Past exam of the cognition course of the year 2 of the psychology (experimental) course of the University of Oxford / 2022 (A16817W1, A16817X1, Part I)

Worked essays for the historical Part I Cognition paper dated 19 April 2022, covering all five choices. Source release: Oxford psychology examination release, member a16817w1.pdf. The cover codes are retained; the last-page footer transposes digits in its X1 code. Each answer develops an argument within the printed 1,500-word limit. The research discussed below was published before the examination; no later findings are silently substituted for the evidence then available.

Delay-period activity explains important properties of working memory when its information content and dynamics predict behavior; elevated activity alone is not a complete storage mechanism. An explanation must account for maintaining an absent stimulus, limited working memory capacity, variable precision, interference and flexible prioritization. Different measurements reveal different aspects: extracellular single-unit recording measures individual neurons, electroencephalography measures aggregated electrical signals, and functional magnetic resonance imaging measures a slower hemodynamic response. None directly reads a unitary store.

The classic case for persistent neural activity came from delayed-response tasks. In Shintaro Funahashi and colleagues' 1989 experiment, monkeys maintained fixation after a briefly presented peripheral cue and later made a memory-guided saccade. Dorsolateral prefrontal cortex neurons could maintain location-selective firing during the intervening blank interval; delay-period activity was reduced or absent on some error trials. This connects a maintained code to a remembered location rather than merely to stimulus presence. It also supplies a plausible bridge between a short-lived sensory event and a later response. However, the remembered location and intended eye movement covaried, so the experiment does not isolate human mnemonic storage from prospective action coding.^[ref]

More explanatory leverage comes from linking population dynamics to the pattern of errors. A bump attractor maintains a localized pattern through recurrent excitation and inhibition. Noise can move its center while leaving a memory trace present, predicting increasing uncertainty and trial-specific directional errors. Klaus Wimmer and colleagues tested such predictions against monkey prefrontal recordings and memory-guided eye movements: neural variability was related to fine-grained response errors in ways predicted by a diffusing representation. This is stronger than a generic activation–accuracy association because competing models need to explain that structured relationship. Nevertheless, it is evidence from a trained spatial task in monkeys; it does not by itself explain verbal rehearsal, binding errors or every source of human forgetting.^[ref]

Human recordings link delay signals to working memory capacity. In Edward K. Vogel and colleagues' visual-memory experiments, people remembered arrays on one side of fixation while ignoring the other. The contralateral delay activity increased with memory load and approached an asymptote associated with each person's behavioral working memory capacity. Its usefulness is the relation between a physiological load function and an independently estimated behavioral limit. It does not prove that the brain has exactly a fixed number of universal slots: stimulus complexity, filtering, allocation and the measurement's sensitivity can influence the signal. A plateau is evidence constraining a working memory capacity mechanism, not the mechanism itself.^[ref]

Delay-period activity also contains specific sensory information. Stephenie A. Harrison and Frank Tong presented oriented patterns and cued which orientation to remember. Multivariate pattern analysis of early visual cortex functional magnetic resonance imaging could identify the remembered orientation during the delay, even when mean activation was weak. This supports sensory recruitment and helps explain why visual memories retain feature-specific detail and may compete with new sensory input. Decoding establishes accessible information, however, not necessarily a local storage site: top-down feedback, attention or task preparation could generate a decodable pattern. Its slow hemodynamics cannot establish uninterrupted moment-to-moment firing.^[ref]

The temporal and priority structure of working memory challenges an exclusively tonic account. Mikael Lundqvist and colleagues recorded monkey prefrontal spiking and local field potentials and found brief gamma oscillation and beta oscillation bursts rather than uniformly sustained oscillatory activity. A smooth trial average can conceal intermittent events. This finding qualifies the interpretation of averaged persistence; it does not show that every neuron becomes silent between bursts, or that recurrent activity is irrelevant.^[ref]

