Neurons in a dish learn to play video games?! Extracorporeal neurons learn and exhibit perceptual abilities in a simulated game world

"From worms to flies to humans, neurons are the starting point for generalized intelligence," said
Brett Kagan, chief scientific officer at the Melbourne Cortex Laboratory in Australia.
"So, the question is, can we interact with neurons and harness this intrinsic intelligence?" Science fiction movies and science fiction have included the idea
of combining human or animal brains with robots to become intelligent robots.
The group of Australian academics envisions taking the first step of integrating neurons in petri dishes into digital systems, hoping to achieve performance
that cannot be achieved with computers alone.
They developed DishBrain, a system that harnesses the "adaptive computing power inherent in neurons" in a structured environment to integrate cultured nerve cells from humans or rodents with computers through a high-density multi-electrode array, and through electrophysiological stimulation and recording, will allow neurons in a dish to "learn" a simulated arcade game "Pong"
.

。
。

background

Harnessing the computing power of living neurons to create synthetic biological intelligence (SBI), previously limited to the realm of science fiction, now seems within reach
.
With attempts to develop biomimetic hardware that supports neuromorphic computing, the superiority of biocomputing has been widely recognized
.
However, no artificial system other than biological neurons can support at least three order complexity (capable of representing three state variables), which is necessary
to reconstruct the complexity of biological neuron networks (BNNs).

The authors attempted to grow neuronal cells from rodent embryos and induced human pluripotent stem cells (hiPSCs) on a high-density multi-electrode array (HD-MEA) to establish a network of functional in vitro biological neurons (BNNs) to demonstrate that these cultured neurons can exhibit biological intelligence—these neuronal cultures meet the formal definition of perception, that is, the ability to "respond to sensory impressions"
through adaptive internal processes.

The system, called DishBrain, harnesses the inherent properties of neurons to share the "language" of electrical activity — the BNN system
that connects computers to this network of biological neurons through electrophysiological stimulation and recording.
The authors focus on which processes lead to intelligent (goal-oriented) behavior
when BNN is embodied through a closed-loop system.
Perceptual behavior in an intelligent system requires two interrelated processes: First, the system must learn how external states affect internal states through "perception," and how internal states affect external states
through behavior.
Second, the system must infer from its sensory state when it should adopt a particular activity and how its behavior will affect the environment
.

To address the first necessity, the authors custom developed software drivers to create a low-latency closed-loop feedback system that simulates the exchange
with the BNN environment through electrical stimulation.
To address the second requirement, the authors used the DishBrain system to test a theoretical framework
for how intelligent behavior arises.
One theory of how intelligent systems in an environment manifest intelligent behavior is the theory of active reasoning in line with the Free Energy Principle (FEP)—that is, the process of active inference minimizes potential free energy, and the brain always tries to minimize the way free energy works (the Free Energy Principle (FEP).

。 FEP proposes a testable implication that at each space-time scale, any self-organizing system that is separate from its environment attempts to minimize its variational free energy (VFE).

The gap between the prediction and the actual feeling ("accident" or "prediction error") can be minimized in two ways: by optimizing the prediction of the environment to bring the prediction closer to reality, or by acting on the environment to make the feeling match its prediction.

According to this theory, BNNs hold "beliefs" about the state of the world, learning to minimize their VFEs, or actively changing the world to make it less surprising
.
If true, this means that BNN behavior should be shaped by simply presenting unpredictable feedback after "incorrect" behavior
.
In theory, BNNs should take action to avoid states that lead to unpredictable inputs, i.
e.
learn to avoid uncertainty
.

Validate the experimental process and results

The growth
of neuronal "wet parts" used for calculation The authors used three methods to construct the BNN system
with neuronal cells obtained by 3 methods1.
Cortex cells from the dissecting cortex of rodent embryos: culture of primary neural cultures from embryonic day 15.
5 (E15.
5) mouse embryos
.

2.
HiPSCs differentiate into monolayers of active heterocortical neurons that show mature functional properties
.
Using dual SMAD inhibition (DSI), the authors cultured long-term cortical neurons
that formed dense connections with supporting glial cells.

3.
A different hiPSC differentiation method, NGN2 direct reprogramming, extends the study
.
Cells obtained by this high-yield method express pan-neuronal markers, showing a high proportion of excitatory glutamatergic cells
.

These cells were co-cultured in nutrient-rich medium with a built-in high-density multi-electrode array HD-MEA, which exhibited complex morphology with a large number of dendritic and axon connections, and the integration of these neuronal cultures on HD-MEAs was confirmed by scanning electron microscopy (SEM), where densely interconnected dendrite networks could be observed, forming a staggered network across MEA regions, which could last for > 3 months
.

Over time, nerve cells exhibit characteristic spontaneous action potentials
The authors mapped the in vitro development of the electrophysiological activity of the nervous system at high spatial and temporal resolution
.
Monitor electrophysiological maturity with daily activity scans
.

