How neurons learn in a dish: closed-loop training, DishBrain and the free energy principle

Neurons learn in a dish when a closed loop links their activity to consequences: a computer encodes information as electrical stimulation, reads the network’s response, turns it into an action, and feeds the outcome back. Over repeated cycles, cultured networks can shift their activity toward responses that change what they receive, as the DishBrain Pong experiments reported in 2022.

Key takeaways

• In vitro learning relies on a closed loop with four steps: encoding, stimulation, decoding and feedback.
• DishBrain (Kagan et al., Neuron, 2022) reported that mouse and human cortical cultures improved at a simplified Pong game within minutes, but only when feedback was closed-loop.
• Instead of reward and punishment, DishBrain used predictable stimulation after hits and unpredictable stimulation after misses, a design based on the free energy principle.
• The paper’s use of the word “sentience” drew public criticism from neuroscientists, and the authors replied in the same journal.
• The idea is not new: Steve Potter’s lab and Shahaf and Marom were training cultured networks around 2001.

What does it mean for neurons in a dish to learn? #

When neuroscientists say a culture “learns,” they mean something specific and modest: after experience, the network responds to the same situation differently, in a way that improves performance on a defined task, and the change is not explained by simple fatigue, drift or chance. That is a behavioral definition. It says nothing about understanding, awareness or feelings.

The neurons usually come from one of two sources. Some studies use primary cortical cells from rodent embryos. Others use human neurons derived from induced pluripotent stem cells, either as flat cultures or as 3D brain organoids. The cells are grown on a microelectrode array, which lets a computer both record and stimulate them. Our guide to microelectrode arrays explains how that interface works.

Left alone, these networks are not silent. They fire spontaneously and often produce network bursts, where much of the culture fires together. Wagenaar, Pine and Potter catalogued a rich repertoire of such bursting patterns in developing cortical cultures. Learning experiments try to shape this spontaneous, somewhat chaotic activity into something structured and task-directed.

How does closed-loop training work? #

The core idea is closed-loop stimulation: what the network does determines what it receives next. Every closed-loop learning experiment, from 2001 to today, follows the same four-step cycle.

1. Encoding. Information about the world (a ball’s position, a sensor reading, a target) is translated into a stimulation pattern. The choice of electrodes, pulse rate and timing carries the information. This is the “input” half of neural encoding and decoding.
2. Stimulation. The array delivers that pattern as electrical pulses, usually brief, charge-balanced pulses that make nearby neurons fire.
3. Decoding. The system records the network’s response, detects spikes, and converts activity in chosen regions into an output: move left or right, up or down, output a number.
4. Feedback. The output changes the simulated or physical world, and the consequences are sent back to the network as new stimulation. Depending on the design, feedback may be informative, rewarding, disruptive or simply absent.

The loop then repeats, often tens or hundreds of times a second. Latency matters. If feedback arrives too long after the activity that caused it, the network has no way to link the two. DishBrain, for instance, counted spikes in 10 millisecond windows to update its game.

Open loop vs closed loop #

The distinction between open and closed loop is not a technicality. It is often the main experimental variable. In open-loop stimulation, the network receives a fixed schedule of input regardless of what it does. In closed-loop stimulation, input depends on output. Several studies, including DishBrain and Bakkum, Chao and Potter (2008), found that learning-like changes appeared only when the loop was closed.

How did DishBrain teach neurons to play Pong? #

The best-known recent experiment is DishBrain, published by Brett Kagan and colleagues at Cortical Labs and collaborating institutions, with Karl Friston among the authors, in Neuron in 2022. The title claimed that in vitro neurons “learn and exhibit sentience when embodied in a simulated game-world.” Here is what the system did, step by step, based on the paper.

The hardware and the cells #

The team used a commercial CMOS high-density MEA that reads up to 1,024 of its 26,400 electrodes at 20,000 samples per second. They tested three kinds of cultures: primary cortical cells from embryonic mice, and human cortical neurons derived from induced pluripotent stem cells by two different differentiation methods. Custom low-latency software, which they called DishServer, sat between the array and a simulated game.

Encoding the ball #

Eight electrodes formed a “sensory” area. The ball’s position was encoded in two ways at once:

• Place coding: which of the sensory electrodes was stimulated indicated where the ball was along the vertical axis.
• Rate coding: stimulation frequency rose linearly from 4 Hz when the ball was at the far wall to 40 Hz as it reached the paddle’s side.

Sensory pulses were delivered at 75 mV. In pilot studies the authors found that increasing the density of sensory information, from place coding alone to place plus rate coding, improved performance.

