The mirror map
Every correspondence in one table. Read down a column for one discipline, across a row for one idea expressed three ways.
| Structure | Biological function | In AI | In robotics |
|---|---|---|---|
| The Cellular Machinery | |||
| The Neuron | The neuron integrates its inputs across space and time and emits an action potential when the membrane potential at the axon hillock crosses threshold. Because dendritic branches compute … | The artificial neuron — a weighted sum through a nonlinearity — abstracts this to its simplest possible form. The correspondence that holds: many weighted inputs, one integrated output, … | In robotics the closest match is not a controller but a sensor node in a distributed network: a unit that integrates local signals, applies a threshold, and transmits … |
| The Synapse | The synapse sets how strongly one neuron influences another, and that strength is the storage medium of learning. Strength depends on release probability, vesicle count, and receptor density … | A synapse is a weight. The parallel is genuinely tight — this is the one place where the biological metaphor earns its keep. Hebbian learning ("fire together, wire … | The closest control-theoretic analogue is an adaptive gain: a coupling coefficient tuned online by experience rather than fixed at design time. Adaptive control systems that adjust their gains … |
| Sparse Coding and the Energy Budget | Sparse codes are metabolically cheap, since spikes dominate the energy budget. They are also computationally useful: sparse representations are more separable, easier to read out with a simple … | This is the direct ancestor of sparse autoencoders in mechanistic interpretability. The setup is nearly identical: an overcomplete dictionary, an L1 sparsity penalty, and a reconstruction objective. Where … | Event-driven sensing and sparse state estimation follow the same logic: transmit and process only what changed. On a power-constrained mobile robot this is not an elegance argument but … |
| Circuits and Plasticity | |||
| Population Coding | The represented quantity is recoverable from the population as a weighted vector sum — Georgopoulos showed in the 1980s that the direction of an arm movement can be … | This is the direct biological precedent for representing features as directions in activation space rather than as individual neurons. Broad tuning is polysemanticity; distributed representation over a fixed … | Sensor fusion and probabilistic state estimation: no single sensor determines the state, and the estimate is a weighted combination whose reliability exceeds any individual input. A population vector … |
| Basal Ganglia and the Dopamine Signal | The basal ganglia select among competing actions and learn which selections were worth making. Schultz's recordings showed that midbrain dopamine neurons fire not to reward itself but to … | Temporal-difference reinforcement learning, and therefore the reward-model machinery in RLHF. The actor-critic architecture maps onto the circuit surprisingly well: the striatum as actor selecting actions, the dopamine signal … | Behavior arbitration in a subsumption or behavior-tree architecture: multiple candidate controllers compete, and a selection layer decides which one gets the actuators. The inhibition-by-default design is also good … |
| Lesions, Compensation, and the Limits of Ablation | The method works, within limits that took decades to characterize. Diaschisis: damage to one region disrupts distant regions that depended on its input, so the deficit maps larger … | Ablation studies in interpretability are lesion studies, and every caveat transfers intact. Backup behavior in transformers — where ablating a head causes another to take over its role … | Fault injection and degraded-mode testing: disable a sensor or actuator and characterize what the system does. Good robotics practice already assumes compensation and redundancy, and measures graceful degradation … |
| Cortical Architecture | |||
| The Cortical Column | The column is the cortex's repeated computational unit. Its layered structure implements a consistent division: receive, transform locally, send onward, and send a signal back down the hierarchy. … | A transformer block is also a repeated, uniform unit stacked many times, with a fixed internal division of labor: route between positions, then transform in place, then pass … | A layered control architecture where each level operates on a different timescale and abstraction — reflexes at the bottom, trajectory planning above, task planning above that — with … |
| The Visual Hierarchy | Each stage builds more abstract and more invariant representations from the previous one. Early stages encode local oriented structure; middle stages encode texture and contour combinations; late stages … | Convolutional networks reproduce this progression without being told to. First-layer filters in a trained CNN are oriented edge and color-opponent detectors closely resembling V1 measurements; deeper layers become … | A perception stack layered from features to objects to scene graph to affordances. The engineering motivation is identical — invariance built in stages, so each stage solves a … |
