Delusions and Imagination

1Delusions and Imagination – Philip R. Corlett

2Imagination, Agency and Predictive Processing
3
4ABSTRACT
5Predictive processing theories of delusions and hallucinations suggest a simple
6dysfunction as the core of these puzzling symptoms. Perception, belief, and action are
7all cast in terms of inferences, wherein prior beliefs are integrated with sensory inputs
8based on their relative reliabilities. Delusions and hallucinations then arise when this
9integration process is perturbed, either toward priors (hallucinations) or sensory inputs
10(delusion formation). However, there are other popular theories of delusions and
11hallucinations, including models that suggest delusions and hallucinations involve
12confusing imagining for believing or perceiving respectively. In this chapter, I will
13attempt to reconcile these accounts with predictive processing theory, appealing to
14associative and reinforcement learning data that integrate imaginative mechanism,
15including representation mediated learning, which may bring source monitoring and
16imagination accounts under the explanatory umbrella of predictive coding theory. I will
17end by speculating on a neurocomputational architecture that may underlie predictive
18processing and makes use of generative, imaginative, processes in service of reality
19perception. Dysfunction in this network can give rise to hallucinations, delusions, and
20other perturbations of conscious experience.
21
22KEYWORDS
23Delusions, Hallucinations, Psychosis, Imagination, Predictive Processing,
24Representation Mediated Learning, Source Monitoring.
25
26

1
 Delusions and Imagination

27INTRODUCTION
28
29This chapter is about delusions, hallucinations, and imagination. Delusions and

30hallucinations are symptoms of psychotic illnesses like schizophrenia, wherein beliefs

31and percepts depart from consensual reality. However, such departures also occur in

32neurological illnesses, closed head injuries, and autoimmune illnesses, as well as in

33attenuated forms along continua in the general (non-clinical) population. My work aims

34to understand this departure in terms of the mechanisms of belief formation and

35updating, with a particular focus on how they are implemented in the brains of people

36with psychosis and people who have hallucination-like experiences and delusion-like

37beliefs.

38

39Thus far, I have taken a reductionist approach, grounded in formal associative learning

40and reinforcement learning theories and, by extension predictive coding, which aims to

41subvert human chauvinism and explain apparently ‘special’ capabilities and

42phenomena, considered higher-level cognitive operations, in terms that are amenable to

43inquiry in simpler model systems (like primates, rodents, and even in silico).

44

45I like things to be as simple as possible (but no simpler) and I believe that

46understanding the implementation of these cognitive operations in cells, circuits, and

47systems, will illuminate the shape of minds and the allowable interactions between

48components of minds.

49

50This attitude means that I am sceptical of explanations that evoke bespoke modular

51gadgets that solve theoretical problems (coalitional threat detection system, cheater
2
 Delusions and Imagination

52detection gadget etc). Instead, I am interested in how simple precision weighted

53learning and belief updating can subsume the functioning of many of those gadgets

54(Heyes and Pearce 2015, Heyes 2018), and how much of delusionality can be

55accounted for with that domain-general learning process.

56

57This has furnished heated debate and disagreement between me and people who

58appeal to domain-specific processes (to confer delusion contents for example). This has

59been immensely generative and allowed my collaborators and I to hone our account

60closer and closer to a more complete explanation of delusions.

61

62Some things are still lacking though (like how delusions have their characteristic

63contents, but see (Corlett, Taylor et al. 2010)).

64

65In this chapter I try to reconcile a different class of explanation with the reductionist

66approach. Some philosophers have pointed out that delusions and hallucinations might

67be understood as confusions between having imagined some stimulus (or

68circumstance) and having perceived (or believed) it (Currie 2000). In what follows, I am

69going to try to adumbrate a path through which the predictive coding theory of psychosis

70might be aligned with these important ideas.

71

72This will involve clarifying how imagination might work in the context of reinforcement

73learning and predictive coding, and what the relationships between associative and

74reinforcement learning and imagination might be. It will entail some consideration of the

3
 Delusions and Imagination

75history of the basic neuroscience of associative learning (with particular emphasis on

76dopamine and glutamate) and how it has recently been extended to imagination. I will

77then discuss the intersection of that work with predictive coding theories and data on

78hallucinations and delusions. I will conclude by speculating about a key role for

79imagination in a novel neural network architecture, the Generative Adversarial Network.

80I will suggest that failures in this network portend different dysfunctions of conscious

81experience, including delusions and hallucinations.

82

83ALGORITHMS FOR LEARNING AND BELIEF

84The psychology and neurobiology of decision making, belief formation, and updating,

85have been enriched by the application of reinforcement learning principles. Put simply,

86we learn to believe things that facilitate choices which increase our adaptive fitness, by,

87for example, increasing contact with rewards and decreasing contact with punishments.

88

89There are many algorithms that might explain or even underwrite how organisms solve

90these problems. Analysing their tractability as well as the ways in which they

91parameterize decisions can yield insights into how beliefs are formed and updated, with

92relevance, not only to the adaptive case, but also instances wherein we err in our

93beliefs.

94

95Presently, I shall argue that whilst imaginative reconstruction of possible outcomes is

96useful for goal-directed action choice, such powerful mechanisms are extremely

4
 Delusions and Imagination

97sensitive to noise and uncertainty, and can perhaps lead to gross mischaracterizations

98of states, and ultimately, to perception, actions, and beliefs that do not reflect reality.

99

100The dopamine rich basal ganglia and their reciprocal connections with the frontal

101cortices and the hippocampus are implicated in reinforcement learning, credit

102assignment (apportioning blame for outcomes to the appropriate actions and agents),

103imagination and psychosis. The late Jeffrey Gray, in collaboration with David Hemsley

104and others at the Institute of Psychiatry in London developed a model integrating the

105neurobiology and neurochemistry of the basal ganglia as well as their input and output

106structures with the cognitive and phenomenological experiences of psychotic patients

107(Gray, Feldon et al. 1991, Hemsley 1992, Gray 1993, Hemsley 1993, Hemsley 1994,

108Hemsley 1994, Gray 1995, Gray 1998, Gray 1998, Hemsley 2005, Hemsley 2005). It

109focused on the hippocampus and its dopaminergic inputs to the nucleus accumbens,

110suggesting that the hippocampus served as a comparator, comparing the perceptual

111inputs to which the organism was exposed with the organism’s internal model of the

112world (the regularities and contingencies which it learned by prior experience (Gray,

113Feldon et al. 1991)).

114

115If there was a mismatch between what was expected and experienced, a novelty or

116salience signal would be sent to the striatum, inhibiting ongoing actions and eliciting an

117orienting response.

