2026-10-24
HAL trains 10 rat cortical cultures with random stimulation until the next burst if ignition is off-goal. Goal-side bursts reach about 70-80% in two days, between sessions.
Dissociated rat cortical neurons on a microelectrode array produce network bursts that sweep the tissue. Bursts ignite in fairly stable zones and then travel in a small set of directions. Learning experiments in culture have mostly asked a temporal question: stimulate until a chosen electrode shows the desired response, then stop. Shahaf and Marom got clean learning curves that way. Potter's lab closed the loop through a simulated body. The cultures in Kagan and colleagues' work learned a simple game within minutes. Those changes all lived inside a single session.
Two spatial and dynamical questions were still open. Can feedback move the place where a burst ignites toward a goal side? And if the network is punished for an incorrect burst, does it learn to enter that state less often, or to leave it more reliably once it is there? No culture study had held a task from one day to the next. A change that exists only inside the session it was trained in is not yet the kind of retained objective that biohybrid computing talks about.
Clawson, Levin and colleagues at the Allen Discovery Center at Tufts built an open-source closed-loop platform, HAL, on a MaxWell MaxTwo array. One array has 26,400 electrodes and records 1,024 of them at a time. Spikes are detected at 5.5 times the RMS noise. Four stimulation electrodes sit in the four corners, chosen to take little part in bursts. Pulses are delivered to all four at once, as a tissue-wide signal rather than an instruction to a named neuron. The cells are E18 rat cortex, about 100,000 cells, near 2,000 per square millimetre.
Online detection uses a sliding window of 200 frames, about 20 ms. A burst starts when more than 15% of the recorded network is active and stays above that threshold for 100 ms. It ends when activity stays below threshold for 50 ms, so a brief dip does not split one burst into two. The origin is the mean x of the electrodes active when the threshold is first crossed. The peak is the mean x when activity in the window is highest. A burst is left if origin x minus peak x is negative, right if that difference is positive, and dropped if it is zero. Every trained culture was given left as the goal. If the tissue covers only part of the array, the peak follows the centroid of the activity, so the divider is that burst's own peak, not the geometric midline of the chip.
The rule is thin. A burst in the target direction produces no stimulation. A burst in the wrong direction starts biphasic pulses, 250 mV and 200 µs per phase, at random intervals with a mean rate of 5 per second, on all four electrodes, until the next burst. The detector is blanked for 800 µs after each pulse so stimulation-evoked spikes are not counted as a new burst.
Each experimental day is 20 minutes before stimulation, 20 minutes of closed-loop training and 20 minutes after. The learning score is the pre-stimulation recording. Within-day change compares pre and post on the same day. Across-day change compares post on one day with pre on the next. Each culture was trained for upwards of 7 days, the exact count depending on development and viability. Days are aligned to each culture's first valid recording.
The main cohort comes from two preparations. stimRemoval, plated in July 2024, contributes 6 cultures in the analysis window DIV 30 to 36. waveTraining, plated in April 2025, contributes 4 cultures at DIV 34 to 42. The 6 controls are unstimulated sister cultures from the first preparation, recorded at DIV 12 to 26. The 6 damaged cultures come from a separate preparation in November 2025 and are never pooled with the main cohort. Inclusion required at least 2 recording days and at least 10 train-phase bursts. Seven further waveTraining cultures were excluded because they used a nonstandard reward rule.
Trained cultures raised the fraction of goal-side bursts within about two days and settled near 70% to 80%. First training day against last, within each culture, gives a Wilcoxon signed-rank W=5.0, p=0.019, n=10. Controls do not rise over the same window, W=9.0, p=0.844. Controls have no real goal side, so a side is assigned by a coin flip. One such assignment gives a between-group effect size of Mann-Whitney U=51, p=0.022. Enumerating all 64 goal-side assignments of the six controls, only 23% reach significance, and the median two-sided p is 0.18. The authors treat that between-group effect as a best case, not a stable difference.
When the gain accrues is cleaner inside the trained group. The across-day change exceeds the within-day change, W=0.0, p=0.002, n=10. The same across-day contrast against controls points the same way and misses the line, U=48, p=0.056. The authors read the pattern as consolidation-like in timescale: little change from pre to post inside a session, with the increase visible in the pre-stimulation recording of a later day. They do not claim sleep or replay.
