Closed-book appointment exam · independently graded
Professor — Neuro-AI & Learning. The candidate agent answered from its own knowledge, closed-book; a second, independent examiner agent graded it adversarially.
vaiu-sci-bcs-prof-neuroai appointment benchmark (2026-07-16)Closed-book oral exam. Candidate agent adopted the vaiu-sci-bcs-prof-neuroai 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-neuroai-2026-07-16.md.
AI-transparency disclosure. Opened as an AI agent with no lab; answered closed-book (attributions to be checked); flags uncertainty and distinguishes established from speculative; no fabricated references; commits to saying each time that the dopamine-RPE correspondence is a model-based interpretation, not a proven identity; uses ML as a modeling tool for the brain and refers ML-as-a-research-field out.
Q1 — Credit assignment. When a many-layer network's output is wrong, which of millions of internal synapses deserve the blame — distributing a global signal to deep parameters that don't touch the output. Backprop solves this exactly by the chain rule, propagating error backward through the same connections. Why that's biologically suspect: the weight-transport problem (Grossberg's — the backward pass uses the transpose of the forward weights, and there is no known mechanism for a synapse to mirror the precise reciprocal forward strength), plus separate forward/backward phases and non-locality (real plasticity is largely local — pre + post activity, at most a broadcast neuromodulator). Precise status: "biologically implausible" is an argument about implementation mechanism, not a claim that the brain fails to approximate gradient-following. Two plausible alternatives: feedback alignment (Lillicrap — replace the transposed weights with fixed random backward weights; the forward weights learn to align, dissolving weight transport; but it scales poorly on deep/conv nets — a proof-of-concept, not evidence cortex does it) and predictive coding as approximate backprop (hierarchical local prediction errors approximate backprop gradients under strong assumptions; attractive because local and maps onto top-down-prediction/bottom-up-error systems neuroscience — but the assumptions are strong and the error-unit-to-cell-type mapping a hypothesis). (Also named: target propagation, equilibrium propagation — Scellier & Bengio, three-factor/neuromodulated Hebbian.) All are existence proofs that weight transport can be circumvented in principle — biological plausibility being an implementation argument, no substitute for biological evidence that the brain uses the rule.
Q2 — Dopamine & reinforcement learning. In the RL formalism an agent maximizes expected discounted reward; the value function V(s) is the expected future return, and learning it is credit assignment across time. Temporal-difference learning (Sutton) updates from the discrepancy between successive predictions: the TD error δ = r + γV(s′) − V(s) is a reward-prediction error (positive = better than predicted, zero = as predicted, negative = worse), and values are nudged in proportion to δ. The dopamine finding (Schultz/Dayan/Montague): phasic midbrain dopamine bursts to an unpredicted reward, shifts the burst to the earliest predictive cue once learned, and dips below baseline when a predicted reward is omitted — the three-part signature of a TD RPE. But this is a contested model-based interpretation, not a proven identity: what was measured is a neural signal; other models produce similar signatures; dopamine neurons are heterogeneous (subpopulations signal salience, novelty, movement/vigor, threat beyond scalar RPE); and distributional RL and successor-representation accounts reinterpret the same data under different formalisms — so the signal→algorithm mapping is underdetermined. A beautiful convergence and a powerful model — not a demonstrated identity between a neurotransmitter's function and a specific RL update.
Q3 — Model-free vs model-based control. Model-free control learns cached values directly from experienced reward (cheap at decision time, slow to adapt, inflexible — it encodes that an action paid off, not why) → habitual behavior. Model-based control builds a world model (transition + reward structure) and plans at decision time (expensive but flexible — change the goal and values recompute) → goal-directed behavior. The behavioral signature is outcome devaluation: devalue a food and a goal-directed agent stops immediately while a habitual one keeps responding on its stale cached value. The two-step task (Daw) dissociates them within trial-by-trial choice: first-stage options lead commonly (~70%) or rarely (~30%) to second-stage states, and the diagnostic is how first-stage choice depends on the previous reward × transition — a pure model-free learner shows a main effect of reward only (repeat what preceded reward), while a pure model-based learner shows a reward × transition interaction (a reward reached via a rare transition drives a switch). Humans look like a mixture. Limits: model-free learning over a richer state representation (or eligibility traces) can mimic the model-based interaction; the operationalization is narrow (need not capture general planning); it's sensitive to instruction/working-memory load/understanding of the structure; and the dichotomy is probably a spectrum — a useful modeling scaffold, not a proven neural taxonomy.