Human perturbation experiments distinguish current priority from continued retention. In Nathan S. Rose and colleagues' task, a temporarily unattended item could cease to be decodable yet remain relevant for a later probe. A transcranial magnetic stimulation pulse reinstated a detectable representation under appropriate relevance conditions. In Michael J. Wolff and colleagues' experiment, a neutral visual impulse during the delay evoked an electroencephalography response carrying information about remembered items, including temporarily unattended ones; instructed forgetting and temporary inattention had different consequences, and the hidden-state measure related to working memory precision. These results explain how an item may remain available without occupying the same active state as the current focus of attention.^[ref][ref]

An activity-silent working memory interpretation proposes temporarily altered network responsiveness, potentially involving short-term synaptic plasticity. Neither an absent decoding result nor a successful perturbation identifies that molecular mechanism uniquely: weak or recoded activity can evade a particular measurement, and perturbation can recruit distributed networks. The strongest conclusion is therefore pluralistic. Persistent neural activity offers a plausible mechanism for some maintenance and precision effects, while intermittent and latent states help explain prioritization and retained but unattended information. A successful theory must jointly predict errors, load effects and perturbation consequences at the appropriate time scale. Delay-period activity provides substantial mechanistic constraints, but human working memory is better explained by interacting representations and control states than by equating storage with continuously elevated firing.

Converging evidence supports partially distinct, interacting dorsal and ventral language pathways; it does not support two isolated, universally mapped cables. The central computational distinction is between mapping heard speech onto articulatory plans and mapping it onto meaning. Gregory Hickok and David Poeppel's dual-stream framework assigns a stronger leftward bias to the dorsal sensorimotor route and a more bilateral organization to ventral speech-comprehension processes. “Parallel” means that an acoustic input can engage both transformations rather than passing through a single obligatory serial chain. It does not mean identical timing, complete independence or that each linguistic operation belongs to only one route.^[ref]

The dorsal language pathway connects posterior auditory–temporal and temporoparietal processing with frontal articulatory systems, using pathways that include the arcuate fasciculus and superior longitudinal system. It is particularly useful when sound must guide an unfamiliar articulatory sequence, as in nonword repetition, and for auditory–motor interaction. The ventral language pathway links temporal representations with systems supporting lexical and conceptual interpretation, with temporofrontal connections through ventral white-matter systems including the extreme capsule. These labels name functional–anatomical families. They should not be equated without qualification with particular textbook stream diagrams from vision, or with a claim that language has only one dorsal and one ventral anatomical connection.

An informative test must connect differences in task demands with differences in connectivity. Dorothee Saur and colleagues combined functional magnetic resonance imaging with diffusion-based tractography. They used sublexical repetition and higher-level comprehension tasks to identify functionally relevant regions, then estimated connections among those regions. Repetition-linked temporal and premotor regions were connected dorsally through the arcuate/superior longitudinal system; comprehension-linked temporal and ventrolateral frontal regions were connected by a ventral route through the extreme-capsule region. The study therefore supplied anatomical evidence for the proposed computational distinction rather than inferring pathways from activation maps alone.^[ref]

This design also has limitations. Nonword repetition places demands on phonological working memory and overt motor output, whereas a comprehension task adds lexical, semantic and decision demands. An activation difference could reflect any correlated component. Moreover, diffusion-based tractography estimates plausible paths from water diffusion; it does not directly observe axons, synaptic direction, information flow or a pathway's functional necessity. Crossing fibers, seed selection and modeling assumptions affect reconstructions. Thus the evidence is stronger when task effects and anatomy converge, but the labels “sound-to-articulation” and “sound-to-meaning” remain hypotheses to test with other methods.