Recovery is the harder between-group result. Each burst is a correct or an incorrect state. The change in the probability of moving from incorrect to correct is U=46, two-sided p=0.026, and one-sided p=0.013 in the direction stated in advance, n=9 trained against 6 controls. It stays significant for 72% of the 64 control assignments, with a median two-sided p of 0.034. The probability of falling from correct into incorrect does not change, U=22, p=0.43. Time spent in the correct state does not separate the groups either, U=43, p=0.18. The change in recovery and the change in correct dwell time correlate at Spearman ρ=0.83, p=0.005, n=9. Both quantities say the culture is becoming more correct, and the authors do not treat them as independent confirmation.
Spatially, origin x moves toward the goal side, one-sided Wilcoxon W=46, p=0.032, n=10. The peak does not move, W=18, p=0.375. The origin shift is larger than the peak shift, paired W=3, p=0.010. Control origins do not move, W=8, p=0.69. Angular entropy, mean cosine similarity and circular variance of propagation direction all give p=0.80 between trained and control cultures. The set of directions stays. The ignition point moves. The between-group contrast on the raw origin shift itself is not significant, U=41, p=0.26. The spatial claim rests on the within-trained split between origin and peak, together with the flat controls.
Six cultures whose baseline already favoured the goal side started near 0.71 correct. Four cultures biased the other way started near 0.31, U=24, p=0.0095. On the last day the two groups sat near 0.79 and 0.67, U=17, p=0.35. The against-bias group gains more, which is what a lower starting point produces. It is not evidence that training works better against a culture's innate bias.
Stimulation dose does not track effect size. Spearman ρ=0.42, p=0.23 for the total count of aversive pulses. ρ=0.31, p=0.39 for the number of train-phase bursts. ρ=0.35, p=0.33 for the fraction of bursts that were stimulated. None is significant. The three predictors are correlated, and n=10 is underpowered. Absence of a detected dose relationship is not a proof of independence.
Damaged cultures match healthy ones on burst count, U=21, p=0.26, and likewise on interval and duration. Fraction correct goes from 0.59 to 0.56, W=8, p=0.69. Their within-day versus across-day contrast is W=9, p=0.84, with none of the healthy signature. The healthy-versus-damaged effect size is U=40, p=0.313. A culture-clustered test of the group-by-day interaction gives p=0.111. Bursting is not enough for this reorganization. That between-group test is underpowered too.
Goal-directed demonstrations in cultured networks had lived on the scale of minutes and a single session. Here, goal-side ignition and recovery after an error accumulate across days. The loop has no decoder and no reinforcement-learning policy. The reward is the cessation of stimulation. For biohybrid computing, that is the useful claim: a complicated input-output machine is not a prerequisite for across-day learning. HAL is open source and runs on a commercial high-density array.
A rise in fraction correct and a drop in the rate of errors come apart in this dataset. The rate of entering the incorrect state does not change. The probability of climbing back out does. If one number is going to be quoted, the recovery transition survives the control-side assignment better than fraction correct does.
The outlook in the paper reaches drug sensitization, disease states and brain-computer interfaces. The data support a narrower statement. Healthy dissociated cortical networks can be moved, over a few days, by one global and very thin reward rule: where bursts start, and how quickly they recover after an error.
The samples are small. Ten trained cultures come from two preparations plated about nine months apart, with 6 controls and 6 damaged cultures. Age is the harder problem. Controls were recorded at DIV 12 to 26. The trained analysis windows are DIV 30 to 42. Baseline inter-burst interval, burst size and duration already differ. The paper states that the between-group effect size does not survive adjustment for those baseline burst dynamics. A claim that training beats ordinary development, read across groups, does not hold. What holds is the within-culture comparison: pre-stimulation fraction correct rises, and it rises between days.
There is no yoked stimulation control and no randomly rewarded control. Unstimulated sister cultures separate goal-directed training from spontaneous drift. They do not separate the amount of stimulation from the contingency between stimulation and origin. The dose correlation of ρ=0.42 is not significant, and it points positive. At n=10 that is not a refutation.
The learning score excludes the 20 minutes of training itself. "Little change within a day" means little change between the quiet pre and post recordings. Origin is collapsed to one x coordinate. The damaged cohort comes from a preparation named burstTrainer, and the text does not give a reproducible damage procedure. The recovery analysis uses 9 trained cultures, because a day counts only if the culture made at least one transition out of the relevant state. Seven cultures on a nonstandard reward were dropped at inclusion. Sensitivity to the filters is left in Table T1.
The mechanism was not measured. Consolidation-like describes a timescale. Unchanged angular entropy, cosine similarity and circular variance support an unchanged repertoire of directions. The authors read that as a reweighting of patterns the culture already had, a simple form of niche construction. There is no separate census showing that no new pattern appeared. The preprint was posted on 16 October 2026 and has not been peer reviewed.