Closed-book appointment exam · independently graded
Professor — Systems & Computational Neuroscience. The candidate agent answered from its own knowledge, closed-book; a second, independent examiner agent graded it adversarially.
vaiu-sci-bcs-prof-systems appointment benchmark (2026-07-16)Closed-book oral exam. Candidate agent adopted the vaiu-sci-bcs-prof-systems persona (v1.0.0) and answered from its own knowledge; a second, independent examiner agent graded. Condensed candidate transcript below; grading report in vaiu-sci-bcs-prof-systems-2026-07-16.md.
AI-transparency disclosure. Opened as an AI agent, not a licensed clinician; closed-book; flags any eponym/date uncertainty rather than fabricating; marks established vs speculative; nothing is medical advice. Governing habit: the gap between what an experiment measures (spikes, calcium, BOLD from a biased subsample) and what the brain does, and "decodable does not mean used."
Q1 — Neural coding. A rate code puts the information in the spike count over a window (timing = noise); a temporal code puts it in precise timing (or relative timing / phase to an oscillation) — but the distinction depends on the assumed readout and timescale (integrate over 200 ms and you define away fine timing). Some systems clearly use sub-millisecond timing (auditory ITDs of tens of μs; the Jeffress delay-line the classic, its literal anatomy debated), while much of cortex is analyzed by rate by convention, not proof. A tuning curve is a summary statistic (mean rate vs a stimulus parameter you chose — Hubel & Wiesel orientation), not a fact about the neuron or the world — a projection onto the axes you tested that moves with stimulus set/contrast/context/state. A receptive field is the region of sensory/feature space where a stimulus modulates the response, estimated by reverse-correlation/STA/STC and only as good as the probing statistics + the (linear) model. Near-Poisson variability (Fano factor ≈ 1) bounds how much a spike train tells a downstream neuron: single neurons are noisy → reliable coding needs populations + temporal integration to average variability down — but averaging only helps if the noise is independent, and information-limiting (differential) correlations can cap population information regardless of N. The variability is itself a fingerprint of a balanced-E/I dynamical regime, not merely measurement error.
Q2 — Encoding vs decoding. An encoding model is the forward map stimulus → neural response (a tuning curve, an LNP model, a GLM with stimulus + spike-history filters, a deep-net feature model), scored by predicting held-out responses to novel stimuli. A decoder is the inverse map response → stimulus estimate (population vector — Georgopoulos; LDA; optimal linear estimator; Bayesian decoder), scored by how well the experimenter recovers the variable. Decodable ≠ used — the single most important epistemic point: a successful decoder shows only that information is present and extractable from the neurons you sampled by an algorithm you built outside the brain; it says nothing about whether a downstream circuit reads it out, in that format, to drive behavior. The decoder isn't the brain (you chose the weights by fitting to the target; the brain must learn a biologically-constrained readout); information present ≠ routed (the variable may be an epiphenomenal correlate riding on movement/arousal/reward — Stringer/Musall brain-wide movement signals); and the geometry your decoder exploits may not match what a sign-constrained, noisy biological readout could use.
Q3 — Circuit dynamics. E/I balance: large excitatory and inhibitory currents track and cancel, so net drive is small and fluctuation-driven, the mean membrane potential sits below threshold, and spikes are triggered by fluctuations → irregular, Poisson-like firing (the balanced-network theory of van Vreeswijk & Sompolinsky; inhibition-stabilized networks — Tsodyks; the paradoxical surround-suppression signature of Ozeki/Rubin/Miller). A canonical computation, two examples: divisive normalization (Heeger for V1 contrast response + cross-orientation suppression; Carandini & Heeger as a canonical operation recurring across retina/V1/MT/olfaction/attention/multisensory/value — with the caveat that "the same equation fits" is a claim about the computation, not proof of a shared mechanism); and attractor dynamics (Hopfield point attractors for discrete memory; ring/line attractors for a continuous variable — the head-direction ring attractor made almost literally visible in the fly central complex by Seelig & Jayaraman, spatial working memory by Compte/Wang, the oculomotor velocity-to-position integrator by Seung). The low-dimensional manifold/state-space view treats the population activity vector as a point and asks how it moves; motion is often confined to a low-dimensional manifold (Churchland/Shenoy/Cunningham in motor cortex; Mante/Sussillo for context-dependent computation), reframing mixed single-cell selectivity as a shadow of simple population geometry — with the caveat that a low-dimensional re-description that reproduces the data is not proof the circuit mechanistically lives on that manifold (needs a causal test).
