What is Brainoware? Brain organoid reservoir computing explained

Brainoware is a system described by Cai et al. in Nature Electronics (2023) in which a human brain organoid on a high-density microelectrode array acted as the reservoir of a reservoir computer. Inputs arrived as electrical stimulation, the organoid turned them into rich activity, and a trained digital readout decoded it for speaker identification and chaotic prediction.

Key takeaways

• Brainoware applied reservoir computing to living tissue: the organoid was the untrained reservoir, and only a digital readout was trained.
• It came from Feng Guo’s group at Indiana University Bloomington and was published in December 2023.
• It was tested on speaker identification from Japanese vowel recordings and on predicting the Hénon map, a chaotic system.
• MIT Technology Review reported about 78 percent speaker accuracy after training, below artificial neural networks on that task.
• Coverage reported that training reshaped the organoid’s connectivity, so the “fixed” reservoir also adapted.

How does Brainoware work? #

A reservoir computer has three parts: an input layer, a reservoir with rich nonlinear dynamics, and a readout trained to map reservoir states to outputs. In a digital echo state network, the reservoir is a random recurrent network in software. Brainoware replaced that software with a brain organoid.

1. Input encoding. Time-dependent data, such as features of an audio clip, were converted into spatiotemporal sequences of electrical pulses.
2. Stimulation and recording. The organoid sat on a high-density microelectrode array, which delivered the pulses and recorded the evoked activity.
3. Readout. Features of the recorded activity were passed to a trained digital model. For the speech task, coverage reported that this was logistic regression.

The organoid’s job was the one a reservoir always has: to project inputs into a high-dimensional, nonlinear space with a fading memory of recent history, so that a simple classifier can separate them. Turning audio into stimulation and spikes back into features is a problem of neural encoding and decoding.

Why use an organoid as a reservoir? #

A network of living neurons already has the properties reservoir theory asks for. It is recurrent, each neuron is strongly nonlinear because it either fires an action potential or does not, and activity triggered by a stimulus reverberates before settling. Crucially, reservoir computing does not require setting the reservoir’s internal weights, which is something nobody can do directly in living tissue. Tanaka et al. (2019, Neural Networks) review the broader idea of physical reservoirs built from optics, electronics, mechanics and other materials. Brainoware is a biological instance of that idea.

What tasks did Brainoware perform? #

The study used two standard benchmarks from the reservoir computing literature.

Speaker identification #

The first task was to identify which of eight male Japanese speakers produced a vowel recording, using 240 audio clips, according to coverage. MIT Technology Review reported accuracy of about 78 percent after training. News-Medical, summarizing the paper, reported that accuracy rose over training from 51 percent to 78 percent. MIT Technology Review also reported that Brainoware was less accurate than artificial neural networks on this task.

Chaotic time-series prediction #

The second task was to predict the Hénon map, a simple two-dimensional rule that produces chaotic behavior. Chaotic prediction is a classic reservoir benchmark; Jaeger and Haas (2004, Science) used chaotic time-series prediction to demonstrate echo state networks. Coverage of Brainoware reported that it outperformed baselines including linear regression and an artificial neural network on this task. Exact figures and baseline architectures are in the paper itself.

Brainoware at a glance #

Was the organoid learning? #

In classic reservoir computing, the reservoir is fixed. Brainoware’s reservoir was not entirely fixed. Coverage reported that repeated training stimulation reshaped the organoid’s functional connectivity, which the authors interpreted as unsupervised learning, and that the effect depended on neural plasticity. An organoid-free control, with no tissue on the array, was reported to perform at a regression score of zero on the prediction task, indicating the computation came from the tissue rather than the electronics alone.

This adaptation links Brainoware to synaptic plasticity and to closed-loop experiments such as DishBrain. The difference is that Brainoware’s changes were driven by exposure to inputs, not by feedback about whether its answers were right. Our article on reservoir computing with brain organoids walks through the experiment in more detail.

How does Brainoware compare with other approaches? #

What open questions does Brainoware raise? #

Brainoware answered one question, whether organoid tissue can serve as a working reservoir, and raised several others that later experiments will need to address.

