Brain Matters Vol. 8 No. 2


Graphic processing units (GPUs) are a major component of artificial intelligence (AI) processing power.
As AI becomes more sophisticated and more processing power is needed to run these intellectual
models, a growing concern of energy demand and extensive AI training becomes an increasing concern.
To create more sophisticated machine learning algorithms, scientists in the field of neuromorphic
computing studied the brain for its ability to efficiently process and store information. Finding a way to
incorporate the brain directly into computing may create novel algorithms to meet the increasing
demands of information processing.

Written by Edward Lin

Abstract

Biological Neural Networks as
the Forefront of AI Processing

Introduction
Graphic processing units (GPUs) are chips within a
computer that process data simultaneously, performing
parallel computations. Making approximately 36,500
calculations per second, GPUs consist of three major
components that use basic arithmetic operations for neural
network processing. The GPU’s tensor core is the forefront
of critical AI operations and AI learning, which uses
geometric transformations and large matrix computations
for AI neural network optimization. As science further
progresses, we obtain more data about the world, which
requires more processing power. This exponential growth in
data requires increasingly more powerful processing
capabilities to interpret this data. However, the current
processing power of modern-day computational systems
are inefficient and expensive, as they have a high energy and
resource demand for running the GPUs and regulating their
temperature as shown in Figure 1. Since the late 1980s,
scientists and engineers have been studying the structural
organization within the brain to help them optimize
information processing, giving rise to the field of
neuromorphic computing. The processing speed of the brain
is similar to that of a supercomputer, and outperforms a
supercomputer in terms of energy efficiency, spatial
optimization, memory, and storage. Rather than mimicking
and studying how brain structures optimize information
processing, it may be effective to explore the direct
integration of these systems, also known as “reverse 

Figure 1. Comparison Between Supercomputer
and Human Brain

1

neuromorphic computing”. Implementing the biological
neuronal networks within the brain directly into a
computer’s processing units can maximize efficiency,
leading to greater information processing and more 



sophisticated machine learning algorithms (Kagan et. al,
2023). As the brain receives new information, it reorganizes
itself and forms new connections. The incorporation of self-
restructuring capabilities to AI opens more dynamic
approaches to AI training. In addition, the incorporation of
neuroplasticity functions in GPUs may allow neural
networks to adjust how they process inputs without
rigorous retraining. Such approaches give engineers the
freedom to develop more powerful algorithms with deeper,
more advanced neural networks. The implementation of
neuroplasticity would enable AI models to process complex,
continuous flows of data more effectively, allowing the
energy-efficient and adaptable data processing of the brain
to be manifested in GPUs.

The implementation of
neuroplasticity would

enable AI models to process
complex,

continuous flows of data
more effectively, allowing
the energy-efficient and

adaptable data
processing of the brain to be

manifested in GPUs.

Biological Neural Networks (BNNs)
Biological Neural Networks BNNs, or biological neural
networks, are clusters of neurons connected to each other
through axons and dendrites (in a sense, they can be
described as miniature brains with the most basic
complexity). Through these axonal and dendritic
connections, BNNs are able to demonstrate characteristics
that mirror those of artificial intelligence: computational
performance and network plasticity (the ability of neurons
to arrange themselves based on the stimulation received).
The vast quantity of connections within a BNN enables the
biological structure to undergo parallel processing across
multiple neuronal signaling pathways and allows stimuli to
be distributed vastly. When integrated within computing
systems, BNNs have the ability to not only process the
information and exhibit a response, but also rearrange
themselves according to the stimulus, displaying plasticity
and small amounts of memory (Dranias et. al, 2014). Memory
consolidation in BNNs can be categorized into two types of
memory processes, fading memory and hidden memory.
Fading memory relies on the firing activity in response to a

Brain Matters Vol. 8 2025

2

stimulus, lasting only for a brief moment, while hidden
memory depends on synaptic plasticity to strengthen the
connections between neurons. Hidden memory allows the
BNN to retain information for prolonged periods. However,
the retention of this information is disrupted or
restructured when the BNN receives a high-intensity
electrical input (stimuli).