118

5
 Delusions and Imagination

119Gray, Hemsley and colleagues argued that in experimental animals treated with

120amphetamine and individuals with psychosis, such mismatch signals occur

121inappropriately, driving attention towards irrelevant environmental stimuli and internal

122thought processes, culminating in the attribution of delusional significance to external

123stimuli and events and in the experience of hallucinations (Gray, Feldon et al. 1991).

124

125Mismatches between expectancy and experience, or prediction errors, dominate

126theories of learning (Rescorla and Wagner 1972, Mackintosh 1975, Pearce and Hall

1271980, Grossberg 1982) and dopamine function (Schultz 2000, Waelti, Dickinson et al.

1282001, Schultz 2004). In the classic studies, dopamine neurons deep in the monkey’s

129midbrain report the surprising presence or absence of rewards, with learning, they

130respond to cues (for example visual stimuli) that predict the presence or absence of

131rewards.

132

133PREDICTION ERRORS

134Psychological models of prediction error driven learning posit that prediction errors

135contribute to learning directly, updating organisms’ expectancies of particular stimuli

136(Rescorla and Wagner 1972) and indirectly, driving the subsequent allocation of

137attention to stimuli that predict important events (Mackintosh 1975, Grossberg 1982)

138and to stimuli that have an unpredictable or uncertain relationship with important events

139(Pearce and Hall 1980, Grossberg 1982), both of which are salient. Gray et al’s thesis

140provided a mechanistic link between brain function and psychotic symptoms via

141dopamine dysregulation and inappropriate attentional learning. However, to borrow

6
 Delusions and Imagination

142Gray’s phraseology, some of that particular wall had also been laid down previously

143(Gray 2004).

144

145Robert Miller first made the connection between dopamine dysfunction, aberrant

146associative learning, and delusion formation, suggesting that increased dopamine

147signalling in the basal ganglia in schizophrenia lowered the significance threshold on

148conclusions that stimuli and events were related and that those relationships were

149important (Miller 1976). Roy King provided mathematical models of chaotic dopamine

150function in schizophrenia that could explain the dopamine responsivity independent of

151cue and context that Shitij Kapur proposed to be the mediator of aberrant salience

152(King, Raese et al. 1981, King, Barchas et al. 1984).

153

154 The elegant tonic/phasic dopamine hypothesis of Anthony Grace offered further

155conciliation between brain dopamine dysfunction in frontal, temporal, midbrain and

156striatal systems, the psychological processes of learning and attentional allocation, and

157psychotic symptoms (Grace 1991). It also highlighted the importance of glutamatergic

158regulation of the balance between tonic and phasic dopamine (Grace 1991). With

159reference to the phasic/tonic dopamine balance and signal-to-noise ratio in the

160dopamine system, Manfred Spitzer outlined an account of delusion formation and

161maintenance, suggesting that under conditions high signal-to-noise ratio, an individual

162would experience “significant events” when merely ordinary events were in fact

163happening and those experience, once generated, would be less likely to passively

164decay or be actively erased – manifest clinically as a delusional belief (Spitzer 1995).

165

7
 Delusions and Imagination

166Phasic dopamine release is held to carry the prediction error signals recorded from

167midbrain and striatal dopamine neurons by Wolfram Schultz. Tonic dopamine levels in

168the synapse are impacted by that release and spill-over, as well as feedback from the

169prefrontal cortex (Grace 1991). They set the gain on the system, modulating the impact

170of a particular phasic, prediction error related release.

171

172In the absence of prefrontal feedback regulation, the signal to noise ratio of the system

173is altered, such that the phasic dopamine component becomes over-responsive,

174precipitating positive symptoms (Grace 1991). Recently, the tonic/phasic scheme has

175been integrated with Gray et al’s model, such that the ventral hippocampus impacts

176upon the responsivity of midbrain dopamine, prediction error coding neurons, increasing

177the number of neurons that are recruited to readily excitable neuron pools in the VTA

178which respond to a particular stimulus (Lodge and Grace 2006). Disruption of this

179hippocampal modulation of VTA neurons, under conditions of dopamine hyperactivity or

180glutamate hypofunction could result in the excessive and chaotic stimulation of the

181striatum and prefrontal cortex, producing positive psychotic symptoms according to the

182Gray et al model (Gray, Feldon et al. 1991, Lisman and Grace 2005).

183

184This aberrant learning model has been tested preclinically, in pharmacological and

185lesion models (Kaye and Pearce 1987, O'Tuathaigh, Salum et al. 2003), as well as

186clinically in human subjects suffering from schizophrenia (Jones, Gray et al. 1992,

187Oades, Bunk et al. 1992, Oades, Zimmermann et al. 1996, Moran, Owen et al. 2007)

8
 Delusions and Imagination

188and in healthy individuals with high scores on schizotypal personality scales (Jones,

189Gray et al. 1990, Moran, Al-Uzri et al. 2003).

190

191ASSOCIATIVE LEARNING

192Two learning phenomena are frequently employed in examinations of the hypothesis;

193both of which measure the allocation of attention to and learning about irrelevant stimuli,

194namely blocking (Kamin 1969) and latent inhibition (Lubow and Moore 1959).

195

196Both of these paradigms involve the adaptive down regulation of the salience of a

197stimulus, either due to pre-exposure to that stimulus in the absence of important

198consequences (latent inhibition) or due to pairing of those important consequences with

199a salient and predictive stimulus which then “blocks” subsequent learning about any

200novel stimulus paired with that pre-trained stimulus. Both blocking and latent inhibition

201are held to reflect the adaptive use of environmental regularities to guide behavior and

202filter out irrelevant and potentially distracting stimulation. However, in psychosis, this

203filtering process is impaired, leading to sensory overload, inappropriate attribution of

204significance to inconsequential of irrelevant stimuli and positive symptoms.

205

206IMAGINATION

207The learning accounts and behavioral phenomena considered so far have been rather

208passive, and anchored in learning about the external world. As noted in the introduction,

209imagining possible future states and the fictive consequences of action choices might

210enhance the learning and adaptation of an agent to its environment.

9
 Delusions and Imagination

211

212Edward Tolman was a neo-behaviorist. He believed even rodents were capable of

213generating elaborate internal models of their environment (Tolman 1932). Guthrie, a

214more traditional behaviorist, quipped that Tolman’s rats would be “buried in thought”,

215incapable of responding because of the elaborate thought processes in which they were

216engaged (Guthrie 1952). In new data, Takahashi and colleagues exhume Tolman’s

217ideas – showing with careful neural and behavioral studies that rats can imagine

218outcomes they have never experienced (Takahashi, Chang et al. 2013).

