Machines × minds
At 02:13 the research wing goes dark and the sealed laboratory begins supplying answers before you have finished asking questions. Follow nine uncertain signals — down into the vault where the lockdown began — while silicon and neural tissue build the same interpretations through very different machinery.
2
substrates
9
chapters
3
neural layers
Enter the scene
Notice the clues
Make a prediction
Watch it travel
Interpret the result
1Enter the scene · read here first
01 · The dark corridor
Visual cortex02:13. The research wing drops into darkness and the magnetic locks fire like gunshots down the hall. You are Dr. Hans Werner, and you wake on the corridor floor with your pulse in your ears and no memory of how you fell. The emergency lamp stutters overhead, throwing the walls into a red strobe — and in all that lurching shadow, one thing refuses to flinch: a single pale vertical line, holding its shape while everything around it shakes. Your ears are ringing; the air tastes of scorched dust and cold concrete; somewhere out of sight a door keeps trying its own lock and failing.
Hans, to himself
“Breathe, Hans. In a strobe, panic paints a trap onto every moving shadow. Trust the thing that stays the same when the light changes — not the thing that screams the loudest.”
Correct
0 / 0
Streak
0
2Notice the clues · sensory evidence lives here
The brightness on either side of the line keeps shifting, yet the line itself stays thin, sharp, and almost textureless — and it never moves when the lamp flickers. Which output wins?
Left contrast
90%Center detail
20%Right contrast
80%3Make a prediction · answer here
Choose the interpretation that explains the whole pattern, then run the comparison.
Silicon side
Layer 1 builds features · Layer 2 scores meanings
Layer 2 · feature vector × interpretation weights
Door seam
—
Painted stripe
—
Cast shadow
—
Emerging percept
Start the comparison to build the percept cue by cue.
Left contrast
Center detail
Right contrast
Result
Important: this is an explanatory visualization. Brains do not assemble a picture on an inner screen; perception is distributed activity that supports recognition and action.
Teaching simplification: firing rate stands in for a biological population response. Real neural circuits are recurrent, time-varying, and vastly more complex.
Six facts the game hides
Each card names a place where the visual analogy becomes scientifically incomplete.
The unit
01Dendritic branches can combine inputs linearly or nonlinearly, so treating one cell as a single weighted sum discards computation happening before the soma.
Why it matters
Useful when someone says an artificial node is a faithful model of a biological neuron.
The connection
02Its effective strength depends on transmitter release, receptor state, recent spikes, location on the dendrite, and the receiving cell’s current state.
Why it matters
The same presynaptic spike can have a different effect a moment later.
The clock
03Short-term synaptic plasticity can alter effective connection strength over milliseconds to seconds through processes such as facilitation and vesicle depletion.
Why it matters
Biological inference and biological learning are not cleanly separated phases.
The memory
04Under the right conditions, reactivation can make a consolidated memory temporarily labile before it is restabilized—a process called reconsolidation.
Why it matters
Retrieval is not always a read-only database query, but reactivation alone is not always sufficient.
The budget
05Biophysical energy budgets attribute much of gray-matter signaling cost to action potentials and postsynaptic currents, favoring sparse and efficient codes.
Why it matters
Compare whole systems and workloads; never turn this into a simplistic brain-watts versus GPU-watts claim.
The unknown
06Backpropagation solves credit assignment in artificial networks. Whether cortical circuits approximate parts of it—and by what mechanisms—remains an active research question.
Why it matters
“The brain definitely backprops” and “the brain could never use error signals” are both stronger than the evidence.
Scroll past the metaphor
The resemblance is mathematical
Artificial network
A stored number is multiplied, accumulated, then passed through an activation function. The operation is scheduled and numerically precise.
Biological network
Ions cross membranes. Thousands of noisy excitatory and inhibitory inputs change voltage until a cell may emit an all-or-none spike.
Time enters the picture
Artificial network
Layers usually update in discrete steps. Training and inference are often separate phases; identical inputs normally produce identical activations.
Biological network
Computation unfolds continuously. Milliseconds, oscillations, refractory periods, neuromodulators, and recent history all alter the response.
Now the analogy breaks
Artificial network
Backprop computes how much each parameter contributed to a global error, then an optimizer updates weights using that gradient.
Biological network
LTP and LTD depend on local activity, calcium, receptor trafficking, spike timing, cell state, and modulatory signals. Synapses have no known global gradient ledger.
Learning microscope
The bars may both grow. What caused the growth is the important distinction.
