Big news! The birth of a machine-organoid hybrid "biological computer"! It may overcome the bottleneck of AI hardware

The human brain, the "command center" of mankind, has about 200 billion cells, connected to each other through trillions of nanometer-sized synapses.

Currently, artificial neural networks powered by artificial intelligence (AI) hardware require about 8 million watts of energy, while the human brain only requires about 20 watts .

Through neuroplasticity and neurogenesis, the brain is also able to efficiently process and learn from noisy data with minimal training cost, thus avoiding the high energy requirements of high-precision computing methods.

Inspired by the structure and function of the human brain, researchers from Indiana University Bloomington, the University of Florida, Cincinnati Children's Hospital Medical Center, and the University of Cincinnati have invented a machine-organoid hybrid computing system - Brainoware .

The system includes traditional computing hardware and brain organoids, and can perform tasks such as speech recognition and nonlinear equation prediction. In addition, the system can flexibly change and reorganize in response to electrical stimulation, which is expected to meet the challenges of current AI hardware in terms of time and energy consumption and heat generation .

The related research paper, titled “Brain organoid reservoir computing for artificial intelligence,” has been published in Nature Electronics, a subsidiary of Nature.

The authors of the paper mentioned that brain organoids are only part of the system, and more complex artificial neural networks remain to be demonstrated .

In a News & Views article published alongside the paper, Johns Hopkins University associate professor Lena Smirnova and colleagues wrote, “As these organoid systems continue to grow in complexity, the neuroethical issues surrounding the study of biocomputing systems involving human neural tissue become increasingly important. While the creation of general-purpose biocomputing systems may be decades away, this research holds promise for providing fundamental insights into mechanisms of learning, neurodevelopment, and neurodegenerative diseases. ”

Voice recognition available

The recent success of AI has been driven primarily by the development of artificial neural networks (ANNs), which process large data sets using silicon computing chips. However, training ANNs on current AI computing hardware is energy-intensive, time-consuming, and the data is physically separated from the data processing unit, i.e., there is a von Neumann bottleneck .

The structure, function, and efficiency of the human brain provide inspiration for the development of AI hardware - the human brain combines data storage and processing in biological neural networks (BNNs), naturally avoiding the von Neumann bottleneck problem.

Inspired by BNNs, scientists have tried to develop efficient and low-cost neuromorphic chips, such as using memristors. However, current neuromorphic chips can only partially mimic brain functions, and it is important to improve their processing capabilities .

In this regard, the study introduces an AI hardware that exploits the reservoir computation and unsupervised learning capabilities of human brain organoid neural networks (ONNs), which are embedded in organoids. This approach is able to process spatiotemporal information and achieve unsupervised learning through the neural plasticity of organoids.

Figure | Brainoware AI computing with unsupervised learning (Source: the paper)

Compared to current 2D in vitro neuronal cultures and neuromorphic chips, Brainoware can provide more insights into AI computing because organoids can provide the complexity, connectivity, neuroplasticity, and neurogenesis of BNNs, as well as low energy consumption and fast learning.

Thanks to the high plasticity and adaptability of organoids, Brainoware is able to flexibly change and reorganize in response to electrical stimulation, highlighting its ability to perform adaptive reserve computing .

The study demonstrated that the method can display physical reserve properties such as nonlinear dynamics, memory decay and spatial information processing, as well as speech recognition and nonlinear equation prediction . In addition, the study also demonstrated that the method can learn from training data by reshaping the functional connectivity of ONNs.

Figure｜Speech recognition (Source: This paper)

However, current Brainoware approaches suffer from several limitations and challenges.

One technical challenge is the generation and maintenance of organoids. Although various protocols have been successfully established, current organoids still suffer from high heterogeneity, low generation efficiency, necrosis/hypoxia, and various activities. In addition, it is critical to properly maintain and support organoids to tap into their computational capabilities.

Although current Brainoware hardware is low energy, it requires additional peripherals that are still quite power hungry. Depending on the development of the electronics industry and system integration, it should be possible in the future to achieve very low energy integration using custom systems for maintaining and interfacing organoids.

The flat, rigid MEA electrodes used in Brainoware to interface with organoids can only stimulate/record from a few neurons on the surface of the organ. Therefore, it is necessary to develop methods to comprehensively interface organoids with AI hardware and software.

Another technical challenge is data management and analysis. Encoding and decoding spatiotemporal information from Brainoware still needs to be optimized, which can be achieved by improving the efficiency of interpreting, extracting, and processing data from multiple sources and modalities. In addition, this new AI hardware is likely to generate a large amount of data, which may require the development of new algorithms and methods for data analysis and visualization.

Broad application prospects

The above research on Brainoware is just an attempt by scientists in the direction of organoids.

As one of the research focuses, organoids refer to micro-organs with three-dimensional structures that can be cultured in an in vitro environment. They have complex structures similar to real organs and can partially simulate the physiological functions of real organs .

In 2009, Hans Clevers' team at the Hubrecht Institute in the Netherlands successfully cultured adult stem cells into small intestinal crypt and villus structures, marking the beginning of organoid technology.

Organoids hold great promise for organ transplantation and drug screening, and they offer the opportunity to create cellular models of human disease that can be studied in the lab to better understand the causes of disease and identify possible treatments. The power of organoids in this regard was first used in a genetic form of microcephaly, where patient cells were used to create brain organoids that were smaller and showed abnormalities in early neurons.

In 2021, a research team from the Austrian Academy of Sciences in Vienna successfully cultivated the world's first in vitro self-organizing cardiac organoid model using human pluripotent stem cells. This model can spontaneously form cavities and beat autonomously without the need for stent support. At the same time, this cardiac organoid can autonomously mobilize cardiac fibroblasts to migrate and repair damage after injury.

Figure | Beating heart organoids (Source: The Mendjan Lab)

Earlier this month, a paper published in the journal Nature Methods showed that scientists from the Institute of Molecular Biotechnology of the Austrian Academy of Sciences successfully developed an organoid model of the dopamine system. This model reveals in detail the complex functions of the dopamine system and its potential impact on Parkinson's disease. It is exciting that this organoid model can be used to improve cell therapy for Parkinson's disease.

Almost at the same time, in a research report published in the scientific journal Cell Reports, scientists from Stanford University School of Medicine and other institutions used the three-dimensional organ tissue model of organoids to screen out genes that cause the growth of many different types of cancers and identified very promising potential targets in oral cancer and esophageal squamous cell carcinoma.

Currently, organoid culture technology is experiencing a stage of rapid technological development and a large number of scientific research results. It has broad application prospects, but it also faces a series of important challenges : including how to effectively use human embryonic stem cells to establish a stable and lasting in vitro model; how to more realistically simulate the human microenvironment; and how to achieve mass production of products with scientific research attributes and successfully transform them into clinical products.

In the future, we expect the continuous development of organoid technology to bring more opportunities and breakthroughs to fields such as medicine, biology, drug development, and AI.

Reference Links:

https://www.nature.com/articles/s41928-023-01069-w

https://en.wikipedia.org/wiki/Organoid#Properties

https://pubmed.ncbi.nlm.nih.gov/37922313/

https://www.nature.com/articles/s41592-023-02080-x