219

220Pavlov’s dogs learned to salivate in response to a bell that predicted food (Pavlov

2211927). Follow-up studies have demonstrated summation; if animals learn that another

222cue, say a light, predicts the same food, and they experience light and bell together,

223then they expect double helpings of reward. This over-expectation, based on summed

224expectancies is manifest as vigorous behavioral responses following the compound of

225the two predictive cues (Thein, Westbrook et al. 2008).

226

227Takahashi et al demonstrate that this over-expectation is mediated by the orbitofrontal

228cortex (OFC) in rodents. The neurons in this region increase in responding as the

229animals learn about the cues. They respond most strongly to the compound cues in a

230manner that predicts subsequent over-expectation behavior. Finally, using optogenetics,

231a technique that allows light to inhibit the over-expecting cells, they show that animals

232summate less when their OFC neurons are inhibited.

233

10
 Delusions and Imagination

234People often seem uncomfortable with the idea of rats having rich representations and

235even being able to imagine. Yet many allow for working memory in rats – the

236maintenance of information over a delay in the absence of sustaining environmental

237stimuli. Imagination is really nothing more than an internally generated representation

238with much in common with that which mediates working memory. Problems with animal

239imagination are further puzzling when we consider the field of translational

240neuroscience studying mental illnesses like schizophrenia and addiction with animal

241models.

242

243Takahashi and colleagues lend a neural reality to a relatively high-level cognitive

244capacity – that of creative expectation or imagination in rodents.

245

246IMAGINATION AND PSYCHOSIS

247Stephen Grossberg’s (2000) computational theory of hallucinations bridged the gap

248between conditioning and incentive models of hallucination (Grossberg 2000). He

249equated hallucinations with the subjective experience of imagination. Grossberg

250conceptualized imagination as the willed deployment of top-down influences on

251perception. Like Frith (Frith 1987), he suggested that in schizophrenia, actions can

252occur in the apparent absence of conscious intentions, due to dysregulated

253dopaminergic volitional signals and, in such situations the sensations generated are

254perceived as externally generated hallucinations (Grossberg 2000). The conditioned

255hallucinations of Seashore, Ellison and others are an example of how top-down

256expectancies can modulate experience (Seashore 1895, Ellson 1941, Kot and Serper

11
 Delusions and Imagination

2572002). In the psychotic individual this process is thought to occur endogenously and

258inappropriately under the influence of an aberrant dopamine system (Gray, Feldon et al.

2591991, Miller 1993, Kapur 2003). There is some evidence that sensory “hallucinations”

260can be conditioned in healthy and psychotic volunteers using a sensory discrimination

261procedure (Seashore 1895, Ellson 1941, Kot and Serper 2002). In brief, subjects are

262instructed that their sensory thresholds are being measured, each trial begins with the

263illumination of a light (the conditioned stimulus, CS), subsequently a tone of increasing

264intensity is played (the unconditioned stimulus, US). Subjects are instructed to make a

265button-push response when they hear the tone. On certain trials, the tone is omitted. An

266hallucination is said to have occurred if subjects report hearing a tone when no tone

267was played. Seashore has demonstrated this phenomenon with olfactory, tactile and

268electrodermal conditioning (Seashore 1895) and, while an explanation could be

269conceived in terms of response biases, these and other data suggest some role of prior

270knowledge and expectancies in generating perceptual experiences, both normal (Frith

271and Dolan 1997) and abnormal (Frith 2000).

272

273More recently, visual-auditory conditioning has been employed to demonstrate that

274voice-hearing schizophrenia patients are significantly more susceptible to this effect

275than patients without hallucinations and controls (Kot and Serper 2002). These

276conditioned hallucinations are mediated by strong prior beliefs, and those priors are

277stronger in people who hallucinate (in a manner that correlates with hallucinations)

278(Powers, Mathys et al. 2017). Furthermore, people with a diagnosed psychotic illness

279are less likely to update those prior beliefs in light of new evidence (Powers, Mathys et

12
 Delusions and Imagination

280al. 2017). Critically, the neural circuit underlying these conditioned phenomena—

281including superior temporal gyrus and insula—largely overlapped with the circuit

282engaged when patients report hearing voices in the scanner (Jardri, Pouchet et al.

2832011, Powers, Mathys et al. 2017). These studies underline the role of learning and,

284more specifically, a bias toward learned top-down information in the genesis of AVHs.

285

286Further support for this strong prior account of hallucinations comes from findings that

287prior knowledge of a visual scene confers an advantage in recognizing a degraded

288version of that image and that patients at risk for psychosis—and, by extension, voice-

289hearing—were particularly susceptible to this advantage, and its magnitude correlated

290with hallucination-like percepts outside of the laboratory (Teufel, Subramaniam et al.

2912015). Similarly, there is a version of this effect in audition; voice-hearing participants

292appear to have an enhanced prior for speech in degraded auditory stimuli even when

293not explicitly instructed (Alderson-Day, Lima et al. 2017). That is, speech is perhaps the

294most salient biological signal for our species, and the auditory system of AVH-prone

295individuals may be predisposed to inferring speech.

296

297This raises an important possible distinction between clinical and non-clinical voice

298hearers. Since non-clinical voice hearers have a ready explanatory framework and have

299spent time cultivating their volitional control over perception [likely over the weighting of

300their prior expectations (regarding some imagined event) relative to the sensory

301evidence (that the event is not occurring)], then they can better anticipate hallucination

302episodes and can understand them through their preferred explanatory lens (Powers,

13
 Delusions and Imagination

303Kelley et al. 2016). This would be consistent with the sensitivity of the non-clinical voice

304hearers to environmental volatility – the rate of change of environmental contingencies:

305they are able to update their beliefs rapidly and appropriately through learning, unlike

306the clinical groups (Powers, Mathys et al. 2017). If one can move between possible

307priors with alacrity, reconfiguring one’s expectations smoothly and accurately, one

308would likely have strong attentional control over learning and belief updating, and by

309extension, imagination.

310

311REINFORCEMENT LEARNING AND IMAGINATION

312Contemporary models of reinforcement learning and artificial intelligence have an

313imaginative, simulation component, that has been evidenced in humans (Gershman,

314Zhou et al. 2017). In some cases, imagined reinforcers can maladaptively bias real-

315world decisions – imagining that you have been rewarded can change your subsequent

316reward seeking (Gershman, Zhou et al. 2017). The adaptive use of imagination

317depends on its obedience to real world constraints (perhaps as observed in non-clinical

318voice hearers). If untethered, then we may find ourselves transcending reality too much

319(Gershman, Zhou et al. 2017), and ultimately failing to learn from the real world

320(Gershman, Zhou et al. 2017) (as observed in patients with psychosis).