Multielectrode Arrays
MEAs, or multielectrode arrays, help researchers integrate
neurons within computational devices. Structurally, they
are a flat surface with multiple microelectrodes embedded
in an array placed underneath a BNN that is in a petri dish.
Each electrode independently records extracellular activity.
The recorded signals are then digitized and processed by a
computer to interpret neural activity. The computer then
generates an electrical stimulus that is sent as a response to
the neurons to evoke another response, generating a real-
time feedback loop.

Figure 2. Multielectrode array structure (left to right)
schematic, 60 electrode array, cultured neurons

Ping-Pong Experiment
Through an experiment revolving around ping-pong,
researchers connected a BNN to an MEA that provided input
of the game’s environment, allowing the BNN to control a
ping-pong paddle in real-time. The neurons received
structured feedback through successful (positive
reinforcement) and missed hits (negative reinforcement),
allowing the BNN to self-correct from an optimal to a more
accurate dynamic state. Through this, the BNN could retain
task-specific information for brief intervals, displaying a
positive correlation between branching ratio
(neuroplasticity) and task performance. This enhanced BNN
performance for that one specific task. When researchers no
longer provided feedback to the BNN, its “hit-to-miss ratio”
decreased (Habibollahi and Kagan, 2023). Once connected to
a computing system through a multielectrode array, BNNs
rearranged themselves until they reached an optimal
dynamic state to efficiently process the structured stimuli.
Despite reaching an optimal state, BNNs displayed low
accuracy but a high level of computing, thus requiring
feedback. Feedback was only in regards to one specific task,
allowing BNNs to process information efficiently and with
high accuracy for one task, making BNNs optimal for
computing one task over and over again. However, if
required to switch tasks, BNNs needed to rearrange
themselves and be given the correct feedback to perform a
task swiftly (Habibollahi and Kagan, 2023).



Neuronal Capabilities of Matrix Processing
The ping-pong experiment's success with parallel
processing and real-time adaptation points to the potential
of larger neural systems. In the brain, neural processing in
the cerebral cortex occurs at intervals of a few milliseconds.
Despite being slower than a computer, the cortex is able to
compensate for this difference through its ability to process
vast amounts of information in parallel (Ballard, 1986).
Resembling GPUs, this parallel processing in the brain is
how the brain receives sensory information and executes
motor skills (Sigman and Dehaene, 2008). In addition,
hierarchical (information flowing through successive layers
of processing) and modular organization (each cortex is
divided into specialized regions for particular tasks) of the
brain is essential for matrix computations. Present artificial
intelligence neural networks follow a similar information
flow, but creating a processing system consisting of
specialized BNNs can structure information for maximal
efficiency and scalability. This hypothetical computational
system, much like the brain, will have areas of the
computer’s processing unit distinguished by function. Each
area will have their own specialization of information
processing. As information is received, the information will
be decomposed and sent to the right processing subregions,
taking into advantage a BNN’s fading memory. Information
processing can then be split into smaller, manageable bits of
information that each subregion of BNNs can process,
allowing for scalability and efficiency.

Ethics
Despite using only BNNs, which in comparison to a human
brain is minuscule in proportion, it is possible to create a
sentient life force when such an organization is scaled to an
extent. Eventually, if such biotechnological methods are
implemented, one will create structures that contain a
neuron count that mirrors certain large mammalian
organisms. In the ping-pong game, researchers were able to
notice a certain degree of self awareness within the BNN,
despite being comparatively small to that of a simple
organism’s (Kagan et. al, 2022). However, if neurons were
integrated into technology, will such a system eventually
have a form of consciousness? Due to exponential growth of
technology, the size of BNN integrated systems will also
have to increase to address this demand. But once a certain
number of neurons are integrated and interacting, it can
eventually have its own life force. It is essential to bring up
such topics when discussing the potential future direction
advancement of biocomputational devices to prevent the
unethical exploitation of sentient beings as inanimate
objects. Furthermore, ethical considerations regarding the
creation of such conscious systems must be explored,
drawing a line as to when a system has developed
consciousness.