The DishBrain system was developed to build a modular real-time platform to leverage neuronal computation
and interact
with neurons in a simulated environment.
The DishBrain environment is a low-latency, real-time system that interacts with vendor MaxOne software to scale its use
.
The system records electrical activity in neuronal cultures and provides "sensory" (non-invasive) electrical stimulation, similar to the generation of action potentials
through activity in neuronal networks.
Using the coding scheme described in STAR Methods, external electrical stimulation can convey a range of messages
.
The authors chose three different categories of information: predictable, random, and sensory
.
DishBrain wants to integrate these functions to "read" information and "write" sensory data to neural cultures in a closed-loop system, so neural "actions" affect future incoming "sensory" stimuli
in real time.
The goal is to represent BNN in a virtual environment and quantify demonstrable learning
.

How do you make these cells "understand" the game?

The basic principle of DishBrain simulating the classic arcade game "Pong" is that there is a slider on the far left that can be moved up and down, the "racket", and the "ball" is shot in
from the right.
If the 'racket' blocks the "ball" as a "hit", the ball repeats the same line after the break, and if the "ball" is not blocked, the ball is shot
from a random position after the break 。 A multi-electrode array embedded in a neuronal culture designates a "predefined sensory region" with eight electrodes for input stimulation (Figure 4D), and defines the other part of the multielectrode array as a motor region where the discharge signal in motion zone 1 is converted to "racket" to move "up" and the activity signal in motion zone 2 is converted to "racket" to move "down", which means that cultured cells will need to adopt different firing modes
by self-organizing.
Real-time collection of electrophysiological activity signals in designated sports areas to move the "racket" in computer games to block the "ball"
.
If the ball is not blocked by the activity at the moment, the program enters an unpredictable stimulus in the sensory area (150mV, 5Hz for 4 seconds; See STAR method), then the "ball" stimulation will start
over on a random vector.
In contrast, if the activity is successfully intercepted at the moment, a predictable stimulus is given a frequency of 100Hz and delivered simultaneously to all electrodes for 10 milliseconds (briefly interrupting the regular sensory stimulus), and then the game continues
in a predictable manner.
Research confirms the following:

• "Learning" was observed in both human and primary mouse cortical neurons BNN, which performed better than mouse neurons after "learning" over time

• With closed-loop feedback, BNN's performance in the gaming environment improves over time, i.
e.
"learns" over time and hits rate increases

• BNN needs to learn feedback, and a system with stimuli but no feedback has no learning

• Increasing the density of sensory information input improves performance•

• Dynamic changes in neurophysiological activity observed during the experiment

The
DishBrain system is able to visualize BNNs from a variety of sources in a virtual environment and measure their responses
to stimuli in real time.
The adaptive responsiveness of neurons to external stimuli is well documented in vivo because it forms the basis of
all animal learning.
This work is the first study
to establish this fundamental behavior for goal-directed behavior in vitro.
This is the first SBI device
to demonstrate adaptive behavior in real time.
The system itself provides the opportunity to extend previous neurobehavioral computer models, such as testing hippocampal and endoolfactory cell models in solving spatial and nonspatial problems
.
The DishBrain platform can make small changes – such as selected cell types, drug administration, and feedback conditions – will enable in vitro testing to obtain information
about how cells are processed and computed that were previously unavailable.
Most importantly, this work has made substantial technical advances in creating a closed-loop environment for BNNs, highlighting the requirement to achieve goal-directed learning in the nervous system, where denser information and more diverse feedback influence performance
.

ConclusionUsing
this DishBrain system, it was demonstrated that cortical neurons in vitro can self-organize their activities to display intelligent and perceptual behavior
.
These findings provide a promising proof for the SBI system to learn
over time in a systematic manner that inputs guidance.
The system provides a fully visualized learning model that can develop unique environments to evaluate the actual calculations
performed by BNN.
This goes beyond purely computer models or predictions
of molecular pathways alone.
Although a lot of hardware, software, and wet engineering is still needed to improve the DishBrain system, this work does demonstrate that the computing power of living neurons can learn adaptively in active exchanges with their sensory organs
.
This is the biggest advance
to date towards SBI that responds with externally defined goal-oriented behavior.

Instantiated synthetic biological intelligence (SBI) may herald a paradigm shift
in biological intelligence research.
Explore how BNN computing helps improve machine learning methods, potentially resulting in biocomputing platforms
that outperform existing hardware.
Theoretically, generalized SBI could precede
artificial intelligence (AGI) due to the inherent efficiency and evolutionary advantages of biological systems.

Future directions for this work have the potential to be in disease modeling, drug discovery, and expanding current understanding of how the brain works and how intelligence is generated
.
"This is the beginning of a new frontier in understanding intelligence," "which involves not only what it means to be human, but also what it means to be alive and intelligent, and what it means
to process information and be sentient in an ever-changing, dynamic world.
" ”