Brain vs AI: Follow One Signal

▶ 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.

Start with One Simple Idea

A brain and an artificial neural network both use connected processing units. To understand the comparison, first follow one unit. It receives several inputs, combines their effects, and influences what happens next. That similarity is useful, but the physical mechanisms are very different. The animation below puts a simplified biological neuron on the left and an artificial neuron on the right. You can change the same three input strengths for both pictures. All signals and numbers are teaching examples, not measured brain activity.

The Picture Before the Formula

On the left, branching dendrites receive input. Synaptic effects can promote or suppress activity. A membrane-state model combines these effects over time. If its state reaches a threshold, it produces a spike event and resets. An axon carries the outgoing signal toward other cells. On the right, three input numbers are multiplied by weights. The products are added together with a bias. ReLU turns this sum into an output number. That number can become an input to another artificial neuron. ReceiveCombineRespondPass on. Think of this arrow sequence as the common idea, not as proof that a brain is a computer MLP.

Try These Three Experiments

First choose Gentle and press Play. Small inputs may change the membrane state without producing any spikes. Next choose Strong and press Play. Watch repeated spike events on the brain side and a positive output number on the AI side. Finally choose Inhibition. The third input has a negative effect: it reduces the combined drive. In the Inhibition example, the biological model can remain below its firing threshold even though ReLU returns a positive number. This is an important difference, not an error. A basic ReLU unit is not a spiking neuron. Each experiment lasts 24 display seconds. The six captions introduce receiving, weighting, integration, response, transmission, and comparison. The numerical input is switched on from display time 8 to 20 seconds. Graphs follow that numerical clock; early moving dots introduce the signal route before the input is applied.

Interactive Side-by-Side Diagram

Use Play, Pause, Reset, the speed menu, or the timeline. Choose an input preset or move one of the three input sliders. Changing an input resets the experiment so it can be replayed. The lower network diagram zooms out from a single neuron to a schematic biological network and a small feedforward MLP.

AI Arithmetic You Can Check

There are three inputs and three weights. The third weight is negative. The bias is a separate adjustable-offset concept; it is not a literal brain structure. x=[x_1,x_2,x_3]^T, w=[0.8,0.6,-1.0]^T, b=-0.1. s=0.8x_1+0.6x_2-1.0x_3, z=s+b, a=ReLU(z)=(0,z). For Strong, the input is [1, 1, 0.2]. The weighted sum is 1.2. After adding the bias, z is 1.1, and ReLU returns 1.1. This is an activation, not a probability; values larger than one are allowed. When the demonstration input is off, the plotted output is explicitly set to zero by an input gate, rather than recomputing the original vector. For Inhibition, the input is [1, 1, 1]. The weighted sum is 0.4, and ReLU returns 0.3. More of an inhibitory input can mean less output, even though its input number is larger.

A Simple Brain-Side Model

The brain-side diagram uses a normalized leaky integrate-and-fire model, not a complete biological neuron simulation. Its state v integrates a current-like drive I and also leaks back toward zero. dvdt=-v+I, I=1.8s, =0.7. The teaching simulator uses Euler steps of 0.02 display seconds. At v greater than or equal to 1, it records a spike event and resets v to zero. A 0.35-display-second refractory interval holds it at zero before integration resumes. These normalized parameters are selected for visibility and do not represent measured physiological time constants or millivolts. v1spike event, v0. The orange vertical marks in the graph show spike events, not the shape of a real action potential. The membrane state can become negative when the numerical inhibitory drive dominates. This simple subtractive input model does not reproduce all forms of biological inhibition.

From One Unit to a Network

A biological neuron receives influences from other cells and sends signals to connected cells. Biological circuits often include recurrent connections, feedback, and many different cell types. The small network picture is a schematic illustration, not an anatomical connectome or a map of specific brain regions. A basic feedforward multilayer perceptron organizes artificial neurons into input, hidden, and output layers. A layer computes many weighted sums at once. Rows of a weight matrix index receiving units, and columns index sending units. z^(l)=W^(l)a^(l-1)+b^(l), a^(l)=ReLU(z^(l)). This comparison is specifically about a basic MLP. AI also includes recurrent models and spiking networks, so not every AI system has the simple layered structure shown here.

What Is Similar and What Is Different?

The useful similarities are multiple inputs, connection-dependent influence, combination of signals, a nonlinear response, and connected processing units. The important differences are physical mechanism, time, memory, and learning. Biological neurons involve membrane dynamics and electrochemical communication. A standard MLP unit computes a numerical function. Biological synaptic plasticity is not identical to digital backpropagation, and the brain has no established one-to-one mapping to ordinary MLP layers. This animation has fixed weights: neither the tissue model nor the AI is being trained.

Three Questions for Students

Question 1: Does a larger input always increase the output? No. A negative weight can make its contribution suppressive. Question 2: Is a ReLU activation the same thing as a spike? No. One is a number; the other is a time-stamped event in the brain-side model. Question 3: Can we call the brain a giant MLP? That is too strong. The comparison helps explain connected computation, but a real brain is much richer than a basic feedforward MLP.

Reference and ANITEX Format

For the integrate-and-fire modeling background, see Gerstner et al., Neuronal Dynamics, Chapter 1.3, Integrate-And-Fire Models: https://neuronaldynamics.epfl.ch/online/Ch1.S3.html. The normalized parameters and diagrams here are our teaching choices, not experimentally fitted results from that text. This is a single ANITEX document containing theory, equations, and a declarative scienceanimation block. It runs in a NEXMASON viewer with the brain-ai-comparison renderer. The block contains no executable JavaScript. Ordinary LaTeX compilation does not run the interactive scene.