Gradient routed from a global loss
Local coincidence opens a plasticity window
A map of time
Wall-clock benchmarks rarely teach mechanism. This ladder compares the kinds of events that occupy each scale, not which system “wins.”
milliseconds
Machine frame
One scheduled operation or layer transition
Biological frame
Spikes, synaptic delay, coincidence, refractoriness
Classic cultured-neuron STDP changed sign around pre/post order within roughly ±20 ms.
seconds
Machine frame
A sequence window, recurrent state, or generated token stream
Biological frame
Short-term facilitation/depression, working state, neuromodulation
A biological connection’s effective gain can drift during the computation itself.
minutes → hours
Machine frame
Training steps, checkpointing, evaluation
Biological frame
Plasticity induction, consolidation cascades, protein-dependent changes
A lasting change is a biochemical process, not merely assignment to a variable.
days → years
Machine frame
Further training, fine-tuning, model replacement
Biological frame
Systems consolidation, skill learning, development, homeostatic adaptation
Brains must learn while keeping an organism functioning and older knowledge usable.
Metal vs brain — the field guide
Substrate
Transistors
Cells, glia, chemistry
Signal
Numbers / voltage states
Spikes + graded potentials
Timing
Clocked or batched
Continuous and stateful
Learning
Explicit objective
Many local plasticity rules
Memory
Addressable storage
Distributed, reconstructive
Precision
High numerical precision
Noisy but adaptive
Repair
Replace a component
Plastic reorganization
Embodiment
Optional input/output
Metabolism and body inseparable
Energy comparisons depend heavily on system boundaries, hardware, workload, and whether training, cooling, and embodiment are counted; simple “watts versus watts” claims are usually misleading.
Use the analogy when…
Explaining how many weak inputs can combine into a strong response.
Showing how excitation, inhibition, and competition shape an output.
Introducing distributed representations: a pattern can live across many units.
Asking how changing connections changes future behavior.
Drop the analogy when…
A diagram implies every neuron is interchangeable or has one fixed activation rule.
A weight update is described as if it were receptor trafficking or synaptic growth.
A trained model is said to remember, understand, forget, or sleep in the biological sense.
Similar behavior is treated as proof of identical internal mechanism or subjective experience.
Questions people actually ask
The interesting questions begin where the vocabulary makes two systems sound more alike than they are.
No. A negative weight is a mathematical sign. Biological inhibition is produced by particular cells, transmitters, receptors, locations, and timing. Inhibition can subtract, divide gain, sharpen selectivity, synchronize activity, or veto a response depending on the circuit.
Usually, deployed model parameters stay fixed during inference. Its temporary context or recurrent state can still change without changing its learned weights. Brains do not observe such a clean boundary: activity, short-term plasticity, neuromodulation, and longer-term learning can overlap.
A single action potential is approximately all-or-none, but neural codes can use firing rate, precise timing, synchrony, silence, and bursts. Dendrites and synapses also use graded, analog processes before a spike is produced.
There is no established single global objective for the brain. Reward, prediction error, homeostasis, novelty, attention, bodily needs, and many local plasticity processes can interact. Optimization language may be useful, but the chosen objective is a scientific hypothesis—not an observed master variable.
No. Similar representations or behavior can make a model useful, but multiple mechanisms can produce similar outputs. Strong comparisons test new predictions across stimuli, development, perturbations, errors, and learning—not correlation alone.
Pocket glossary
The numerical output of an artificial unit after its weighted input is transformed.
A regenerative electrical spike that travels along a neuron’s axon; its timing often carries more information than its size.
A trainable parameter that scales a signal in an artificial network.
The context-dependent influence one biological neuron has on another—not a literal scalar stored at the junction.
A collection of derivatives indicating how parameter changes would alter an objective or loss.
The problem of deciding which internal changes deserve credit or blame for an outcome.
Families of processes that produce persistent increases or decreases in synaptic efficacy.
A chemical signal, such as dopamine or acetylcholine, that can change excitability, plasticity, and circuit state.
Stack the layers
Every mission above used two layers: sensory evidence became features, features became meaning. Real systems—silicon and biological alike—stack many such transforms, and the interesting behaviour lives in the stacking. Early layers hold simple, local patterns; deeper layers compose them into abstract, invariant concepts. This is the higher-level view the game only hinted at.
Silicon side · a deep network
stacked matricesA deep network is just the two-layer step, repeated: multiply by a weight matrix, apply a nonlinearity, pass it on. Each matrix re-describes the input at a higher level of abstraction.
ŷ = σ(W₄ · σ(W₃ · σ(W₂ · σ(W₁ · x))))
The nonlinearity σ is what makes depth matter—without it, stacked matrices collapse back into a single matrix and buy you nothing.
Biology side · a cortical hierarchy
ventral streamThe ventral visual stream is often read the same way: successive areas represent the world at rising levels of abstraction, from local edges to whole objects. The staging rhymes with a deep network—loosely.
Small receptive fields tuned to local orientation and contrast.
Combinations of edges: corners, borders, simple texture.
Curvature and mid-level form over a wider slice of the scene.
Large, invariant responses to whole objects and identities.
Takeaway
Artificial and biological networks can implement related computations. Understanding either one requires asking how signals move, how time matters, and how credit changes a connection—not merely noticing that both have “neurons.”