Reading time: approximately 6–7 minutes · 27 October 2026 · Neuron (2022)

Can neurons growing on a chip learn to play Pong? And how can a dish of cells possibly know where the ball is?

Kagan and colleagues built DishBrain, a system that connects living cortical neurons to a simple virtual game through electrical stimulation and recording. The most interesting part is not the game itself. It is the mechanism: when a neural network’s actions change the sensory input it receives next, its activity can adapt within minutes.

1. Giving cultured neurons a body inside a game

The researchers used two sources of cortical neurons: human induced pluripotent stem cells (hiPSCs) differentiated into neurons, and cortical neurons from embryonic mice. The cells were grown on a high-density microelectrode array (HD-MEA), which could both stimulate the culture and record its electrical activity.

Imagine a Pong game: a ball moves across the screen, and a paddle moves vertically to intercept it. The cells have no eyes and no hands. So the computer translates the game into a language the neurons can receive: electrical pulses in, neural spikes out.

The entire loop is:

Ball position → electrical stimulation → living neural network → recorded spikes → paddle movement → hit or miss → new stimulation.

Original Figure 1 — DishBrain culture preparation and feedback arrangements

Figure 1 (excerpt). The experimental protocol connects human- or mouse-derived cortical neurons on an HD-MEA to Pong. The lower panels distinguish stimulation, silent feedback, no feedback and rest.

That final arrow—the consequence of an action flowing back to the cells—is what turns a recording setup into a closed-loop environment.

2. How does a dish of neurons “see” and move?

DishBrain used an intentionally simple encoder–decoder interface. There were no visual images shown to the cells; the ball’s position became a pattern of electrical pulses delivered through eight sensory electrodes.

Encoding the ball (sensory input). Two coding schemes were combined:

  • Place coding: which electrode was stimulated represented the ball’s vertical position (y).
  • Rate coding: stimulation frequency represented the ball’s horizontal position (x). The frequency increased from roughly 4 Hz to 40 Hz as the ball approached the paddle’s side of the screen.

So a ball at a particular height, moving closer to the paddle, produced stimulation in a particular location at an increasing pulse rate.

Decoding the paddle (motor output). Other electrodes were assigned to two motor regions. Activity in motor region 1 moved the paddle up; activity in motor region 2 moved it down. The system compared activity in the two regions and compensated for differences in their baseline firing so one consistently more active region would not dominate the paddle. The decoder itself was kept simple: the network had to change its own spike patterns to improve control.

The loop was therefore direct and fast. When the neural network changed its activity, the paddle moved, changing what the network would receive next.

Original Figure 4 — software loop, electrode layout and stimulation patterns

In Figure 4D, look for the sensory electrodes and the motor regions; Figure 4C shows how stimulation changes during gameplay.

3. The clever trick: make success predictable and mistakes surprising

If the cells control a paddle, what encourages them to hit the ball?

The authors drew on active inference and the free energy principle. The basic idea is that a system can organize its actions to make its incoming signals more predictable.

DishBrain expressed this idea through its feedback:

  • Successful hit: the system delivered a predictable, structured electrical pattern and the ball continued moving.
  • Miss: it delivered unpredictable, randomized stimulation, and the ball reset.

Think of trying to operate a machine by touch. One action produces a familiar, repeatable response; another produces an unexpected burst of changing signals. Over repeated attempts, the system may reorganize toward actions associated with a more predictable sequence of inputs.

In DishBrain, the stimulation was not an explicit label saying “good” or “bad.” Its timing and predictability provided the learning signal. The living network experienced the consequence of its own activity through the next incoming pulses.

4. What did the neurons actually learn?

The main experiment analyzed 399 game sessions across five groups: mouse cortical neurons (MCC), human cortical neurons (HCC), media-only controls (CTL), in-silico/random-output controls (IS), and resting cultures without sensory game information (RST).

Performance was measured by dividing a 20-minute session into T1 (0–5 min) and T2 (6–20 min). The researchers tracked three intuitive metrics:

  • Average rally length: how many consecutive returns the paddle made.
  • Aces: how often the ball passed the paddle without any return.
  • Long rallies: sequences of more than three consecutive hits.

Original Figure 5 — Pong performance across experimental and control groups

Figure 5. Start with panel B: average rally length rises from T1 to T2 in both mouse (MCC) and human (HCC) cultures. Panels C–D show fewer aces and more long rallies.

The result was striking: both human and mouse cultures improved their average rally length, while the control conditions did not show the same time-dependent improvement. The main comparisons reported p = 5.27 × 10⁻⁸ for MCC and p = 3.92 × 10⁻¹⁹ for HCC. The two biological groups also had fewer aces and more long rallies as play continued.

Notice what the experiment measures: the researchers did not train a conventional software classifier to predict paddle movement. Instead, the biological neural network adjusted its activity while connected to the game.

5. Why does closed-loop feedback matter?

The most important follow-up comparison changed what happened after a hit or miss:

Stimulus: predictable stimulation after a hit; randomized stimulation after a miss.

Silent: the game still changed after an action, but the explicit feedback pulse could be replaced by silence.

No-feedback: the cells received sensory stimulation, but the stream was not appropriately linked to the consequences of their paddle actions.

The results in Figure 6 show that the Stimulus condition produced the strongest improvement in rally length. The Silent condition performed better than No-feedback, while No-feedback did not produce the same learning trend. In other words, simply sending electrical pulses was less important than allowing neural activity to change what happened next.

Original Figure 6 — comparing Stimulus, Silent and No-feedback

Figure 6A visualizes the three stimulation schedules; Figure 6C compares their changes in rally length.

The authors also looked directly at neural dynamics. In Figure 7, electrophysiological activity reorganized during gameplay; information entropy was generally lower during gameplay than rest, while unpredictable feedback increased measured neural-response entropy. This links changing game performance with changing patterns of neuronal activity.

Takeaway

DishBrain’s essential contribution is a working read–write interface between a living neural network and a virtual environment. Eight electrodes encoded the ball; motor-region spikes controlled the paddle; the consequences of each action came back as electrical signals.

A neural culture was able to improve its behavior because it was part of a closed loop. This experiment offers a useful starting point for biological computing and for designing stimulation protocols that shape neural network activity through interaction.

References / 参考资料