• AI neural networks require about one million times more power than human brains use.
• Previous neural networks using live neurons were inefficient, so researchers found a viable solution in a polymer scaffold that allows neurons to function properly.
• While this new type of biological neural network still needs to be scaled up, it shows promise for faster computation and lower energy consumption in the future.

Computers are often thought of as artificial brains. That said, they’re nowhere close to being as complex or energy-efficient as human brains. AI consumes an enormous amount of electricity, and it’s constantly demanding more as it keeps endlessly morphing and advancing. How can our power supply catch up to a future of neural networks and digitized intelligence? Well, maybe the answer lies in merging living brain cells with a programmable electronic system.

Previous attempts to use actual neurons as the brain of a computer have run into problems. 2D neural cultures—in which the flattened neurons showed abnormal interactions and gene expression—couldn’t survive for long, and these structures were ultimately unable to replicate the connections and activity that occur in vivo. More advanced in vitro neural networks have tried to compensate for some of those problems by mirroring the structure and function of the brain with organoids. Despite some improvements, brain organoids (clumps of stem cells engineered to turn into neurons) are inconsistent and prone to both hypoxia and necrosis.

Alternative 3D neural networks known as biological neural networks (BNNs) could still be a viable option. Such a system would ideally take the form of an in vitro model that reconstructs the brain’s networks, can be reproduced, and actually lasts. It would also feature both dense and sparse neural connections (not unlike those in the hippocampus) to prevent too much data from moving around at once. In an effort to create a fusion of biology and machinery, researchers Tian-Ming Fu, James Sturm, and Kumar Mritunjay from Princeton University used electrodes and microscopic metal wires to create a 3D polymer mesh scaffold flexible enough for tens of thousands of living neurons to grow into a network that could operate with minimal energy.

“Understanding the brain’s network structures and working principles could help in the development of general-purpose computing with improved data and energy efficiencies, as well as provide insights into the brain’s physiology and pathology,” the researchers said in a study recently published in Nature Electronics.

Fu, Sturm, and Mritunjay began this experiment to gain more insight into other lingering questions about brain function, but soon saw its potential as a biological neural network, and 3D-MIND (3D Micro-Instrumented Neural network Device) was born. Taking inspiration from origami, the researchers initially created the device in two dimensions, embedding precisely enough electronic sensors to match the soma of a neuron before folding it into 3D layers. Neurons were then integrated into the system. While this hasn’t been done with human neurons yet, rat neurons from the hippocampus—which is critical for learning and memory—were extracted and cultured on the scaffold.

Finally, the entire device was covered in a thin gel coating. Protective and practical, the coating contained proteins that would provide extra support for neurons in forming strong connections with glial cells—cells that not only hold these neural structures together, but supply nutrients, perform immune functions, regulate chemistry, produce the fatty insulation for axons known as myelin, and keep the surrounding environment clear for signaling.

Eventually, the researchers observed neurons positioning themselves and forming connections in three dimensions throughout the structure. These neurons were also stable enough to be tracked for extended periods of time, and the team managed to record growth, development, and action potentials—electrical impulses that neurons use to communicate.

The researchers admit that it will be challenging to scale this system up, but it’s definitely promising, especially when compared to current energy-guzzling AI networks.

“The interfaced system can then provide a physiologically relevant understanding of the brain’s 3D network connectivity,” the team wrote. “[It has] the ability to track a 3D neural network [and] could be of use in understanding the efficiency and versatility of the brain’s computational capabilities.”