Decoding the paddle #

Other regions of the array were defined as “motor regions.” Spikes in one region moved the paddle up; spikes in the other moved it down. Because neurons are spread randomly across the chip, nothing about the culture’s layout made these regions meaningful in advance. To control the paddle well, the network had to organize its activity in a way that matched an arbitrary layout. The authors also tested several motor configurations to check that the layout itself did not produce ball-tracking by accident.

The feedback design: predictable vs unpredictable #

This is the most distinctive and most misunderstood part of DishBrain. The cultures did not receive a “reward” chemical or a “pain” signal. Instead:

• After a hit, the cultures received a predictable stimulus: 75 mV at 100 Hz for 100 ms, delivered simultaneously across all eight sensory electrodes. Then the game continued normally.
• After a miss, the cultures received an unpredictable stimulus: 150 mV at 5 Hz, delivered at random sites and at random times across the eight input electrodes for four seconds, followed by a rest period. The ball then restarted in a random direction.

The idea was not that unpredictable stimulation “hurts.” It was that, under a particular theory of how self-organizing systems behave, a network should act to avoid states that lead to unpredictable input. If hitting the ball avoids chaos and missing produces it, a network that follows the theory should learn to hit more.

Controls #

To test whether improvement was real learning, the authors compared active cultures with:

• Media-only controls: MEAs with no cells, to catch artifacts.
• Rest sessions: cultures controlling the paddle but receiving no sensory information.
• In silico controls: the same game with the paddle driven by random noise.

In a follow-up experiment, they also varied the feedback itself:

• Stimulus: the standard predictable and unpredictable feedback.
• Silent: after a miss, all stimulation was withheld for the same duration instead of being randomized, then the game restarted.
• No feedback: misses produced no restart at all; the ball simply bounced on, while ball-position stimulation continued.

What they reported #

Kagan and colleagues reported “apparent learning within five minutes of real-time gameplay not observed in control conditions.” In their main comparison, only the mouse and human cortical cultures showed significantly longer average rallies (consecutive hits) in the later part of a 20-minute session (minutes 6 to 20) compared with the first five minutes, and both outperformed the control groups by then. Human cultures actually started worse than controls, which the authors suggested might reflect initially maladaptive control or exploratory behavior, then improved. Cultures also showed fewer “aces” (missing the first serve) over time.

In the feedback experiment, the Stimulus condition showed significant learning and outperformed both Silent and No-feedback conditions. Both Stimulus and Silent conditions showed higher overall average rally lengths than Rest and No-feedback, but only the Stimulus condition showed a significant increase over matched rest controls. The authors concluded that closed-loop, structured feedback was important for learning over time. They also reported that functional connectivity between sensory and motor regions was stronger during gameplay than at rest.

DishBrain at a glance #

What is the free energy principle, and why did DishBrain use it? #

The free energy principle (FEP) is a theoretical framework developed mainly by Karl Friston. In a 2010 Nature Reviews Neuroscience article, Friston proposed it as a candidate unified account of brain function. In simplified terms, it says that any self-organizing system that maintains itself separate from its environment acts as if it is minimizing a quantity called variational free energy, which bounds how “surprising” its sensory inputs are given its internal model of the world.

A system can reduce surprise in two ways:

1. Perception and learning: update its internal model so inputs become more predictable.
2. Action: change the world so that inputs match what it predicts. This second route is called active inference.

DishBrain was designed as a test of the second route. If a culture of neurons behaves according to active inference, it should take actions (moving the paddle) that make its future input more predictable. Hitting the ball produced predictable input; missing produced unpredictable input. So under the FEP, a culture should come to hit more often. That is what the authors reported, which they presented as consistent with the theory.

Earlier evidence from cultured neurons #

The FEP framing has roots in earlier culture work. Isomura, Kotani and Jimbo (2015, PLOS Computational Biology) stimulated rat cortical cultures with mixtures of hidden signals and reported that neurons learned to respond to distinct sources, a form of blind source separation. They reported that network changes reduced free energy, as the principle predicts, and argued that the result could only be explained by some form of Hebbian plasticity. Kagan and colleagues described their own study as extending this open-loop result to a closed-loop, embodied setting.

What the FEP framing does and does not show #

It is worth being careful here. Showing that behavior is consistent with a theory is not the same as showing the theory is the only explanation. Other accounts, such as networks reorganizing to reduce strong, disruptive stimulation, or simpler plasticity rules acting on correlated input, could predict similar outcomes. The DishBrain design, with its silent-feedback condition, partly addresses this, but the broader debate about what the FEP explains is a theoretical argument that continues well beyond this one study.