• How much does the tissue contribute? The organoid-free control suggests the tissue matters, but comparing the same readout against a well-tuned digital reservoir on identical encoded inputs is the clearest way to measure its contribution.
• Which encoding works best? The way audio or time series are turned into stimulation shapes what the reservoir can represent. Small changes in pulse timing, amplitude or electrode choice could change results.
• How stable is the reservoir? If an organoid’s responses shift from day to day, a readout trained on Monday may fail on Friday. Tracking drift is as important as reporting peak accuracy.
• Can adaptation be steered? Training appeared to reshape connectivity. Whether that change can be guided deliberately, for example with closed-loop stimulation, is an open research direction.
• What scale is needed? Larger or more mature organoids, or several organoids linked together, might give richer dynamics, but they are also harder to keep healthy and to record.

Each of these is an experimental design problem as much as a biological one, which is why careful logging and repeatable protocols matter.

What are the limitations of Brainoware? #

The authors and commentators were clear that Brainoware is an early demonstration. Smirnova, Caffo and Johnson’s commentary in Nature Electronics placed it in the wider program of organoid intelligence. Practical challenges include:

• Maintenance. Organoids need continuous culture and life support.
• Variability and drift. No two organoids are identical, and each one changes over time, so readouts may need retraining.
• Interface limits. A flat electrode array samples only the surface of a three-dimensional tissue.
• Power accounting. The tissue uses little energy, but the incubators, amplifiers and computers around it use far more.
• Scope. Two benchmarks were tested; long-term memory and multi-task use were not.

Running organoid reservoir experiments with every stimulation sequence, readout and culture lineage logged, so they can be repeated and compared, is one of the workloads we are designing for at Organoid Programming.

Conclusion #

Brainoware showed that a human brain organoid can serve as a working reservoir in a hybrid computer, with a trained digital readout turning its activity into answers. Its accuracy was modest and below conventional neural networks on speech, but it gave biocomputing a concrete, published example of living tissue doing useful signal processing. For the hardware side of such systems, see our article on bio-silicon hybrid computing.

Frequently asked questions

What is Brainoware in simple terms?

Brainoware is a hybrid computer in which a living human brain organoid does part of the processing. Data are turned into patterns of electrical pulses and delivered to the organoid through an electrode array. The organoid’s responses are recorded, and a simple trained program reads them to produce an answer. Cai and colleagues described it in Nature Electronics in 2023.

Who built Brainoware?

Brainoware was built by Hongwei Cai and colleagues in the group of bioengineer Feng Guo at Indiana University Bloomington. Their paper, Brain organoid reservoir computing for artificial intelligence, appeared in Nature Electronics in December 2023, alongside a commentary by Lena Smirnova, Brian Caffo and Erik Johnson titled Reservoir computing with brain organoids.

How accurate was Brainoware?

MIT Technology Review reported that Brainoware identified speakers from Japanese vowel recordings with about 78 percent accuracy after training, and that it was less accurate than artificial neural networks on that task. With eight speakers, chance is 12.5 percent. The authors presented the work as a proof of concept rather than a competitive replacement for digital machine learning.

Did the organoid in Brainoware hear speech?

No. The organoid never received sound. Audio recordings were converted into sequences of electrical stimulation, and the organoid responded to those pulses. The classification itself was performed by a trained digital readout using features of the organoid’s recorded activity. Lena Smirnova of Johns Hopkins University made this point in coverage by MIT Technology Review.

How is Brainoware different from DishBrain?

DishBrain used flat cultures of neurons that directly controlled a game paddle, with feedback depending on performance. Brainoware used a three-dimensional brain organoid as a reservoir: it transformed inputs into rich activity, and a separate digital readout was trained to interpret it. DishBrain tested closed-loop behavior; Brainoware tested whether living tissue could serve as a computing substrate.

What are the limitations of Brainoware?

Organoids need continuous life support, vary from one to another and change over time, so results are hard to reproduce exactly. Flat electrode arrays sample only part of a three-dimensional tissue, and the supporting hardware uses far more power than the tissue. The study tested two benchmark tasks and did not address long-term memory or multiple tasks.

Sources

1. Cai et al. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics.
2. Smirnova, Caffo & Johnson (2023). Reservoir computing with brain organoids. Nature Electronics.
3. MIT Technology Review (2023). Human brain cells hooked up to a chip can do speech recognition.
4. News-Medical (2023). Brainoware: a breakthrough AI approach using brain organoids for advanced computation.
5. Jaeger & Haas (2004). Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science.
6. Tanaka et al. (2019). Recent advances in physical reservoir computing: a review. Neural Networks.