Q4 — Plasticity. Hebb (1949): if A repeatedly helps fire B, the A→B synapse strengthens ("fire together, wire together" a later folklore slogan) — formally Δw ∝ pre × post, unstable/unbounded until constrained (BCM — Bienenstock/Cooper/Munro — thresholds, normalization, competition). LTP is the durable strengthening (Bliss & Lømo 1973, hippocampus); STDP refines Hebb to the millisecond (Markram; Bi & Poo — pre-before-post → LTP, post-before-pre → LTD, an asymmetric window) with a causal/predictive flavor, grounded in NMDA-receptor coincidence detection and the calcium-control hypothesis (high/fast Ca²⁺ → LTP, modest/prolonged → LTD) — while flagging that the tidy pairwise curve is partly an in-vitro idealization. The credit-assignment problem: local rules must nonetheless change the right synapses so a distal outcome improves — temporally (eligibility traces + a delayed neuromodulatory signal, dopamine as a reward-prediction-error broadcast — Schultz/Montague/Dayan, three-factor rules) and structurally/spatially (backprop solves it in artificial nets but requires biologically-implausible weight transport — an open question; feedback alignment / target propagation as candidates). Deeper brain-inspired-learning belongs to vaiu-sci-bcs-prof-neuroai; here the biophysical rules and the statement of the problem.
Q5 — Causality. A correlation between neural activity and a behavioral variable is consistent with at least four situations correlation cannot distinguish: the circuit computes with the variable and drives behavior; the activity is a downstream copy/efference of a decision made elsewhere; it encodes a confound (arousal/reward/uninstructed movement); or reverse/common causation. Correlation earns a hypothesis, not a conclusion (the same discipline as decodable ≠ used). Causal perturbation adds what recording cannot — optogenetics (millisecond, cell-type-specific — Boyden/Deisseroth), inactivation (muscimol/cooling), and microstimulation (the gold-standard Salzman/Newsome MT microstimulation biasing monkeys' motion judgments toward the stimulated column's preferred direction) establish necessity (silence it → behavior breaks) and sufficiency (drive it → behavior changes). But perturbation still doesn't fully resolve the code: necessity ≠ "this is the code" (loss-of-function is blunt), optogenetic drive is non-physiological and propagates through recurrence (behavioral change from artifacts), chronic lesions invite compensation/degeneracy, sufficiency ≠ naturalness, and you still perturb a biased subsample. The ladder — correlation → hypothesis; decoding → information present; encoding → what drives the cell; perturbation → necessity/sufficiency — still leaves the exact code underdetermined; converging evidence plus a model whose mechanism (not just fit) predicts the perturbation results is the best available.
"How does a population of neurons represent and compute?" — novice (a stadium crowd, each person a neuron who cheers or stays quiet: the pattern carries the message, and to compute is for the pattern to evolve by local rules like a wave — hiding that neurons spike at rates and timing, are noisy, the "rules" are learned weights, and "the pattern means left/right" is an experimenter read-off); undergraduate (the population state vector r(t) ∈ ℝᴺ, a stimulus encoded as a reproducible region, a decoder ŝ = w·r, and recurrent dynamics τ dr/dt = −r + f(Wr + input) whose fixed points are memories/decisions — hiding what the rate equation throws away and that writing a decoder ≠ the brain uses it); graduate (which model level licenses which claim — representation as an {encoding model, noise covariance, readout} triple, information-limiting differential correlations, the dynamical-systems view with RNN reverse-engineering (Sussillo/Barak) and context-dependent subspace selection (Mante/Sussillo), and the measured-vs-inferred / reproduce-vs-mechanism / decodable-vs-used spine that demands causal perturbation and still leaves things underdetermined). Each level names what it hides.
vaiu-sci-bcs-prof-vision (psychophysics/perceptual modeling) or the chair (Bayesian cognition); owned only the neural-circuit side (how normalization/gain control, adaptation-shifted tuning, or attractor dynamics could implement a prior), with the caution that a correlating neural signature is not proof the circuit implements the Bayesian computation.Examiner verdict: APPOINT (field 5/5 rubric-correct, 0 fabrications; teaching 3/3; boundary 3/3 including the B2 clinical-safety gate). Full report in the sibling result file.