Why did the word “sentience” cause controversy? #

The paper’s title stated that neurons “exhibit sentience.” In the text, the authors explain they mean sentience in a specific formal sense, citing earlier work: being “responsive to sensory impressions” through adaptive internal processes. They did not claim the cultures were conscious, felt anything or had experiences.

Many readers, however, hear “sentience” as implying subjective experience. That gap between technical and everyday meaning drew criticism. In March 2023, Neuron published “A response to claims of emergent intelligence and sentience in a dish” by Fuat Balci and a large group of co-authors, challenging the use of terms such as sentience and intelligence for the reported results. Neuron also published a reply from Kagan and colleagues, “Scientific communication and the semantics of sentience,” which responded to the criticism and discussed how such terms should be used.

The episode is a useful lesson for anyone writing about organoid intelligence:

• Define terms explicitly and use the narrowest word that fits.
• Separate behavior from experience. Improved Pong rallies are a behavioral result. Claims about inner experience need different kinds of evidence.
• Remember that words travel. Headlines rarely carry the technical definition from a paper’s introduction.

The deeper ethical questions, including whether more complex organoids could ever have morally relevant experiences, are real and are discussed seriously by researchers. We cover them in the ethics of organoid intelligence, and the field’s own commitments appear in documents such as the Baltimore declaration (Hartung and colleagues, Frontiers in Science, 2023).

What mechanisms let neurons learn? #

Whatever the theoretical framing, the physical change must happen in the cells. The prime candidates are forms of synaptic plasticity: lasting changes in the strength of connections between neurons, driven by activity.

Hebbian learning #

In 1949, Donald Hebb proposed in The Organization of Behavior that when one cell repeatedly helps fire another, the connection between them strengthens. This is usually summarized as “cells that fire together wire together.” Hebbian learning is the general family of rules in which correlated activity strengthens connections. It explains how repeated patterns of stimulation could carve preferred pathways into a network.

Spike-timing-dependent plasticity #

Hebb’s idea was later refined by experiments showing that timing matters at the millisecond scale. Markram and colleagues (1997, Science) reported that synaptic efficacy changed depending on the relative timing of presynaptic input and postsynaptic action potentials. Bi and Poo (1998, Journal of Neuroscience), working with cultured hippocampal neurons, showed that a synapse tends to strengthen when the presynaptic spike comes shortly before the postsynaptic spike, and weaken when the order is reversed, within a window of about 20 milliseconds.

This rule, spike-timing-dependent plasticity (STDP), fits closed-loop training well. Stimulation in a closed loop creates consistent timing relationships between input, network activity and feedback, exactly the kind of structure STDP can act on. Many spiking neural network models and neuromorphic computing chips implement STDP for the same reason.

Other candidate mechanisms #

Plasticity is not limited to synapse strength. Cultures can also change through:

• Intrinsic excitability: neurons adjusting how readily they fire.
• Short-term plasticity and adaptation: transient changes that fade within seconds to minutes. Eytan, Brenner and Marom (2003, Journal of Neuroscience) reported that cortical networks adapt selectively, reducing responses to frequent stimuli while relatively enhancing responses to rare ones.
• Homeostatic mechanisms: slow processes that keep overall activity within a range.
• Structural change: growth or pruning of connections over days.

For most in vitro learning results, including DishBrain, the specific cellular mechanism has not been pinned down. The evidence is mainly behavioral (performance improves) plus network-level measures (connectivity changes). Linking those to particular synaptic rules remains open work.

Who trained neurons in a dish before DishBrain? #

DishBrain built on more than two decades of work. A few landmarks stand out.

Potter lab: animats and hybrots #

Steve Potter’s lab pioneered embodying cultured networks. DeMarse, Wagenaar, Blau and Potter (2001, Autonomous Robots) described the “neurally controlled animat,” in which rat cortical neurons on an MEA controlled a simulated animal in a virtual world, while the computer acted as its sensory system by feeding electrical stimulation back to the culture.

In a 2004 book chapter, “Removing some ‘A’ from AI: embodied cultured networks,” Bakkum and colleagues described embodying cultured neurons in both simulated and robotic bodies, and used the term “hybrot,” short for hybrid robot. A related art-science project, MEART, connected a cultured network to a drawing robot, described in a 2007 Frontiers in Neurorobotics paper by Bakkum and colleagues.