321

322SOURCE MONITORING

323It is important to distinguish reality testing (discriminating between real and imagined

324stimuli in perception) and reality monitoring (discriminating between real and imagined

325stimuli in memory). The bulk of the the work in psychosis has – ironically perhaps –

14
 Delusions and Imagination

326focused on monitoring (i.e. a mnemonic process) rather than the more perceptual –

327testing or filtering of reality.

328

329Imagining leaves a trace in memory, and we must then be able to discriminate between

330these memories and memories of observed stimuli (Johnson 1981). Deficits in such

331discrimination have been related to delusions (Simons, Henson et al. 2008) and

332hallucinations, though not consistently (Alderson-Day, Smailes et al. 2019), and

333perhaps not in non-clinical voice hearers (Garrison, Moseley et al. 2017).

334

335One animal learning phenomenon that may unite prior weighting and source monitoring

336views of psychosis is representation mediated learning, wherein the value of an

337expected cue is updated even though that cue is never presented with a salient

338outcome (like a food reward). Through associations, an expectation of a stimulus (S2) is

339engendered, contingent on the presentation of a stimulus that predicts it (S1). The

340expected stimulus (S2) is not presented, however, it is, through those learned

341associations retrieved from memory when its predictor S1 is presented. Representation

342mediated learning occurs when that retrieved representation is updated in light of some

343change in the value of S1 (pairing it with reward or punishment for example). Mediated

344learning probes the extent to which animals can distinguish between real percepts and

345internally retrieved representations of cues (McDannald and Schoenbaum 2009,

346McDannald, Whitt et al. 2011, Kim and Koh 2016, Koh, Ahrens et al. 2018, Fry, Russell

347et al. 2020, Koh and Gallagher 2020).

348

15
 Delusions and Imagination

349It typically proceeds through three phases: 1) associative phase: odor (stimulus 1; S1) +

350flavor (stimulus 2; S2) pairings, 2) devaluation phase (e.g., aversion training): the flavor

351(S2) is devalued by an injection of lithium chloride (LiCl) to induce gastric malaise to

352alter its “value”, and, 3) a test phase to determine if both the associated odor (S1), in

353addition to the experienced flavor (S2), are devalued (i.e., mediated and direct learning,

354respectively). Patterns of neural activity that correspond to the representation of the S1

355cue, recorded during the devaluation phase when this cue is absent, but expected, can

356predict stronger expression of mediated learning at test (Kerfoot, Agarwal et al. 2007,

357Saddoris, Holland et al. 2009). Mediated learning bridges perceptual decision-making

358and conditioned hallucination tasks because it assesses retrieved cue representations,

359with translational relevance (Wimmer and Shohamy 2012, Zeithamova, Dominick et al.

3602012, Barron, Reeve et al. 2020, Wang, Howard et al. 2020).

361

362Various manipulations that produce psychosis-like phenotypes in rodents, including

363chronic ketamine (Koh, Ahrens et al. 2018), increase mediated learning (McDannald

364and Schoenbaum 2009, McDannald, Whitt et al. 2011, Busquets-Garcia, Soria-Gómez

365et al. 2017, Koh, Ahrens et al. 2018, Fry, Russell et al. 2020, Wu, Haberman et al.

3662020). Furthermore, antipsychotic medications can rescue the phenotype in some

367models (Busquets-Garcia, Soria-Gómez et al. 2017, Koh, Ahrens et al. 2018, Fry,

368Russell et al. 2020). The potential of mediated learning has yet to be explored in

369humans with psychosis, however, the preclinical data and theory suggest that

370representation mediated learning ought to be enhanced in people who hallucinate, and

371perhaps those who harbor delusional beliefs too. Critical to the phenomenon is a

16
 Delusions and Imagination

372confusion of having remembered a stimulus (perhaps very richly) with having perceived

373it. In the final section of this chapter, I will speculate on how the brain might distinguish

374imagination from perception, and how those mechanisms might breakdown, portending

375various changes in the sense of reality of conscious perception. I believe these

376examples may be relevant to psychosis, and taken together, this approach proffers the

377exciting possibility of unifying source monitoring and learning based explanations of

378psychotic symptoms (Griffin and Fletcher 2017).

379

380SIMPLE MECHANISMS, COMPLEX PHENOMENA

381Throughout this piece I have suggested that basic reinforcement learning mechanisms,

382co-opted towards perception and belief, give rise to delusions and hallucinations. I have

383suggested that something like simple cued-associative retrieval can evoke

384representations into a space akin to a simple form of imagination, and that those evoked

385representations can be confused for real percepts, especially in neurochemical and

386neurobiological contexts that resemble psychosis. I’d like to try to relate my argument

387back to delusions and I will use paranoia and persecutory delusions as an example.

388

389Recent work in the cognitive science of paranoia has suggested that paranoia – the

390belief that others intend to harm one – is the purview of an evolved mechanism for

391detecting coalitional threats (Raihani and Bell 2019). This mechanism is held to be

392defective in people who are paranoid. The hypothesis has largely been examined in the

393context of game theory based iterative interactions, and indeed, people who are

394paranoid seem to have trouble generating and using representations of the agents they

17
 Delusions and Imagination

395are interacting with (Barnby, Raihani et al. 2022). Theorists of this bent often appeal to

396mechanisms like recursive mentalizing, thinking a number of moves ahead about the

397counterfactuals: “what should I do if they do this, and how will they respond?” (Barnby,

398Raihani et al. 2022). This type of mentalizing has much in common with theory of mind

399tasks, performance of which, perhaps surprisingly, is not consistently or strongly related

400to paranoia (Pinkham, Harvey et al. 2016) (though of course it is possible that these

401older tasks were not recursive enough).

402

403In work from my lab on this topic, we find that people who are paranoid have

404appropriate prior expectations about the advice of minimal-group partners; they expect

405that someone on their team will help them, and that someone competing might mislead

406them. Neither effect is exaggerated in paranoia (Rossi-Goldthorpe, Leong et al. 2021).

407Rather, paranoid people will rely in advice from collaborators or competitors, particularly

408when decisions are difficult, and they do so especially when they find their own

409perceptual decisions to be unreliable (Rossi-Goldthorpe, Leong et al. 2021). This

410centers paranoia not in group cognition, but rather in noisy personal level perceptual

411decision-making. Further work from my group suggest that this may be a fundamental

412expectation in people who are paranoid (including patients with schizophrenia who have

413persecutory delusions (Sheffield 2022)); they expect that the world will be volatile and

414changeable (Reed, Uddenberg et al. 2020, Suthaharan, Reed et al. 2021). We even find

415evidence in basic preclinical animal models: rats treated with methamphetamine have

416stronger prior beliefs about volatility (Reed, Uddenberg et al. 2020). We contend that

417what looks like mentalizing may actually – like representation mediated learning - be a

18
 Delusions and Imagination

418much simpler version of cue driven learning, and that delusions involve aberrations of

419this basic learning mechanism, rather than mentalizing proper. Of course it remains an

420empirical question whether these simple mechanisms undergird more baroque social

421cognition, but, following occams razor, I prefer not to evoke something specifically or

422separately social before we have exhausted the reach of simpler mechanisms.