Lesion studies offer a complementary test because intact performance after damage can challenge claims of necessity and selective deficits can identify vulnerable transformations. Dorothee Kümmerer and colleagues studied one hundred patients with acute left-hemisphere aphasia, deriving repetition and comprehension measures from language tests and relating them to damage. Repetition deficits were associated more strongly with dorsal temporoparietal connections, while comprehension deficits implicated ventral temporal–frontal connections. This convergence with healthy-participant imaging supports a functionally differentiated network. Studying patients soon after injury limits some effects of long-term reorganization, although it introduces sensitivity to acute physiological dysfunction.^[ref]

The relevant dissociation is stronger than the observation that two tasks activate different places, but it is not an experiment that cuts one tract in isolation. Stroke lesions follow vascular territories; nearby cerebral cortex and several fiber systems can be damaged together. General severity, incomplete lesion coverage and disconnection can affect apparent associations. A patient's conduction aphasia, with disproportionately poor repetition, is consequently consistent with dorsal network dysfunction without proving that all such cases arise from an isolated arcuate lesion. Nor does relatively preserved repetition imply completely normal comprehension or vice versa.

Lateralization further qualifies the claim. The ventral framework includes bilateral auditory–speech analysis, whereas detailed lexical, sentence-level and semantic operations can show stronger leftward biases. Evidence for a left ventral temporofrontal connection is not evidence that all ventral processing is left-only. Conversely, a strongly left-biased dorsal network does not imply that the right hemisphere never contributes to speech. The roles of stimulus type, noise, task and level of representation should be specified before assigning a function to a hemisphere or stream.

The same logic prevents “dorsal equals syntax, ventral equals semantics” from becoming an unsupported substitute theory. A sentence requires recognition, retention, structure building and interpretation. Corianne Rogalsky and colleagues' lesion study associated sentence-comprehension difficulties particularly with temporal and inferior-parietal damage, while damage to Broca's area alone did not produce the predicted obligatory comprehension pattern. That finding is compatible with distributed and connected language processing; it is not a direct assignment of every syntactic computation to one ventral tract.^[ref]

The most persuasive evidence is therefore the convergence of computational distinctions, task-linked structural connections and differential lesion associations, with each method controlling different weaknesses. Stronger tests would match task difficulty, separate auditory recognition from repetition and rehearsal, combine individual functional localization with connectivity, and examine whether damage predicts the relevant computation beyond general impairment. Parallel dorsal and ventral processing is a useful network-level explanation of speech and language, provided it remains a graded division of labor with interaction, rather than an exclusive anatomical partition.

Broca's area contributes to sentence processing, but the evidence does not justify identifying it with a single necessary-and-sufficient syntax processor. “Syntax” includes representing hierarchical relations, combining words, maintaining unresolved dependencies and using structure in interpretation or production. A region can support one of these operations, or the memory and control demands accompanying it, without implementing all of them. The anatomical label itself usually refers to parts of the left inferior frontal gyrus, especially regions corresponding approximately to areas 44 and 45; it is not a functionally homogeneous unit.

The historical association is important but logically limited. Paul Broca's 1861 observations linked severe impairment of articulate speech with left frontal injury. Loss of fluent output is not a selective test of syntactic knowledge, and a large lesion cannot localize every component of its resulting syndrome to the surface region that is most visible. In 2007, Nina F. Dronkers and colleagues used high-resolution magnetic resonance imaging to re-examine the preserved brains from the historic cases and found damage extending substantially beyond the traditionally identified superficial region. This undermines a simple inference from a broad historical syndrome to a small isolated syntax center; it does not invalidate the importance of left frontal language mechanisms.^[ref]

There is positive experimental evidence for frontal sensitivity to syntax. David Embick and colleagues contrasted visually presented sentences containing grammatical word-order errors with sentences containing spelling errors, asking participants to identify the number of errors. The grammatical condition elicited greater activity in Broca's area, with a larger grammatical–spelling difference there than in comparison language regions. Matching lexical material and requiring error detection in both conditions make this more informative than comparing difficult sentences with an unrelated baseline. It supports a contribution to syntactic processing beyond merely seeing words.^[ref]

Nevertheless, grammatical violations require detection, reanalysis, conflict resolution and response selection. Matching overt task format does not necessarily equate these internal demands. A greater functional magnetic resonance imaging response therefore does not establish that the region stores grammar rules, is syntax-exclusive or is necessary for ordinary successful comprehension. Those stronger claims would require converging tests separating syntactic computation from accompanying processes. This is the problem of reverse inference: activation associated with a task component does not uniquely identify the process generating it.