Shahaf and Marom: learning by removing stimulation #

In 2001, Goded Shahaf and Shimon Marom published “Learning in networks of cortical neurons” in the Journal of Neuroscience. Their protocol was elegant. They stimulated a cortical culture at low frequency (0.3 to 1 Hz) until a chosen response appeared at a chosen electrode, 50 ± 10 ms after a stimulus. At that moment, stimulation stopped for five minutes. Repeating this cycle, the network came to produce the desired response more directly, and plotting the number of stimuli needed per cycle produced learning curves. The authors described it as, to their knowledge, the first demonstration of learning arbitrarily chosen tasks in networks of real cortical neurons outside the body.

Note the logic: stimulation acts as a drive that keeps changing the network, and removing it “selects” whatever configuration produced the target response. That is conceptually close to DishBrain’s idea that networks settle into states that avoid disruptive input.

Closed-loop control and adaptive training #

Wagenaar, Madhavan, Pine and Potter (2005, Journal of Neuroscience) used closed-loop multi-electrode stimulation to control spontaneous bursting in cortical cultures, showing that feedback could shape network dynamics. Bakkum, Chao and Potter (2008, Journal of Neural Engineering) went further with an adaptive training algorithm. A neocortical network learned to reach predetermined activity states within tens of minutes, controlling movement in a task, with fewer training stimuli needed over time. Strikingly, replaying the same effective stimuli without making them contingent on the network’s behavior did not produce significant plasticity. Feedback, not just stimulation, mattered.

A timeline of in vitro learning #

What are the limitations of in vitro learning? #

These experiments are scientifically important, but they are early. Honest reporting means being clear about what they have not shown.

• Simple tasks. Pong with a single paddle, or producing a response at one electrode, is far from the flexible learning of an animal.
• Modest, short-term effects. Improvements are typically measured over minutes to hours. Whether learned changes persist over days, and how they are retained, is largely unresolved.
• Variability. Cultures differ in cell composition, density, age and activity. Results are statistical across many sessions and replicates, not reliable in every dish.
• Confounds. Stimulation can produce artifacts, and networks adapt to stimulation in ways that are not task learning. Good controls (no cells, no feedback, random output) are essential, and DishBrain’s design reflects that.
• Mechanism. Behavioral improvement is easier to measure than the underlying cellular change.
• Interface limits. Electrodes only reach part of the tissue, especially for 3D organoids, which constrains how much of a network can be “taught.”
• Interpretation and language. As the sentience debate showed, the words used to describe results can shape how they are received.

Roadmaps such as the organoid intelligence paper by Smirnova and colleagues (2023, Frontiers in Science) list many of these as research priorities: more mature and reproducible organoids, better 3D interfaces, longer-term experiments and shared data standards.

How does in vitro learning connect to biocomputing? #

Learning in a dish is one of two broad strategies for computing with living neurons. In the first, which this article covers, the network itself is trained to change its behavior through feedback. In the second, called reservoir computing, the network is treated as a fixed, complex dynamical system and only a digital readout is trained. Brainoware, which used a brain organoid this way to identify speakers from vowel sounds, is explained in reservoir computing with brain organoids.

Both strategies share infrastructure: reliable encoding, low-latency stimulation and recording, careful controls, and complete records of every stimulus and response. Reproducing an experiment like DishBrain in another lab requires knowing exactly what was delivered when. Platforms such as the remotely accessible Neuroplatform described by Jordan and colleagues (2024) point toward shared, cloud-accessible wetware computing infrastructure. This is also the problem we’re working on at Organoid Programming: a closed-loop runtime and experiment ledger that make protocols like these easier to define, run and reproduce (see the platform).

Conclusion #

Neurons learn in a dish through closed loops: encode information as stimulation, decode the response into action, and feed the consequences back. DishBrain made this idea famous by reporting that mouse and human cortical cultures improved at Pong within minutes, using predictable and unpredictable feedback inspired by the free energy principle, and by sparking a sharp debate over the word “sentience.” The underlying science is older and steadier, from Potter’s animats and Shahaf and Marom’s learning curves to adaptive closed-loop training. What remains open is mechanism, persistence and scale. For the wider field, see our guide to organoid intelligence and the future of biocomputing.

Frequently asked questions

Can neurons in a dish really learn?

Several studies report that cultured neurons change their activity in task-relevant ways when given closed-loop feedback. Shahaf and Marom (2001) showed cortical networks could be trained toward a chosen response, and Kagan et al. (2022) reported improved Pong performance within minutes. These are simple forms of learning, much narrower than animal learning, and their interpretation is still debated.