423

424There are many other complexities, beyond symptom contents, that demand our

425attention. In what remains, I will try to sketch a version of how the predictive coding

426theory of mind and psychosis might address the sense of reality of percepts relative to

427imagination, and, by extension, hallucinations.

428

429THE REALITY OF PERCEPTION

430Where might the sense of reality (or imagination) arise in a purely inferential model?

431Anil Seth has argued that sensorimotor counterfactuals are critical (Seth 2014). People

432perceive something as real to the extent that their perceptual hypothesis testing is

433satisfied (Seth 2014). For example, real things rarely move in perfect synchrony with our

434eye movements. If an object in the visual field moves when our eyes move, we will likely

435infer that it is not real. Furthermore, real things are unlikely to retreat in synchrony with

436our investigative reaching movements. This is a perspicuous idea, however, people who

437are unable to move their eyes may experience low sensorimotor contingency, but can

438still discriminate real from imagined stimuli.

439

440Perhaps a more in-depth consideration of our sense of visual reality might help clarify?

19
 Delusions and Imagination

441
442In the neuropsychology of vision there are patients whose brain damage and

443subsequent dysfunction help us think about the perception of reality. Blindsight

444describes the covert visual abilities of brain damaged individuals who deny conscious

445visual perception. It is contrasted with Anton’s syndrome, wherein blind individuals claim

446to be sighted and behave as though they are (walking through the world, but colliding

447with objects in it). Both of these syndromes are characterized by acquired damage to

448the visual cortex in individuals whose development (and interaction with the visual

449world) was otherwise typical. The syndromes are consistent with a Bayesian Prediction

450Error Minimization Model of conscious perception, wherein candidate percepts are

451predicted and visual experiences are consistent with those model predictions. The

452syndromes differ in the richness with which those predictions are experienced.

453Blindsight has low richness which conflicts with behavioral detection of stimuli. Anton’s

454has high richness that conflicts with reality. How is this possible in the model? And, how

455are these cases instructive with regards to blindness and psychosis?

456

457The computational motif of generative adversarial networks may be instructive. These

458involve:

459

460 1) A training set of images (containing e.g. dogs)

461 2) A Generator Network

462 3) A Critic Network

463

20
 Delusions and Imagination

464First, the generator is trained with the training set (so that it learns a given examplar is a

465picture of a dog). Then it is inverted to produce novel content (images of dogs). The

466Critic decides whether a generated image belongs amongst the training set (is this a

467dog?). It is reinforced for correctly rejecting what the Generator produces. The

468Generator is reinforced for fooling the Critic. Thus, the two networks train one another.

469This arrangement is one mechanism through which predictive coding could implement

470reality monitoring, and give rise to conscious perception (Lau 2008).

471

472In blindsight, it is possible that the generator is impaired. And in Anton’s syndrome, the

473critic may be impaired. In Charles-Bonnet syndrome, characterized by visual

474hallucinations, wherein the retina is degenerating, the generator maybe hyper-engaged

475(Reichert, Series et al. 2013) however, the critic is somewhat intact, since Charles-

476Bonnet syndrome cases often appreciate that they are hallucinating. The congenitally

477blind never achieve a fully formed generative network, and as such they do not suffer

478the deleterious impact of perturbations within that network. And endogenous psychosis

479involves both noise and compensation within the generative network and a failure of

480reality monitoring by the critic.

481

482This sounds rather like a 2-factor theory of delusions and psychosis – which demands

483two independent deficits, one in perception (a broken generator), and one in belief

484evaluation (a broken critic) (Gershman 2019). However, lets consider the information

485flow within a GAN. The premier 2 factor theory of delusions – based on monothematic

486delusions secondary to neurological damage – holds that people with these delusions

21
 Delusions and Imagination

487have a deficit in perception (Factor 1) and belief evaluation (Factor 2) (Coltheart 2010).

488Under the encapsulated modularity of two factor theory, the flow of information is

489unidirectional, from perception, to belief (Ross 2016). For 2-factor theory, beliefs cannot

490influence perception (otherwise only damage to the mechanisms of belief would be

491necessary for delusions to arise (Corlett 2019)).

492

493As an aside, such an architecture would demand quite a specific and unlikely set of

494insults, in which odd experiences led to a final instance of belief updating, and thence

495beliefs were concretized (McKay 2012, Corlett 2019).

496

497Now, careful consideration of GANs reveals that they have no such unidirectional flow

498of information. The generator also learns from the critic – particularly its inaccuracies (or

499prediction errors perhaps), which tell the generator how convincingly it has emulated the

500data distribution pertaining to reality (Gershman 2019). It is possible then, that although

501GANs appear aligned with modularity and two factor theory, they are actually

502underwritten by mechanisms that share more with predictive processing. Indeed,

503computer scientists argue that the ‘clamping’ of layers within a hierarchy that permits

504inception – as in Google Deep Dream – could instantiate the type of generative activity

505required for adversarial functionality (Sun and Orchard 2020).

506

507The question remains, how would such a network ‘know’ that it was in generative

508mode? Does the precision of prediction errors in the critic change when we imagine?

509This would allow prior beliefs to dominate our inferences.

22
 Delusions and Imagination

510

511These are exactly the circumstances that Martins and Pacherie argue obtains during

512hypnosis (Martin and Pacherie 2019). Perhaps then, there is the potential for

513rapproachment – the Grossberg account has its prior beliefs that predominate, Frith has

514intention being brought to the fore (through the down regulation of prediction error – the

515suspension of disbelief). This is the state that clairaudient psychics might achieve, as do

516those who practice tulpamancy (Powers, Kelley et al. 2016). The difference between

517them and individuals with psychotic illness, is that those with an illness wield less

518control over the precision of their prediction errors and as such, they do not control the

519phenomenology they experience to the same degree (Powers, Kelley et al. 2017).

520

521In summary then, there is much to be gleaned from considering hallucinations and

522delusions through the lens of imagination and belief. In so doing, we learn more about

523the potential functions imagination might serve (simulation, planning, tuning our

524generative model of the world) and the sorts of algorithms that might underwrite it. It

525remains to be shown definitively, however, evidence mounts that beliefs might well

526influence perception and that perhaps strict distinctions between these two are

527challenged by the data on imagination, as well as on hallucination and delusion.