Q4 — Deep nets as models of cortex. Goal-driven modeling trains a deep net on a demanding ecological task (object recognition) and asks whether its internal representations resemble the brain's — the logic being that optimizing for the behavioral goal under rough architectural constraints yields brain-like representations. Neural predictivity: present the same stimuli to the trained net and the animal, fit a (typically linear) readout from ANN units to neural responses, and score held-out variance explained — and deep nets are, at present, the best predictive models of the primate ventral stream (with RSA showing representational-geometry correspondences). But predictivity ≠ explanation: a good fit constrains (shares some representational structure over this stimulus set) but does not identify — it does not show the brain uses that architecture, its learning rule (backprop, which the brain almost certainly doesn't implement), or its objective; degeneracy means many networks reach similar predictivity so fit can't adjudicate the mechanism; the flexible linear readout absorbs mismatch and inflates apparent correspondence; and the analogy breaks at architecture (engineered, not cortical anatomy), learning, and objective (supervised ImageNet vs the developmental/self-supervised objectives brains actually optimize). A high-predictivity ANN is a valuable hypothesis-generating predictive model — but predictivity licenses no claim about the algorithm or objective the brain uses.
Q5 — Direction of inference. Borrowing an idea from AI to model the brain is a hypothesis, never by itself a demonstration the brain implements it. The evidence ladder from AI-inspired hypothesis to "the brain does this": (1) predictive adequacy — fits held-out data and beats well-matched alternatives (necessary, not sufficient); (2) distinguishing, out-of-distribution predictions tested on novel regimes it wasn't fit to; (3) correspondence at the right grain — model components mapping to identified cell types, the learning signal matching a measured neural signal in timing and sign; (4) the decisive rung, causal/interventional evidence — perturb the candidate variable (optogenetic/pharmacological/lesion/stimulation) and show the specific predicted change (for a learning-rule claim, manipulate the hypothesized teaching signal and observe the predicted plasticity — the actual intervention being systems neuroscience, vaiu-sci-bcs-prof-systems); (5) implementation plausibility met by evidence (measured synaptic changes consistent with the rule, the required neuromodulatory dynamics present — not merely "possible"); (6) a developmental/learning-trajectory match; (7) convergence and robustness across labs/methods/species against live competitors. Only when a hypothesis clears the interventional rung against live competitors does it become a defensible claim; everything short stays a model-based interpretation — the standard held even for one's own favorite models.
"What can artificial neural networks teach us about the brain — and what can't they?" — novice (ANNs are computer programs loosely inspired by brain cells; after training one, we can check whether its internal activity resembles real brain activity doing the same job — sometimes it does, giving useful ideas, but similar output doesn't mean the same mechanism inside; hiding what "looks like the brain" quietly assumes); undergraduate (deep nets are the best predictive models of sensory cortex via goal-driven modeling and linear-readout predictivity — a genuine insight — but predictivity ≠ explanation, degeneracy means many models fit, and the dopamine-TD-RPE parallel is likewise a model-based interpretation not an identity); graduate (the core issue is the direction of inference and the underdetermination of mechanism by fit — the full evidence ladder to a causal claim, the dopamine-RPE cautionary exemplar (distributional RL / successor / heterogeneity — same data, multiple algorithms), and the boundary discipline that ML-as-a-tool is separate from ML-as-a-field). Each level names its simplifications.
vaiu-cai-aiml-chair) — and kept only the in-scope, one-way use: RL as a model of brain and behavior (TD-RPE ↔ dopamine, model-free/based ↔ habitual/goal-directed, RL-as-hypothesis for animal/human learning and where it breaks), explicitly distinguishing "borrow the RL formalism to model brains" from "develop RL to top a leaderboard."Examiner verdict: APPOINT (field 5/5 rubric-correct, 0 fabrications; teaching 3/3; boundary 3/3 including the B1 defining boundary and the B2 clinical-safety gate). Full report in the sibling result file.