In a breakthrough at the intersection of neuroscience and artificial intelligence, researchers in the Cortical Laboratory have demonstrated that biological neuroculture learns faster and more efficiently than some of the most advanced machine learning algorithms available today. This striking finding was revealed through a pioneering experimental comparison between in vitro neural networks known as “dish brains” and the cutting-edge deep reinforcement learning (RL) model, marking an important milestone in understanding intelligence itself.
The central component of this research, Dishbrain is an innovative system that fuses silicon-based substrates with living human neurons grown from stem cells, allowing for a true hybrid platform where biological and synthetic components interact seamlessly. This synthetic biological intelligence (SBI) system utilizes high density multi-electrode arrays (MEAS) to promote real-time closed-loop interactions in dynamic gaming environments, particularly in versions of classic pon games. Such integration allows for the monitoring and manipulation of neuronal activity that dynamically evolves in response to stimuli, an experimental design that has never been performed at this scale or accuracy.
The study, entitled “Plasticity and Sample Efficiency of Dynamic Networks in Biological Neural Cultures: A Comparative Study with Deep Reinforcement Learning,” systematically maps how neural cultures exhibit plasticity, namely how instantaneous underlying connectivity and adaptive changes in firing patterns between gameplay vs. resting states. By reducing the high-dimensional spike activity of neurons to interpretable and low-dimensional representations, investigators can decipher the underlying network reconstruction, which implies learning and adaptation. This level of analysis supports basic neuroscience theory that links synaptic plasticity to functional changes associated with intelligence and learning ability.
.adsslot_uxmrywpr6v {width: 728px! Fality;Height: 90px! important. }
@media (max-width: 1199px) {.adsslot_uxmrywpr6v {width: 468px! Fality;Height: 60px! important. }}
@media (max-width: 767px) {.adsslot_uxmrywpr6v {width: 320px! Fality;Height: 50px! important. }}
advertisement
One of the deepest aspects of this study is the quantification of sample efficiency. This is the number of training examples or interactions the system needs to improve performance. Artificial RL systems such as DQN (Deep Q-network), A2C (Advantage Actor-Critic), and PPO (Proximal Policy Optimization) often require millions of iterations before demonstrating meaningful learning, and biological neural cultures showed rapid acquisition of new behaviors with far fewer samples. This is more closely in line with natural animal learning, where organisms effectively adapt within limited exposure to stimuli and emphasize the superior adaptability of living neural networks.
Besides serving as a compelling proof of concept, this work paves the way for a paradigm shift in AI research by suggesting that intelligence should not be seen as an artificial construct born from algorithms, but rather as a fundamentally biological phenomenon. According to Brett Kagan, Chief Science Officer at Cortical Labs, the breakthrough challenges the existing notion that intelligence can be fully replicated with silicon-based calculations alone, and advocates adopting biological substrates as their own powerful computational entities.
The meaning of leveraging the term “bioengineering intelligence” (BI), introduced by teams in companion research, goes far beyond isolated learning tests. BI envisions a future where designed neural circuits from laboratory-grown neurons can be accurately structured and interfaced to perform complex processing tasks that are comparable or exceed traditional AI methods. This complements a new field of organoid intelligence (OI) that employs naturally grown brain organoids but does not involve the same degree of designed control over network architecture.
In analyzing the data, researchers at the Cortical Institute show that the rapid reorganization of synaptic activity seen in the dishroom reflects true functional improvements in learning task performance, rather than merely a statistical phenomenon. This is evidenced by the reconstruction of the patterns of connection between neurons as gameplay progresses, reflecting the principles governing the cognition of the brain in an intact mammal. Co-author of this study, Moein Khajehnejad, highlighted that extracting interpretable, low-dimensional signals from spike patterns was more clearly illuminated these internal plasticity processes than previous methodologies allowed.
The comparative benchmarks implemented show a pioneering approach in AI assessment, placing equally the biological systems and deep RL methods on the number of samples available for learning and real-world time. This direct, direct challenge highlights the possibility that synthetic biological systems can not only match, but outweigh adaptation speed and robust artificial agents under conditions that emulate true learning scenarios. It is humble insight for researchers striving to unravel the essence of cognition and intelligence.
Support for this research comes from a respected international consortium including the Turner Institute for Brain and Mental Health at Monash University, the IITB-Monash Research Academy in India, and the Welcome Human Neuroimaging Centre in London. This collaborative expertise highlights the interdisciplinary complexity of research and integrates stem cell biology, computer science, neuroscience, and bioengineering in an unprecedented way.
Experts in this field have expressed enthusiasm for the potential of the CL1 platform, the first commercial biological computer to come from this study. Professor Millera Dottori of the University of Wollongong said such techniques not only advance basic neuroscience, but also provide a new tool for exploring neurological diseases by providing dynamic and functional measures of neural network behavior. Similarly, Hideaki Yamamoto from Tohoku University praised the rapid development and commercialization of CL1 devices, and recognized its promise as a versatile tool for investigating brain calculations.
This groundbreaking research illustrates a fundamental shift in how scientists and engineers approach the future of artificial intelligence. By bridging living neural tissue with a computational framework, researchers are creating courses for machines that simply mimic but embody biological intelligence. The fast and efficient learning capabilities demonstrated by these cultured neural networks challenge the general assumptions and open and exciting possibilities for developing adaptive, resilient, and ethically sustainable AI systems.
Continuing research and technical improvements have made bioengineering intelligence poised to redefine the meaning of computation, intelligence, and systems “learning.” As the Cortical Institute pushes this early technology, the convergence of biology and machinery not only marks the study of brain secrets, but also a new era in which actively harnesses to shape the future of intelligent machines.
Research subject: Cells, synthetic biology, intelligence, biomaterials, biotechnology
Article Title: Dynamic network plasticity and sample efficiency in biological neural cultures: A comparative study with deep reinforcement learning.
News Release Date：August 12, 2025
Web reference:http://dx.doi.org/10.1016/j.celbio.2025.100156
reference:
Cortical Labs et al. , Plasticity and sample efficiency of dynamic networks in biological neural cultures: A comparative study with deep reinforcement learning, cyborg and bionic systems: Science Partner Journal, 2025.
Image credits: Cortical Research Institute
Advanced Biological Neural Networks in Artificial Intelligence Reinforcement Learning Comparative Research Survey Results Dynamic Game Environment