528
529 References
530
531Alderson-Day, B., C. F. Lima, S. Evans, S. Krishnan, P. Shanmugalingam, C. Fernyhough and S. K.
532Scott (2017). "Distinct processing of ambiguous speech in people with non-clinical auditory
533verbal hallucinations." Brain 140(9): 2475-2489.
534Alderson-Day, B., D. Smailes, J. Moffatt, K. Mitrenga, P. Moseley and C. Fernyhough (2019).
535"Intentional inhibition but not source memory is related to hallucination-proneness and
536intrusive thoughts in a university sample." Cortex 113: 267-278.

23
 Delusions and Imagination

537Barnby, J. M., N. Raihani and P. Dayan (2022). "Knowing me, knowing you: Interpersonal
538similarity improves predictive accuracy and reduces attributions of harmful intent." Cognition
539225: 105098.
540Barron, H. C., H. M. Reeve, R. S. Koolschijn, P. V. Perestenko, A. Shpektor, H. Nili, R. Rothaermel,
541N. Campo-Urriza, J. X. O’Reilly, D. M. Bannerman, T. E. J. Behrens and D. Dupret (2020).
542"Neuronal Computation Underlying Inferential Reasoning in Humans and Mice." Cell 183(1):
543228-243.e221.
544Busquets-Garcia, A., E. Soria-Gómez, B. Redon, Y. Mackenbach, M. Vallée, F. Chaouloff, M.
545Varilh, G. Ferreira, P. V. Piazza and G. Marsicano (2017). "Pregnenolone blocks cannabinoid-
546induced acute psychotic-like states in mice." Molecular psychiatry 22(11): 1594-1603.
547Coltheart, M. (2010). "The neuropsychology of delusions." Ann N Y Acad Sci 1191: 16-26.
548Corlett, P. R. (2019). "Factor one, familiarity and frontal cortex: a challenge to the two-factor
549theory of delusions." Cogn Neuropsychiatry: 1-13.
550Corlett, P. R., J. R. Taylor, X. J. Wang, P. C. Fletcher and J. H. Krystal (2010). "Toward a
551neurobiology of delusions." Prog Neurobiol 92(3): 345-369.
552Currie, G. (2000). Chapter: Imagination, delusion and hallucinations. Pathologies of belief.
553Malden, Blackwell Publishing: 167-182.
554Ellson, D. G. (1941). "Hallucinations produced through sensory conditioning." Journal of
555Experimental Psychology 28: 1-20.
556Frith, C. and R. J. Dolan (1997). "Brain mechanisms associated with top-down processes in
557perception." Philos Trans R Soc Lond B Biol Sci 352(1358): 1221-1230.
558Frith, C., Dolan, R.J. (2000). The role of memory in the delusions associated with schizophrenia.
559Cambridge, Harvard University Press.
560Frith, C. D. (1987). "The positive and negative symptoms of schizophrenia reflect impairments in
561the perception and initiation of action." Psychol Med 17(3): 631-648.
562Fry, B. R., N. Russell, R. Gifford, C. F. Robles, C. E. Manning, A. Sawa, M. Niwa and A. W. Johnson
563(2020). "Assessing Reality Testing in Mice Through Dopamine-Dependent Associatively Evoked
564Processing of Absent Gustatory Stimuli." Schizophrenia bulletin 46(1): 54-67.
565Garrison, J. R., P. Moseley, B. Alderson-Day, D. Smailes, C. Fernyhough and J. S. Simons (2017).
566"Testing continuum models of psychosis: No reduction in source monitoring ability in healthy
567individuals prone to auditory hallucinations." Cortex 91: 197-207.
568Gershman, S. J. (2019). "The Generative Adversarial Brain." Front Artif Intell 2: 18.
569Gershman, S. J., J. Zhou and C. Kommers (2017). "Imaginative Reinforcement Learning:
570Computational Principles and Neural Mechanisms." J Cogn Neurosci 29(12): 2103-2113.
571Grace, A. A. (1991). "Phasic versus tonic dopamine release and the modulation of dopamine
572system responsivity: a hypothesis for the etiology of schizophrenia." Neuroscience 41(1): 1-24.
573Gray, J. A. (1993). "Consciousness, schizophrenia and scientific theory." Ciba Found Symp 174:
574263-273; discussion 273-281.
575Gray, J. A. (1995). "Dopamine release in the nucleus accumbens: the perspective from
576aberrations of consciousness in schizophrenia." Neuropsychologia 33(9): 1143-1153.
577Gray, J. A. (1998). "Abnormal contents of consciousness: the transition from automatic to
578controlled processing." Adv Neurol 77: 195-208; discussion 208-111.
579Gray, J. A. (1998). "Integrating schizophrenia." Schizophr Bull 24(2): 249-266.

24
 Delusions and Imagination

580Gray, J. A. (2004). "On biology, phenomenology, and pharmacology in schizophrenia." Am J

581Psychiatry 161(2): 377; author reply 377-378.
582Gray, J. A., J. Feldon, J. N. P. Rawlins, D. Hemsley and A. D. Smith (1991). "The neuropsychology
583of schizophrenia." Behav Brain Sci 14: 1-84.
584Griffin, J. D. and P. C. Fletcher (2017). "Predictive Processing, Source Monitoring, and
585Psychosis." Annu Rev Clin Psychol 13: 265-289.
586Grossberg, S. (1982). "Processing of expected and unexpected events during conditioning and
587attention: a psychophysiological theory." Psychol Rev 89(5): 529-572.
588Grossberg, S. (2000). "How hallucinations may arise from brain mechanisms of learning,
589attention, and volition." J Int Neuropsychol Soc 6(5): 583-592.
590Guthrie, E. R. (1952). The Psychology of Learning. New York, Harper and Brothers.
591Hemsley, D. R. (1992). "Cognitive abnormalities and schizophrenic symptoms." Psychol Med
59222(4): 839-842.
593Hemsley, D. R. (1993). "A simple (or simplistic?) cognitive model for schizophrenia." Behav Res
594Ther 31(7): 633-645.
595Hemsley, D. R. (1994). "A cognitive model for schizophrenia and its possible neural basis." Acta
596Psychiatr Scand Suppl 384: 80-86.
597Hemsley, D. R. (1994). Percpetual and Cognitive Abonromalities as the Bases for Schizophrenic
598Symptoms. The Neuropsychology of Schizophrenia. A. S. David, Cutting J.C., Larence Erlbaum
599Associates: 97-118.
600Hemsley, D. R. (2005). "The development of a cognitive model of schizophrenia: placing it in
601context." Neurosci Biobehav Rev 29(6): 977-988.
602Hemsley, D. R. (2005). "The schizophrenic experience: taken out of context?" Schizophr Bull
60331(1): 43-53.
604Heyes, C. (2018). "Précis of Cognitive Gadgets: The Cultural Evolution of Thinking." Behav Brain
605Sci 42.
606Heyes, C. and J. M. Pearce (2015). "Not-so-social learning strategies." Proc Biol Sci 282(1802).
607Jardri, R., A. Pouchet, D. Pins and P. Thomas (2011). "Cortical activations during auditory verbal
608hallucinations in schizophrenia: a coordinate-based meta-analysis." Am J Psychiatry 168(1): 73-
60981.
610Johnson, M. K., Raye, C.L. (1981). "Reality Monitoring." Psychological Review 88(1): 67-85.
611Jones, S. H., J. A. Gray and D. R. Hemsley (1990). "The Kamin blocking effect, incidental learning
612and psychoticism." Br J Psychol 81 ( Pt 1): 95-109.
613Jones, S. H., J. A. Gray and D. R. Hemsley (1992). "Loss of the Kamin blocking effect in acute but
614not chronic schizophrenics." Biol Psychiatry 32(9): 739-755.
615Kamin, L. J. (1969). Predictability, surprise, attention and conditioning. Punishment and aversive
616behavior. R. Church and B. Campbell. New York, Appleton-Century-Crofts: 279-296.
617Kapur, S. (2003). "Psychosis as a state of aberrant salience: a framework linking biology,
618phenomenology, and pharmacology in schizophrenia." Am J Psychiatry 160(1): 13-23.
619Kaye, H. and J. M. Pearce (1987). "Hippocampal lesions attenuate latent inhibition and the
620decline of the orienting response in rats." Q J Exp Psychol B 39(2): 107-125.
621Kerfoot, E. C., I. Agarwal, H. J. Lee and P. C. Holland (2007). "Control of appetitive and aversive
622taste-reactivity responses by an auditory conditioned stimulus in a devaluation task: a FOS and
623behavioral analysis." Learn Mem 14(9): 581-589.