Corianne Rogalsky and colleagues directly tested one alternative mechanism. Participants comprehended subject-relative and more demanding object-relative sentences while performing concurrent speech articulation, a manual control task, or no secondary task. Articulation selectively interfered behaviorally with the harder sentence condition. In imaging, concurrent articulation eliminated the complexity effect in pars opercularis, whereas the manual task did not; pars triangularis showed a different secondary-task profile. These results support an articulatory rehearsal contribution to maintaining information during demanding comprehension and show that the frontal anatomical region contains distinguishable operations.^[ref]

This is evidence against treating every complexity-related frontal response as syntax-specific, not evidence that syntax is entirely reducible to rehearsal. Articulation may alter strategies and general task demands, and absence of a difference under dual-task conditions can have several causes. Conversely, a syntactic account must explain why an independently motivated memory manipulation changes the response. A satisfactory theory should predict when syntactic dependencies need rehearsal or control and when successful structure building proceeds without those extra demands.

Individual functional anatomy supplies another constraint. Evelina Fedorenko and colleagues identified language-selective and domain-general subregions side by side within the broad anatomical Broca's area. They then tested responses across linguistic and nonlinguistic tasks, including harder and easier memory and control conditions. The distinct profiles were detectable within individuals and did not simply align with the traditional 44/45 boundary. Group averaging across anatomically variable people can therefore merge adjacent populations and produce an apparent multifunctional center. Importantly, language selectivity does not equal syntax selectivity: a region responding to sentences may support lexical–semantic combinations as well as structure.^[ref]

Lesion evidence challenges strict necessity claims. In their 2018 study, Corianne Rogalsky and colleagues tested patients using sentence–picture matching and plausibility judgments, including canonical and noncanonical constructions. Analyses implicated superior temporal and inferior-parietal regions in impaired comprehension; patients with frontal damage did not uniformly show the predicted agrammatic comprehension pattern. A traditional diagnostic label and damage to Broca's area cannot therefore be substituted for an independently measured syntactic deficit.^[ref]

Yet preserved performance after a chronic lesion does not prove that the region never contributes in an intact brain: spared tissue, compensation, task sensitivity and unequal lesion coverage matter. Likewise, comprehension findings should not automatically settle the neural implementation of syntactic production, where planning and sequencing impose different demands. The strongest interpretation combines these limitations rather than choosing imaging or lesions as universally decisive.

Broca's area is a heterogeneous participant in a distributed language network. The evidence supports frontal contributions to syntactic and sentence-level processing, with some responses attributable to rehearsal and control and others belonging to language-selective tissue. It weighs against both an exclusive syntax organ and a claim of no language-specific frontal contribution. Progress requires explicit computational hypotheses, individual localization, matched comparison conditions and convergent perturbation or lesion evidence, while distinguishing representations from the processes that operate on them.

Human categories show robust category typicality gradients, but those gradients do not uniquely demonstrate that knowledge is stored as a single concept prototype. A concept prototype is a central or especially representative category representation. In prototype theory, new cases are judged through their similarity to that summary. This challenges a simple classical theory of concepts in which every familiar category is represented solely by equally necessary defining features. The crucial distinction is between evidence that members are psychologically unequal and evidence for the particular representation producing that inequality.

Eleanor Rosch's studies showed that category members are not interchangeable in judgments and processing. People agree that some examples are better representatives than others, and a category name can differentially prepare the processing of good and poor members. Her semantic priming experiments went beyond explicit category typicality ratings: a category-name cue preceded judgments on example pairs, and even physically identical pairs were processed differently depending on their representativeness. Convergence across ratings and performance is important because it reduces the chance that category typicality is only an artifact of being asked to rate goodness.^[ref]

