machines to make systems more energy- and data-efficient.
- Machine-learning software scours database of already available drugs that could treat COVID-19 infections
- Lab grown stem cells emit brain waves like newborns – and boffins build robot worm to slither through heads
- Not just you in the night: Tiny bugs use superpropulsion to eject huge volumes of pee
- Fossil brain undoes 350 million years of scientific understanding
This does not mean that biocomputing will replace all machine learning
“Biocomputing will use far less power, can learn with less data so can make choices quickly in real-time, are likely far more flexible, and can provide us useful insights into how our own brains work, so can be useful for disease modelling and drug discovery,” Brett Kagan, chief scientific officer at Cortical Labs, a startup that created the DishBrain neuron system trained to play Pong, told The Register.
“This does not mean that biocomputing will replace all machine learning, but it could be a useful tool for certain tasks it is optimized for.”
“From here on, it’s just a matter of building the community, the tools, and the technologies to realize [organoid intelligence’s] full potential,” he added. ®
A new field of research dubbed “organoid intelligence” is emerging as scientists look to build computers from lumps of brain cells grown in a petri dish.
These organoids are cultured from stem cells harvested from skin samples, and are tiny lumps of brain cells containing a jumble of neurons. Researchers believe they can store information, and can be trained to learn simple tasks like a computer that operates much more efficiently than artificial neural networks.
AI can now outperform humans at multiple tasks, but humans are still much better at processing information and learning new things, academics who coined the term “organoid intelligence” argued in a paper in Frontiers in Science.
DeepMind’s AlphaGo system beat the world’s best player at a competition in 2016, but required intensive training on samples generated from hundreds of thousands of games. The neural network was trained for weeks on 50 GPUs slurping approximately 40 billion joules, roughly the same amount of energy to sustain the metabolism of an adult for a decade, they said. Humans are over a million times better at handling data efficiency than AI algorithms.
“The brain is still unmatched by modern computers,” Thomas Hartung, co-author of the study and a professor of environmental health sciences at the Johns Hopkins University (JHU), said in a statement.
“Frontier … is a $600 million, 6,800-square-feet installation,” Hartung said, referring to the Oak Ridge National Lab’s supercomputer. “Only in June of last year, it exceeded for the first time the computational capacity of a single human brain – but using a million times more energy.”
Researchers believe some of the brain’s computational abilities and power could be replicated in organoids. Unlike traditional computers, these brain cells can’t be programmed using software. Instead, their electronic signals have to be manipulated somehow and used to control an output device.
The fascinating ability of the brain to make sense of millions of incoming neurons firing with information is very different to the logic of a computer program, which is executing a flow of zeros and ones
“What we humans learn is how to react to certain inputs,” Lena Smirnova, first author of the paper and an assistant professor at JHU, explained to The Register.
“The fascinating ability of the brain to make sense of millions of incoming neurons firing with information is very different to the logic of a computer program, which is executing a flow of zeros and ones.”
Smirnova believes organoids can be manipulated through electrode arrays where input signals can influence output signals. These jumbled mass of cells could, in theory, be trained to carry out simple tasks.
“The electrical signals exchanged between brain cells can be recorded by electrodes in contact with the organoid. An impulse or better a specific pattern of such signals can be translated into an action, for example of a connected robot or an action in a virtual game environment,” Smirnova said.
Researchers have already demonstrated that these brain cells can learn how to play Pong, but transforming them into biocomputers is a formidable challenge. Each brain organoid is about one three-millionth the size of a human brain; it’s unknown how much memory they can store; and it’s not clear how we will decode its electrical signals.
Advancing organoid intelligence will involve developing new machine learning and statistical algorithms, and scaling these biological structures to larger sizes – from 50,000 cells to 10 million. Organoids also require oxygen and nutrients to survive; their function and abilities will depend on their architectures too.
Still, researchers believe it’s worth pursuing even though brain cells in a dish are not comparable and could never replace their silicon-based counterparts. Instead, these biocomputers can be used as an interface with other machines to make systems more energy- and data-efficient.
- Machine-learning software scours database of already available drugs that could treat COVID-19 infections
- Lab grown stem cells emit brain waves like newborns – and boffins build robot worm to slither through heads
- Not just you in the night: Tiny bugs use superpropulsion to eject huge volumes of pee
- Fossil brain undoes 350 million years of scientific understanding
This does not mean that biocomputing will replace all machine learning
“Biocomputing will use far less power, can learn with less data so can make choices quickly in real-time, are likely far more flexible, and can provide us useful insights into how our own brains work, so can be useful for disease modelling and drug discovery,” Brett Kagan, chief scientific officer at Cortical Labs, a startup that created the DishBrain neuron system trained to play Pong, told The Register.
“This does not mean that biocomputing will replace all machine learning, but it could be a useful tool for certain tasks it is optimized for.”
“From here on, it’s just a matter of building the community, the tools, and the technologies to realize [organoid intelligence’s] full potential,” he added. ®