25
 Delusions and Imagination

624Kim, H. J. and H. Y. Koh (2016). "Impaired Reality Testing in Mice Lacking Phospholipase Cβ1:
625Observed by Persistent Representation-Mediated Taste Aversion." PLoS One 11(1): e0146376.
626King, R., J. D. Barchas and B. A. Huberman (1984). "Chaotic behavior in dopamine
627neurodynamics." Proc Natl Acad Sci U S A 81(4): 1244-1247.
628King, R., J. D. Raese and J. D. Barchas (1981). "Catastrophe theory of dopaminergic
629transmission: a revised dopamine hypothesis of schizophrenia." J Theor Biol 92(4): 373-400.
630Koh, M. T., P. S. Ahrens and M. Gallagher (2018). "A greater tendency for representation
631mediated learning in a ketamine mouse model of schizophrenia." Behavioral neuroscience
632132(2): 106-113.
633Koh, M. T. and M. Gallagher (2020). "Using internal memory representations in associative
634learning to study hallucination-like phenomenon." Neurobiology of Learning and Memory 175:
635107319.
636Kot, T. and M. Serper (2002). "Increased susceptibility to auditory conditioning in hallucinating
637schizophrenic patients: a preliminary investigation." J Nerv Ment Dis 190(5): 282-288.
638Lau, H. C. (2008). "A higher order Bayesian decision theory of consciousness." Prog Brain Res
639168: 35-48.
640Lisman, J. E. and A. A. Grace (2005). "The hippocampal-VTA loop: controlling the entry of
641information into long-term memory." Neuron 46(5): 703-713.
642Lodge, D. J. and A. A. Grace (2006). "The hippocampus modulates dopamine neuron
643responsivity by regulating the intensity of phasic neuron activation."
644Neuropsychopharmacology 31(7): 1356-1361.
645Lubow, R. E. and A. U. Moore (1959). "Latent inhibition: the effect of nonreinforced pre-
646exposure to the conditional stimulus." J Comp Physiol Psychol 52: 415-419.
647Mackintosh, N. J. (1975). "Blocking of conditioned suppression: role of the first compound trial."
648J Exp Psychol Anim Behav Process 1(4): 335-345.
649Mackintosh, N. J. (1975). "A theory of attention: Variations in associability of stimuli with
650reinforcement." Psychological Review 82(4): 276-298.
651Martin, J. R. and E. Pacherie (2019). "Alterations of agency in hypnosis: A new predictive coding
652model." Psychol Rev 126(1): 133-152.
653McDannald, M. and G. Schoenbaum (2009). "Toward a model of impaired reality testing in
654rats." Schizophr Bull 35(4): 664-667.
655McDannald, M. A., J. P. Whitt, G. G. Calhoon, P. T. Piantadosi, R.-M. Karlsson, P. O'Donnell and
656G. Schoenbaum (2011). "Impaired reality testing in an animal model of schizophrenia."
657Biological psychiatry 70(12): 1122-1126.
658McKay, R. (2012). "Delusional Inference." Mind & Language 27(3): 330-355.
659Miller, R. (1976). "Schizophrenic psychology, associative learning and the role of forebrain
660dopamine." Med Hypotheses 2(5): 203-211.
661Miller, R. (1993). "Striatal dopamine in reward and attention: a system for understanding the
662symptomatology of acute schizophrenia and mania." Int Rev Neurobiol 35: 161-278.
663Moran, P. M., M. M. Al-Uzri, J. Watson and M. A. Reveley (2003). "Reduced Kamin blocking in
664non paranoid schizophrenia: associations with schizotypy." J Psychiatr Res 37(2): 155-163.
665Moran, P. M., L. Owen, A. E. Crookes, M. M. Al-Uzri and M. A. Reveley (2007). "Abnormal
666prediction error is associated with negative and depressive symptoms in schizophrenia." Prog
667Neuropsychopharmacol Biol Psychiatry.