The family resemblance account explains what can make an example representative. In Eleanor Rosch and colleagues' work, participants listed attributes of members of natural categories. More prototypical examples tended to share many attributes with other members of their category while sharing fewer with contrasting categories. Thus an item's usefulness as a representative depends on distributions of features both within and across categories, rather than on one feature occurring in every member. Such structure can support efficient generalization from incomplete evidence. But a descriptive relation between feature overlap and category typicality is not itself an algorithm for learning or storing a concept prototype.^[ref]

Artificial-category experiments make the abstraction claim more testable. In Michael I. Posner and colleagues' 1968 study, participants learned to classify distortions of dot patterns without seeing the underlying prototypes. They subsequently classified the unseen prototypes especially well relative to control patterns from the learned categories, and training variability influenced generalization to new distortions. The result is striking because direct familiarity with the test concept prototype cannot explain its advantage. A summary preserving the regularities among training examples supplies a natural explanation.^[ref]

It is not, however, a unique explanation. In exemplar theory, a new object is compared with stored examples. A central unseen item can resemble many stored examples at once and therefore receive greater summed support than a less central old item. Robert M. Nosofsky's identification–categorization experiments used multidimensional perceptual stimuli and linked classification to stored instances and selective attention to diagnostic dimensions. This provides an explicit alternative to treating every concept prototype advantage as proof of an abstract average. Theories must be distinguished by predictions across learning, transfer and individual trials, not merely by whether they can generate a category typicality effect.^[ref]

A second challenge concerns the difference between identification and membership. Sharon Lee Armstrong and colleagues applied category typicality-rating and verification-time methods to categories including mathematically well-defined ones. Even these categories produced graded judgments and processing differences: an example can seem more representative while being no more genuinely a member. All odd integers, for example, meet the same parity criterion, irrespective of their familiarity or apparent category typicality. Category typicality can therefore reflect an identification heuristic, accessible examples or feature salience rather than the full constitutive content of a concept.^[ref]

This observation limits, rather than abolishes, concept prototype explanations. It remains possible that many everyday categories have probabilistic feature structure while some concepts require rules or deeper knowledge. It also creates a challenge of conceptual combination: a theory should explain how familiar concepts combine into a new one, including cases where the combination's typical instances are not predicted by simply averaging the separate prototypes. If category typicality is all that a representation contains, logical and compositional inference may become difficult to explain.

Context and goals provide a further constraint. Lawrence W. Barsalou compared common taxonomic categories with goal-derived categories. Central tendency, ideals associated with a goal and frequency of encountering an item could make different contributions to category typicality; central tendency was not a sufficient general account of the goal-derived cases. A food may be a good member of a goal-defined category because it is especially suited to the goal, rather than because it is statistically average. This makes a permanently fixed, context-free central representation implausible as a complete account.^[ref]

The theoretical implication is a distinction among levels of explanation. Category typicality gradients and family-resemblance structure are phenomena to explain; concept prototype, exemplar, rule-based and knowledge-sensitive models propose mechanisms that can overlap in their predictions. Strong tests manipulate the distribution of training cases, diagnostic features, goals and novelty, then evaluate predicted generalization and errors while controlling familiarity and task demands. A hybrid or flexible system is plausible, but should earn its flexibility by predicting when different information will guide behavior.

The evidence decisively rejects the idea that familiar category members are psychologically equivalent; it does not establish a universal concept prototype-only architecture. Prototypes can organize efficient recognition and generalization, while instance memory, explicit criteria and goal-dependent knowledge constrain how that organization is used. Confusing graded category typicality with graded truth of membership would overstate what the experiments show.

People often underweight base rates when a vivid case description competes with them, but the extent of neglect depends on representation, relevance and task. The normative benchmark is conditional on a known sampling process. If a case is randomly drawn from a population, a prior class proportion $P(H)$ must combine with the case evidence's likelihood ratio. In Bayes' theorem,Changing prior odds while holding diagnostic evidence fixed should ordinarily change posterior odds. A judgment driven only by resemblance to a class violates that benchmark when the stated premises apply.

