Revolutionizing AI Hardware: Bridging the Gap with Brain-Inspired Technologies


Artificial Intelligence (AI) is dramatically transforming industries, medicine, society, and education, heralding a new era of technological advancement. The core driver of this revolution is the advancement of Artificial Neural Networks (ANNs), which utilize big data and silicon computing chips.

However, the full potential of AI is yet to be unlocked due to significant limitations in current AI hardware, which is energy-intensive, time-consuming, and is approaching its theoretical limits. A paradigm shift is proposed through the development of Brainoware, a novel concept that mimics the human brain’s structure and efficiency using brain organoids.


AI’s widespread impact across various sectors (1) is a testament to its transformative power. The success of AI in these fields is largely attributed to the rise of ANNs, which handle vast datasets efficiently. However, the hardware used for these networks is currently a bottleneck, creating a demand for innovation.

The Challenges of Current AI Hardware

The limitations of existing AI hardware are multifaceted (4-6, 8-10). Primarily, the immense heat generation and the excessive energy and time requirements for training ANNs are major concerns. Additionally, the hardware’s development pace is slowing down, diverging from the predictions of Moore’s Law. The ‘von Neumann bottleneck’—the separation of data and processing units—further hinders efficiency.

Inspiration from the Human Brain

The human brain, a complex biological network of approximately 200 billion cells connected by trillions of synapses (12-15, 17-18), operates with remarkable efficiency, consuming about 20 watts of power. In contrast, AI hardware needs about 8 million watts for similar operations ( 6). The brain’s ability to process and learn from noisy data with minimal training costs, thanks to neuronal plasticity and neurogenesis, positions it as an ideal model for AI hardware.

Neuromorphic Chips: A Step Forward

Pioneering efforts in developing high-efficiency and low-cost neuromorphic chips, such as memristors (11, 19-30), represent a significant advancement. These chips can store electrical states and support short-term memory functions. They are currently employed in areas like computer vision and speech recognition. Despite this progress, neuromorphic chips are limited in their ability to fully replicate brain functions, especially in processing real-life uncertainties and energy efficiency.

Brainoware: The Future of AI Hardware

Brainoware is an innovative concept that uses brain organoids—3D structures derived from human stem cells that replicate aspects of a developing brain’s structure and function ( 31-34). This living AI hardware leverages the computational and learning capabilities of biological neural networks. Brainoware offers advancements in complexity, connectivity, neuroplasticity, and neurogenesis, all achieved with lower energy consumption and faster learning compared to current technologies.


The development and demonstration of Brainoware, a novel form of living AI hardware based on human brain organoids, represent a significant step forward in the field of artificial intelligence. Brainoware leverages the unique characteristics of human brain organoids, which possess the remarkable ability to self-organize and form functional biological neural networks (BNNs).

These BNNs exhibit complexity and diversity, offering the potential to mimic the capabilities of the human brain, thus inspiring the creation of more sophisticated and human-like AI systems. The inherent plasticity and adaptability of brain organoids enable Brainoware to change and reorganize in response to electrical stimulation, highlighting its capacity to learn and adapt over time, a crucial attribute for the development of advanced AI systems.

As a living brain-like AI hardware, this approach naturally addresses the challenges of time consumption, energy consumption, and heat production that plague current AI hardware. Furthermore, it demonstrates physical reservoir properties such as nonlinear dynamics, fading memory, and spatial information processing, and has been successfully implemented in real-world applications, including solving non-linear equations. This section discusses the implications, limitations, and challenges associated with the Brainoware approach.

Unique Advantages of Brainoware: Brainoware introduces a groundbreaking approach to AI computing by tapping into the computational capabilities of living neural networks within human brain organoids. This unique approach offers several distinct advantages:

  • Mimicking Human Brain Complexity: The complexity and diversity of organoid neural networks hold the potential to mimic the intricate workings of the human brain more closely, opening up possibilities for the development of highly advanced and human-like AI systems.
  • Flexibility and Adaptability: Brainoware exhibits a high degree of plasticity and adaptability, enabling it to change and reorganize in response to electrical stimulation and training. This adaptability is essential for the continual improvement of AI systems.
  • Addressing Hardware Challenges: The living nature of Brainoware naturally mitigates the time-consuming, energy-consuming, and heat production challenges inherent in current AI hardware.
  • Real-World Applications: Brainoware has been successfully applied to real-world tasks, such as speech recognition and solving non-linear equations, showcasing its practical utility in various domains.