26
 Delusions and Imagination

668O'Tuathaigh, C. M., C. Salum, A. M. Young, A. D. Pickering, M. H. Joseph and P. M. Moran

669(2003). "The effect of amphetamine on Kamin blocking and overshadowing." Behav Pharmacol
67014(4): 315-322.
671Oades, R. D., D. Bunk and C. Eggers (1992). "Paranoid schizophrenics may not use irrelevant
672signals. The use of measures of blocking and of urinary dopamine." Acta Paedopsychiatr 55(3):
673183-184.
674Oades, R. D., B. Zimmermann and C. Eggers (1996). "Conditioned blocking in patients with
675paranoid, non-paranoid psychosis or obsessive compulsive disorder: associations with
676symptoms, personality and monoamine metabolism." J Psychiatr Res 30(5): 369-390.
677Pavlov, I. P. (1927). CONDITIONED REFLEXES: AN INVESTIGATION OF THE
678PHYSIOLOGICAL ACTIVITY OF THE CEREBRAL CORTEX.
679Pearce, J. M. and G. Hall (1980). "A model for Pavlovian learning: variations in the effectiveness
680of conditioned but not of unconditioned stimuli." Psychol Rev 87(6): 532-552.
681Pinkham, A. E., P. D. Harvey and D. L. Penn (2016). "Paranoid Individuals with Schizophrenia
682Show Greater Social Cognitive Bias and Worse Social Functioning Than Non-Paranoid Individuals
683with Schizophrenia." Schizophr Res Cogn 3: 33-38.
684Powers, A. R., 3rd, M. S. Kelley and P. R. Corlett (2017). "Varieties of Voice-Hearing: Psychics
685and the Psychosis Continuum." Schizophr Bull 43(1): 84-98.
686Powers, A. R., III, M. Kelley and P. R. Corlett (2016). "Hallucinations as top-down effects on
687perception." Biol Psychiatry Cogn Neurosci Neuroimaging 1(5): 393-400.
688Powers, A. R., C. Mathys and P. R. Corlett (2017). "Pavlovian conditioning-induced hallucinations
689result from overweighting of perceptual priors." Science 357(6351): 596-600.
690Raihani, N. J. and V. Bell (2019). "An evolutionary perspective on paranoia." Nat Hum Behav
6913(2): 114-121.
692Reed, E. J., S. Uddenberg, P. Suthaharan, C. D. Mathys, J. R. Taylor, S. M. Groman and P. R.
693Corlett (2020). "Paranoia as a deficit in non-social belief updating." Elife 9.
694Reichert, D. P., P. Series and A. J. Storkey (2013). "Charles Bonnet syndrome: evidence for a
695generative model in the cortex?" PLoS Comput Biol 9(7): e1003134.
696Rescorla, R. A. and A. R. Wagner (1972). A theory of Pavlovian conditioning: variations in the
697effectiveness of reinforcement and non-reinforcement. . Classical conditioning II: Current
698research and theory. A. H. Black and W. F. Prokasy. New York, Appleton-Century-Crofts.
699Ross, R. M., McKay, R., Coltheart, M., Langdon, R. (2016). "Perception, Cognition, and Delusion."
700Behavioral and Brain Science 39: 47-48.
701Rossi-Goldthorpe, R. A., Y. C. Leong, P. Leptourgos and P. R. Corlett (2021). "Paranoia, self-
702deception and overconfidence." PLoS Comput Biol 17(10): e1009453.
703Saddoris, M. P., P. C. Holland and M. Gallagher (2009). "Associatively Learned Representations
704of Taste Outcomes Activate Taste-Encoding Neural Ensembles in Gustatory Cortex." The Journal
705of Neuroscience 29(49): 15386-15396.
706Schultz, W. (2004). "Neural coding of basic reward terms of animal learning theory, game
707theory, microeconomics and behavioural ecology." Curr Opin Neurobiol 14(2): 139-147.
708Schultz, W., Dickinson, A. (2000). "Neural coding of prediction errors." Annual Review of
709Neuroscience.(23): 473-500.
710Seashore, C. E. (1895). "Measurements of illusions and hallucinations in normal life." Studies
711from the Yale Psychological Laboratory 3: 1-67.

27
 Delusions and Imagination

712Seth, A. K. (2014). "A predictive processing theory of sensorimotor contingencies: Explaining the
713puzzle of perceptual presence and its absence in synesthesia." Cogn Neurosci 5(2): 97-118.
714Sheffield, J., Suthaharan, P., Leptourgos, P. (2022). "Belief Updating and Paranoia in Individuals
715with Schizophrenia." Psycharxiv Preprint https://psyarxiv.com/3gyde/.
716Simons, J. S., R. N. Henson, S. J. Gilbert and P. C. Fletcher (2008). "Separable forms of reality
717monitoring supported by anterior prefrontal cortex." J Cogn Neurosci 20(3): 447-457.
718Spitzer, M. (1995). "A neurocomputational approach to delusions." Compr Psychiatry 36(2): 83-
719105.
720Sun, W. and J. Orchard (2020). "A Predictive-Coding Network That Is Both Discriminative and
721Generative." Neural Comput 32(10): 1836-1862.
722Suthaharan, P., E. Reed, P. Leptourgos, J. Kenney, S. Uddenberg, C. Mathys, L. Litman, J.
723Robinson, A. Moss, J. Taylor, S. Groman and P. Corlett (2021). "Paranoia and belief updating
724during the COVID-19 crisis." Nature Human Behaviour In Press.
725Takahashi, Y. K., C. Y. Chang, F. Lucantonio, R. Z. Haney, B. A. Berg, H. J. Yau, A. Bonci and G.
726Schoenbaum (2013). "Neural estimates of imagined outcomes in the orbitofrontal cortex drive
727behavior and learning." Neuron 80(2): 507-518.
728Teufel, C., N. Subramaniam, V. Dobler, J. Perez, J. Finnemann, P. R. Mehta, I. M. Goodyer and P.
729C. Fletcher (2015). "Shift toward prior knowledge confers a perceptual advantage in early
730psychosis and psychosis-prone healthy individuals." Proc Natl Acad Sci U S A 112(43): 13401-
73113406.
732Thein, T., R. F. Westbrook and J. A. Harris (2008). "How the associative strengths of stimuli
733combine in compound: summation and overshadowing." Journal of experimental psychology.
734Animal behavior processes 34(1): 155-166.
735Tolman, E. C. (1932). Purposive Behavior in Animals and Men. New York, Appleton-Century-
736Crofts.
737Waelti, P., A. Dickinson and W. Schultz (2001). "Dopamine responses comply with basic
738assumptions of formal learning theory." Nature 412(6842): 43-48.
739Wang, F., J. D. Howard, J. L. Voss, G. Schoenbaum and T. Kahnt (2020). "Targeted Stimulation of
740an Orbitofrontal Network Disrupts Decisions Based on Inferred, Not Experienced Outcomes." J
741Neurosci 40(45): 8726-8733.
742Wimmer, G. E. and D. Shohamy (2012). "Preference by association: how memory mechanisms
743in the hippocampus bias decisions." Science 338(6104): 270-273.
744Wu, J. L., R. P. Haberman, M. Gallagher and M. T. Koh (2020). "Probing for Conditioned
745Hallucinations Through Neural Activation in a Ketamine Mouse Model of Schizophrenia."
746Neuroscience bulletin 36(8): 937-941.
747Zeithamova, D., A. L. Dominick and A. R. Preston (2012). "Hippocampal and ventral medial
748prefrontal activation during retrieval-mediated learning supports novel inference." Neuron
74975(1): 168-179.
750

28