In Daniel Kahneman and Amos Tversky's prediction experiments, descriptions of people were paired with different proportions of engineers and lawyers in the source group. Judgments based on an occupationally suggestive description were surprisingly insensitive to those proportions despite their relevance to a random draw. Without the individualized description, participants could use the stated proportions. This contrast is especially informative: the issue was not simply inability to read a base rate, but the way case-specific evidence displaced it. Similarity and probability can agree in ordinary situations, which helps explain why the substitution is tempting.^[ref]

The first account is the representativeness heuristic, potentially supplemented by dual-process theory. An occupational description quickly evokes similarity to a stereotypical engineer or lawyer. That judgment does not inherently encode class prevalence or evidential reliability, so it can dominate the probability estimate. More controlled reasoning can integrate the base rate, but may fail to engage, misunderstand the statistical structure or fail to override the compelling response. This account predicts errors especially when representativeness and prior probability conflict, with possible improvement when people deliberately compare alternative hypotheses and their frequencies.

It is too strong to infer from a biased answer that no statistical information was processed. Wim De Neys and colleagues contrasted problems where descriptions and base rates conflicted with control problems where they agreed. Although verbal reports often emphasized the description, measures including decision time, later memory for the base rates and reconsideration behavior indicated sensitivity to the conflict. Statistical information can therefore influence processing without determining the final answer. This challenges a simple account in which analytic monitoring is always absent; it is compatible with incomplete integration or unsuccessful correction.^[ref]

The second account emphasizes ecological rationality and information format. Gerd Gigerenzer and Ulrich Hoffrage argued that algorithms adapted to natural frequencies can perform Bayesian updating more easily than algorithms given separately normalized probabilities. Natural sampling represents the relevant nested sets directly: how many cases belong to each class, and how many within each class produce the evidence. The denominator can then be the whole set of evidence-positive cases, without reconstructing joint counts by multiplying conditional probabilities.

Their 1995 experiments compared probability and frequency versions of Bayesian problems while also varying the information menu. Frequency formats substantially increased the use of Bayesian solutions across the problem set, assessed using solution processes as well as final answers. This supports a representational contribution to apparently poor reasoning; it does not show that everyone is spontaneously Bayesian or that every frequency phrasing helps. The preserved nesting and common reference population matter. Normalizing both classes to independent “out of one hundred” samples can remove precisely the base-rate information the natural-frequency representation was meant to preserve.^[ref]

An original numerical illustration makes the distinction explicit. Suppose a randomly sampled population contains $200$ category-$H$ cases and $800$ non-$H$ cases. A cue occurs in $75\%$ of the first class and $25\%$ of the second. There are then $150$ cue-positive $H$ cases and $200$ cue-positive non-$H$ cases, so $P(H\mid E)=150/(150+200)=3/7$, approximately $0.429$. Reporting $0.75$ would confuse the cue's conditional rate in class $H$ with the probability of class $H$ given the cue. These are constructed teaching counts, not results from a cited experiment.

The two accounts agree that the task representation matters and that naïve answers need not obey the normative model. They disagree in emphasis: the heuristic account highlights substitution and control, while the ecological account highlights a mismatch between the supplied representation and an available inference algorithm. Frequency success alone cannot decide between them, because a clearer nested-set display can also make deliberate correction easier. Conversely, conflict sensitivity constrains an absolute “base rates were never noticed” story without proving successful Bayesian computation.

A discriminating design would cross numerical format with description–base-rate conflict, preserve the sampling assumptions and information content, and measure answers together with process indicators. It should distinguish a failure to encode prevalence, a failure to combine encoded evidence and a rational belief that the stated reference group is irrelevant. Verbal descriptions seldom specify a likelihood ratio, so prior-insensitivity is easier to diagnose than a unique numerical Bayesian answer for every description. If descriptions were selectively rather than randomly sampled, or a base rate were not applicable to the individual, the simple benchmark would change. Accordingly, base-rate neglect is a reproducible conditional pattern, not evidence that people invariably ignore statistics or possess no Bayesian competence; both heuristic competition and representational structure help explain when it occurs.