Limitations and Challenges: While Brainoware represents a promising advancement in AI hardware, several limitations and challenges must be addressed:

  • Organoid Generation and Maintenance: The generation and maintenance of brain organoids pose technical challenges, including high heterogeneity, low generation throughput, and viability concerns. Advances in differentiation protocols and microenvironment manipulation are needed to improve the standardization and scalability of organoid production.
  • Electrode Interface: Current Brainoware implementation relies on flat and rigid MEA electrodes, limiting the number of neurons that can be stimulated or recorded on the organoid surface. Developing brain-machine interfaces, soft electrodes, and neural implants is imperative to enable comprehensive interaction between AI hardware and the entire organoid, facilitating information exchange and activity manipulation.
  • Data Management and Analysis: Encoding and decoding temporospatial information to and from Brainoware require optimization, including improvements in data interpretation, extraction, and processing across multiple sources and modalities. The generation of substantial data volumes necessitates the development of new algorithms and methods for efficient data analysis and visualization.

Conclusion: In summary, the development of Brainoware represents a pioneering effort in the creation of brain-inspired living AI hardware with on-chip learning capabilities. The successful real-world applications of Brainoware in speech recognition and problem-solving demonstrate its potential to revolutionize AI hardware. However, addressing the limitations and challenges outlined above is crucial for further advancement and broader adoption of this innovative approach. Brainoware offers new perspectives and insights into the future of AI hardware, potentially leading to the development of highly advanced, adaptable, and energy-efficient AI systems.

TABLE 1 – Functional Biological Neural Networks (BNNs)

Functional Biological Neural Networks (BNNs) are an emerging field at the intersection of neuroscience, biology, and artificial intelligence. They represent an advanced form of computing, inspired by the structure and functioning of biological brains. Here’s a detailed overview of the current research, applications, and challenges in this domain:

Research and Development

  • Brain-Inspired Computing Hardware: Traditional brain-inspired silicon chips, despite emulating brain structure and principles, have limitations in fully mimicking brain function. A novel approach using adaptive reservoir computation of biological neural networks in brain organoids, termed “Brainoware,” shows potential. This technique involves sending and receiving information from a brain organoid using a high-density multielectrode array. It can achieve nonlinear dynamics, fading memory properties, and unsupervised learning, reshaping the organoid’s functional connectivity. Applications demonstrated include speech recognition and nonlinear equation prediction​​.
  • Oracle-Supervised Neural Engineering Framework (osNEF): The osNEF method trains biologically-detailed spiking neural networks to perform cognitively-relevant dynamical systems. It uses a variety of neuron models and synaptic models to realize computations found in cognitive systems, like communication, multiplication, harmonic oscillation, and gated working memory. This approach aligns with animal data in performance and forgetting rates, demonstrating the possibility of training functional brain models using biologically-detailed components​​.
  • Biological Detail in Neural Models: The incorporation of synaptic and cellular mechanisms in cognitive models allows exploration of psychology aspects, such as mental disorders and their pharmacological treatment. However, simulating biological detail increases the complexity of building and analyzing cognitive models. There’s a trade-off between biological realism and cognitive capacity in these models. To bridge this gap, a balance between biophysical detail and cognitive capacity is essential​​.


  • Neuro-Robotic Systems: In vitro BNNs, when embodied in robots, exhibit complex behaviors like supervised and unsupervised learning, memory, object tracking, obstacle avoidance, and playing games. These BNN-based neuro-robotic systems utilize a stable bidirectional connection, transferring sensor signals from robots to BNNs and decoding neural activities of BNNs into robot control commands. These systems display a range of intelligent behaviors, including learning and memorization, which can be classified based on computing capacity or network plasticity​​​​.
  • Computing Capacity-Dependent Behaviors: In some cases, the BNN acts merely as an information processor, generating specific neural activities in response to stimuli without the necessity of learning. This aspect is crucial in tasks where adaptability and learning are not required​​.

Challenges and Future Directions

  • Fabrication of 3D BNNs: A significant challenge is creating 3D BNNs that closely resemble their in vivo counterparts. This advancement is crucial for enhancing the functionality and realism of BNNs.
  • Training Robot-Embodied BNNs: Training BNNs is complex due to the absence of various neuromodulators found in biological systems. Transplanting training methods used in animals to BNNs is thus challenging. Additionally, while BNNs can perform simpler tasks, higher-level cognitive tasks remain a significant hurdle​​.

In conclusion, Functional Biological Neural Networks represent a significant step forward in emulating biological brain functions for various applications, particularly in robotics and AI. However, the field faces challenges in enhancing the realism of these networks and effectively training them for complex cognitive tasks. Continued research and development in this area are likely to yield groundbreaking advancements in understanding and replicating brain functions.

reference link:


Please enter your comment!
Please enter your name here

Questo sito usa Akismet per ridurre lo spam. Scopri come i tuoi dati vengono elaborati.