Conclusion
Although the idea of integrating neurons into the processing 

Biological Neural Networks as the Forefront of AI Processing

system of computers seems efficient and beneficial to the
further advancement of AI, it is still fairly new and many
caveats have yet to be considered. For example, how would
BNNs be integrated at a scale into computers/super
computers that will provide enough processing power to
match that of the computer? How much energy and
resources would it take to maintain the functionality of a
BNN once integrated into the computer’s processing? Such
topics are hard to answer and current research on the topic
is in its infancy. In addition, with the global emergence of AI
made public in only the past few years. But the benefits of
such computational systems are apparent and mitigate the
flaws of current information processing systems in
computers.The ability of the brain to process information in
a hierarchical and modular manner, as well as BNN’s
plasticity and hidden memory, can allow information
processing systems to be optimally efficient. Without the
rigid structure of AI neural networks, such a system allows
for fluidity of adaptability, while taking on an energy-
efficient approach.

References
1. Smirnova, L., Caffo, B. S., Gracias, D. H., Huang, Q., Pantoja,
I. E. M., Tang, B., Zack, D. J., Berlinicke, C. A., Boyd, J. L.,
Harris, T. D., Johnson, E. C., Kagan, B. J., Kahn, J., Muotri, A.
R., Paulhamus, B. L., Schwamborn, J. C., Plotkin, J., Szalay, A.
S., Vogelstein, J. T., . . . Hartung, T. (2023). Organoid
intelligence (OI): the new frontier in biocomputing and
intelligence-in-a-dish. Frontiers in Science, 1.
https://doi.org/10.3389/fsci.2023.1017235
2. Roberts, T. P., Kern, F. B., Fernando, C., Szathmáry, E.,
Husbands, P., Philippides, A. O., & Staras, K. (2019). Encoding
Temporal Regularities and Information Copying in
Hippocampal Circuits. Scientific reports, 9(1), 19036.
https://doi.org/10.1038/s41598-019-55395-1
3. Kagan, B. J., Gyngell, C., Lysaght, T., Cole, V. M., Sawai, T.,
& Savulescu, J. (2023). The technology, opportunities, and
challenges of Synthetic Biological Intelligence.
Biotechnology Advances, 68, 108233.
https://doi.org/10.1016/j.biotechadv.2023.108233
4. Dranias, M. R., Westover, M. B., Cash, S., & VanDongen, A.
M. (2015). Stimulus information stored in lasting active and
hidden network states is destroyed by network bursts.
Frontiers in integrative neuroscience, 9, 14.
https://doi.org/10.3389/fnint.2015.00014
5. Sigman, M., & Dehaene, S. (2008). Brain mechanisms of
serial and parallel processing during dual-task performance.
The Journal of neuroscience : the official journal of the
Society for Neuroscience, 28(30), 7585–7598.
https://doi.org/10.1523/JNEUROSCI.0948-08.2008
6. Nelson, M. E., & Bower, J. M. (1990). Brain maps and
parallel computers. Trends in Neurosciences, 13(10), 403–
408. https://doi.org/10.1016/0166-2236(90)90119-u
7. Ballard, D. H. (1986). Cortical connections and parallel
processing: Structure and function. Behavioral and Brain
Sciences, 9(1), 67–90.
https://doi.org/10.1017/s0140525x00021555

3



8. Kagan, B. J., Kitchen, A. C., Tran, N. T., Habibollahi, F.,
Khajehnejad, M., Parker, B. J., Bhat, A., Rollo, B., Razi, A., &
Friston, K. J. (2022). In vitro neurons learn and exhibit
sentience when embodied in a simulated game-world.
Neuron, 110(23), 3952–3969.e8.
https://doi.org/10.1016/j.neuron.2022.09.001

Brain Matters Vol. 8 2025

About the Author
Edward is a Sophomore at the University of Illinois majoring in
Neural Engineering. Through his studies, he aspires to implement
biological mechanisms/systems into computers and explore AI-
neural network connections. Some of his interests include playing
volleyball, filming, and going on road trips. After graduation, he
hopes to attend graduate school.

4



5