Organoid Nodes and Matrix Computation: Living Neural Networks for Hybrid AI

▶ Open interactive ANITEX · equations and animations

This accessible text edition preserves the document's narrative. See the interactive edition for typeset equations, diagrams and playback.

What the Brain-Shaped Nodes Mean

This interactive teaching model represents each organoid module as one brain-shaped node. An organoid is a small three-dimensional neural-tissue model containing many cells; it is not an entire brain and is not a single biological neuron. The four icons are four hypothetical modules, not four human brains. A scalar node value summarizes a population response measured through an interface. The external connections form a matrix-computation diagram. They do not assert that intact organoids naturally arrange themselves into conventional MLP layers. The blue and orange connections are programmable signal couplings in this surrogate; the animation is not a physical wiring diagram or evidence of a functioning multi-organoid device.

A Hybrid Architecture

Sensor inputStimulus encoder Organoid modulesMEA recording Digital MLPOutput. The graphical architecture is 4 sensor channels, 4 biological-surrogate module responses, 4 digital hidden neurons, and 3 output scores. Only the middle digital network is an ordinary MLP. Brain-shaped nodes illustrate biological population processing, while ordinary circular nodes illustrate numerical processing on a computer. One cannot identify an entire organoid with a standard artificial neuron without making a substantial abstraction.

Input Vector and Coupling Matrices

The editable sensor vector has four entries between zero and one. The stimulus encoder uses a 4-by-4 matrix. Rows index receiving nodes; columns index sending nodes. xR^4, W_inR^44, rR^4. d_i=_j=1^4 W_in,iju_j. Every input can contribute to every organoid-module summary. A negative weight subtracts from the numerical drive. It is not a claim that an electrode directly implements an inhibitory synapse.

Biological Dynamics: A Numerical Surrogate

To show memory and a gradual response, the animation uses a bounded recurrent rate model. All quantities here are dimensionless, except the illustrative simulation clock. drdt=-r+ (W_inu +0.35Rr), =0.12\ s. The hyperbolic tangent acts elementwise. The simulator integrates this equation with a fixed Euler step of 0.05 seconds during the stimulate-and-record stages. It holds the recorded response during digital readout. Each demonstration cycle starts with a reset response; this is a software reset, not an experimental procedure for resetting tissue. r_n+1=r_n+ t. The pulsing tissue icons visualize response magnitude. The MEA raster is a deterministic synthetic display with an illustrative rate proportional to response magnitude; it is neither recorded electrophysiology nor a validated spike-generation model. Synthetic spikes and display time are deliberately slowed down for teaching.

Digital MLP: Weighted Sum and Activation

The recorded module responses feed a digital hidden layer. The first weight matrix has shape 4 by 4 and the output matrix has shape 3 by 4. z=W_1r+b_1, a_i=ReLU(z_i)=(0,z_i). q=W_2a+b_2, p_i=(q_i-m)_j=1^3(q_j-m), m=_j q_j. The animation displays moving contributions during each matrix multiplication, then displays ReLU values and normalized outputs. The three outputs A, B, and C have no learned meanings. Fixed illustrative matrices are used, so the device is not recognizing a validated category and its softmax values are not calibrated confidence.

Closed-Loop Feedback Across Two Cycles

The first cycle uses the original sensor input. At the beginning of the second cycle, a simple external controller compares the first output with a teaching target distribution. p_target=[0.7,0.2,0.1]^T, e=p_target-p_cycle\ 1. u=0.3[e_1,e_2,e_3,e_1-e_3]^T, u_cycle\ 2=clip(x+u,0,1). Disable feedback to compare two identical cycles. This controller changes the next input stimulus, not the matrices or tissue parameters. It is not backpropagation into biological tissue, not demonstrated biological learning, and not guaranteed to reduce the output error.

Interactive Animation

Press Play to follow encoding, stimulation, recording, digital hidden-layer computation, softmax output, and feedback. Pause freezes the state. Reset restarts both cycles. Use the timeline to inspect any instant and use the sensor sliders to test another vector. Changing a slider resets the experiment. Expand the matrix panel to inspect every weight and numerical intermediate value. On small screens, scroll the network horizontally.

Research Context and Limitations

Organoid intelligence research investigates combining living neural cultures with electronic interfaces. Brainoware is a published example of brain-organoid reservoir computing using a high-density multielectrode array. That research motivates the biological-reservoir-plus-digital-readout distinction; it does not validate the four-module architecture, equations, weights, timing, or feedback law used in this animation. This single ANITEX source contains theory, equations, and declarative animation data. The NEXMASON viewer must support the organoid-hybrid model to execute it. Ordinary LaTeX tools do not execute the animation block. No executable JavaScript or biological experiment instructions are embedded in the document.

References

Smirnova et al. (2023), Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science. DOI: 10.3389/fsci.2023.1017235. Cai et al. (2023), Brain organoid reservoir computing for artificial intelligence. Nature Electronics 6, 1032--1039. DOI: 10.1038/s41928-023-01069-w.