Rat Neurons Trained With ML

The new result out of Japan sounds like science fiction until you look at what the researchers actually built. They took cortical neurons from rats, grew them in culture, wired them to a high-density microelectrode array, and then trained that living network inside a machine-learning setup. The work came from Tohoku University and Future University Hakodate, and it was published in *PNAS* in March 2026 as “Online supervised learning of temporal patterns in biological neural networks under feedback control” (pnas.org, tohoku.ac.jp). What they trained the neurons to do was not image recognition or language. It was something stranger and, in this context, more important: generate time-varying signals. That matters because many real systems, from movement to speech to control loops, depend on patterns that unfold over time. The team used a machine-learning approach called reservoir computing, which does not try to tune every connection in the network. Instead, it treats the network itself as a rich dynamical system and trains only a simple output layer that reads the network’s activity and feeds part of that signal back in (pnas.org, wpi-aimr.tohoku.ac.jp). That distinction is the whole story. In ordinary AI, the “neurons” are mathematical units inside software. Here, the reservoir was made of living cells. The researchers recorded voltage traces from the cultured network, converted those signals into a continuous reservoir state, and then used FORCE learning, a real-time error-correction method, to train the readout. The biological network was not simulating dynamics on a chip. It was the dynamical system. The hardware and the wet tissue were the same thing (pnas.org, techxplore.com). The hard part was not keeping the neurons alive. It was stopping them from becoming too simple. Neurons grown in a dish often synchronize too much, firing together in ways that wipe out the messy internal activity reservoir computing needs. So the team used microfluidic devices to shape how the cells grew and connected, creating modular architectures instead of one overlinked blob. That top-down control gave the cultures more varied internal dynamics, which made them trainable in the first place (pnas.org, wpi-aimr.tohoku.ac.jp). Once that system was in place, the neurons learned to produce several kinds of target signals. The paper and university release say the cultures generated sine, triangle, and square waves, and even reproduced the Lorenz attractor, a classic chaotic system. The same setup also learned sine waves with periods from 4 to 30 seconds. That is the striking part. The experiment was not just a one-off demonstration of a simple oscillation. It showed that a living neural culture could be pushed into multiple stable temporal behaviors under closed-loop training (pnas.org, techxplore.com). This does not mean rat neurons are about to replace GPUs. The paper supports a narrower claim. Living neural networks can serve as trainable physical reservoirs for certain signal-generation tasks. That is interesting for two reasons. It gives neuroscientists a controllable way to study how real networks produce structured dynamics, and it gives engineers another example of computation happening in matter rather than in abstract code. The researchers frame it as a step toward biologically inspired, energy-efficient neuromorphic systems, but the concrete achievement is simpler than the hype: a dish of rat neurons, guided by microfluidics and a readout algorithm, learned to trace a chaotic trajectory that mathematicians know as the Lorenz attractor (pnas.org, tohoku.ac.jp).