| Attention Networks | Attention allocates limited processing capacity. Attending to a location or feature increases the gain of neurons representing it and suppresses competitors, improving the effective signal-to-noise ratio for the … | Self-attention shares the name and part of the concept — content-based selection of what to read — and the query-key-value formulation is a reasonable abstraction of "search for … | Active perception: a robot with a steerable camera and a compute budget must decide where to look and what to process, and it faces the biological version of … |
| Memory, Motor Control, and Embodiment | |||
| The Cerebellum and Internal Models | The cerebellum learns internal models of the body and its interactions. A forward model predicts the sensory consequence of a motor command, which allows correction to begin before … | Supervised learning with an explicit error signal, and — since the cerebellum is also engaged in cognitive and language tasks, not only motor ones — a candidate for … | This is the mirror the embodied-control track was built for. A cerebellar inverse model does what inverse kinematics does: given a desired end state, produce the commands that … |
| Motor Cortex and the Degrees-of-Freedom Problem | Motor cortex converts intended movement into descending commands, and the population code described earlier is how that intention is represented. The system faces Bernstein's degrees-of-freedom problem: the body … | Overparameterized networks face the same structure: many parameter settings fit the training data equally well, and which one you land on is determined by implicit biases of the … | Null-space projection on a redundant manipulator is the engineering answer to Bernstein's problem: achieve the task with the pseudoinverse, then use the remaining freedom for a secondary objective. … |
| Proprioception: The Sense That Makes Control Possible | Proprioception supplies the state estimate that every internal model requires. Without it, movement is possible but nearly unusable: the rare patients who have lost proprioception through sensory neuropathy … | The nearest analogue is a model's access to its own state. This is thinner than the other mirrors in this section, and the thinness is the point: a … | Joint encoders, force-torque sensors, and IMUs feeding a state estimator. Every robot has this and every roboticist knows the system is only as good as its state estimate. … |
| Hippocampus, Consolidation, and Replay | The hippocampus binds the elements of an experience into an episode that can be retrieved as a unit, and it does so fast — one exposure is often … | Complementary learning systems theory — a fast episodic learner training a slow statistical one — is the direct ancestor of experience replay in deep reinforcement learning, which was … | SLAM — simultaneous localization and mapping — solves the problem place and grid cells appear to solve, and the resemblance was noticed early enough that hippocampus-inspired mapping systems … |
| Working Memory and Its Hard Limit | Working memory holds a small amount of information in an actively accessible state for immediate use. The capacity limit is famously small — around four items, not the … | A recurrent network's hidden state is a maintained buffer with a fixed capacity, and its failure mode — earlier content overwritten by later — is recognizably similar. A … | The state buffer in a behavior tree or task planner: the small set of facts about the current situation that the controller keeps live. Robotics engineers deliberately keep … |
| Failure, Narrative, and the Limits of Analogy | |||
| Confabulation and the Interpreter | The system that generates explanations is not the system that performs actions, and it does not have privileged access to its causes. It builds a plausible account from … | A model asked something outside its knowledge produces the most plausible continuation, fluently and without hedging. A model asked to explain its own reasoning produces a plausible explanation … | A post-hoc explanation module that narrates a controller's decisions from its inputs and outputs rather than from its internal state. Such systems produce satisfying explanations that are not … |
| Homeostasis, Sleep, and Why Grounding Matters | Homeostasis grounds motivation. An organism's goals are not arbitrary: they terminate in physical requirements for continued existence, and reward signals are ultimately calibrated against those requirements. This grounding … | There is no analogue, and the absence is the most important entry in this whole section. An artificial system's objective is written by someone. Nothing anchors it to … | Battery management and thermal limits are the closest thing an artificial system has to a homeostatic drive, and it is instructive how much more robust robot behavior becomes … |
How to read this table
The rows are correspondences, not equivalences. Each full entry carries a "where the analogy breaks" section, and in several cases that section is the most useful thing on the page — the absence of a mirror is itself a finding.