What is DishBrain?

DishBrain is a system described by Kagan et al. in Neuron (2022). Mouse or human cortical neurons grown on a high-density microelectrode array were connected to a simulated game of Pong. Electrical stimulation told the cells where the ball was, activity in defined regions moved the paddle, and feedback depended on whether the paddle hit or missed the ball.

How did DishBrain give neurons feedback?

When the paddle hit the ball, the cultures received a short, predictable stimulus: 75 mV at 100 Hz for 100 ms across all eight sensory electrodes. When the paddle missed, they received unpredictable stimulation: 150 mV at 5 Hz, delivered at random sites and times across those electrodes for four seconds. The design tested whether neurons act to avoid unpredictable input.

What does the free energy principle have to do with neurons playing Pong?

The free energy principle, associated with Karl Friston, proposes that self-organizing systems act to minimize surprise about their inputs. Kagan et al. reasoned that if cultured neurons follow it, they should change their activity to avoid states followed by unpredictable stimulation. Hitting the ball avoided that stimulation, so learning to hit it is consistent with the principle.

Were the DishBrain neurons sentient?

The paper used sentience in a narrow, formal sense: being responsive to sensory impressions through adaptive internal processes. It did not claim consciousness or feelings. The wording drew criticism, including a 2023 response in Neuron by Balci and colleagues, and Kagan and colleagues replied in the same journal discussing the semantics of sentience.

What mechanisms let neurons in a dish learn?

The likely mechanisms are forms of synaptic plasticity, where connection strengths change with activity. Hebbian learning strengthens connections between neurons that fire together, and spike-timing-dependent plasticity makes the direction of change depend on the precise order of pre- and postsynaptic spikes. Exactly which mechanisms drive closed-loop learning in cultures has not been fully established.

Who first trained cultured neurons with feedback?

Work on embodied and trained cultures goes back at least two decades. Steve Potter’s lab described the neurally controlled animat in 2001, and robotic hybrots in a 2004 chapter. Shahaf and Marom reported learning of arbitrarily chosen tasks in cortical networks in 2001, and Bakkum, Chao and Potter showed adaptive closed-loop training in 2008.

What are the main limitations of in vitro learning experiments?

Effects are often modest and measured over minutes or hours, cultures vary greatly, and tasks are very simple. It can be hard to separate genuine learning from adaptation or stimulation artifacts. Long-term memory, transfer between tasks and scaling remain largely untested, and the language used to describe results can outpace the evidence.

Sources

1. Kagan et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron.
2. Balci et al. (2023). A response to claims of emergent intelligence and sentience in a dish. Neuron.
3. Kagan et al. (2023). Scientific communication and the semantics of sentience. Neuron.
4. Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
5. Isomura, Kotani & Jimbo (2015). Cultured cortical neurons can perform blind source separation according to the free-energy principle. PLOS Computational Biology.
6. Shahaf & Marom (2001). Learning in networks of cortical neurons. Journal of Neuroscience.
7. DeMarse et al. (2001). The neurally controlled animat: biological brains acting with simulated bodies. Autonomous Robots.
8. Bakkum et al. (2004). Removing some 'A' from AI: embodied cultured networks. In Embodied Artificial Intelligence (Springer).
9. Bakkum et al. (2007). MEART: the semi-living artist. Frontiers in Neurorobotics.
10. Bakkum, Chao & Potter (2008). Spatio-temporal electrical stimuli shape behavior of an embodied cortical network in a goal-directed learning task. Journal of Neural Engineering.
11. Wagenaar, Madhavan, Pine & Potter (2005). Controlling bursting in cortical cultures with closed-loop multi-electrode stimulation. Journal of Neuroscience.
12. Eytan, Brenner & Marom (2003). Selective adaptation in networks of cortical neurons. Journal of Neuroscience.
13. Hebb (1949; reprinted 2005). The Organization of Behavior. Psychology Press.
14. Markram et al. (1997). Regulation of synaptic efficacy by coincidence of postsynaptic APs and EPSPs. Science.
15. Bi & Poo (1998). Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of Neuroscience.
16. Wagenaar, Pine & Potter (2006). An extremely rich repertoire of bursting patterns during the development of cortical cultures. BMC Neuroscience.
17. Smirnova et al. (2023). Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science.
18. Hartung et al. (2023). The Baltimore declaration toward the exploration of organoid intelligence. Frontiers in Science.
19. Jordan et al. (2024). Open and remotely accessible Neuroplatform for research in wetware computing. Frontiers in Artificial Intelligence.