1、A Neural Network Model of Subliminal Priming,Howard Bowman Computing Laboratory, University of Kent at Canterbury,Collaborators: Friederike Schlaghecken (Psychology, Univ Warwick) Adam Aron (Psychiatry, University of Cambridge) Prof. Martin Eimer (Psychology, Birkbeck College) Prof. Phil Barnard (Co
2、gnition and Brain Sciences Unit),Contents,Connectionism Now Subliminal Priming Eimer and Schlaghecken Pattern Masked Priming Lateral Inhibition and Opponent Processes The Model Predictions and Future Directions,Connectionism - The Claim,Connectionism claims to offer a bridge between cognition and th
3、e brain. How complex behaviour emerges from the interaction of a multitude of computationally unsophisticated (one might even say dumb) units.,Is this claim justified?,Arguments against Connectionism Neural networks are not actually biologically plausible, e.g. problems with backpropagation How are
4、computational implementations related to psychological theories? Does the model work because of the theory it realises or because of hidden implementation assumptions? Models can do anything, e.g. backpropagation can learn any computable function!,Connectionism Now,Biological plausibility all models
5、 are abstractions connectionist abstractions are becoming more grounded, e.g. spiking neurons and biologically plausible learning re-circulation algorithms instead of backpropagation Hebbian learning, sparse representations and sparsification techniques,Reducing the degrees of freedom,work closely w
6、ithin context of existing cognitive theory; apply biological constraints; systematic and transparent parameter setting (e.g. do statistics on model); make clear the key mechanisms involved; testable predictions from the models (especially counter-intuitive ones).,Theoretical Background,Two central t
7、heoretical issues the role of conscious control in visuomotor performance the “cognitive levels” at which inhibition functions,Issue 1: Consciousness and Visuomotor Performance,Tight coupling of vision and action actions planned on basis of visual information action execution guided by vision The de
8、bate, “To what extent is conscious experience a prerequisite for the control of visuomotor performance?” “Is there a direct, below conscious, link from vision to action?”,Direct Parameter Specification Hypothesis Neumann & Klotz,94,Tentative Support for Direct Parameter Specification (1),Blindsight
9、Weiskrantz et al,74 above chance visual guided action in absence of visual awareness Visual Illusions Carey,01 movements (e.g. grasp apertures) resist visual illusions Visual Form Agnosia Milner et al,91 profound deficits in object recognition, but intact visuomotor performance,Note: dissociation do
10、rsal (where) stream ventral (what) stream but subcortical routes also significant,Support for Direct Parameter Specifiction (2),More direct evidence provided by masking experiments. Two varieties, Metacontrast masked priming Pattern masked priming,Metacontrast Masked Priming,Fehrer & Raab (1962), Ne
11、umann & Klotz (1994), Vorberg (2002) For example, Neumann & Klotz,94 subjects not told of presence of prime subjects either respond to diamonds or squares respond left or right depending upon target position target stimulus metacontrast masks the prime strict criteria for perception of prime - signa
12、l detection,Neumann and Klotz Metacontrast Priming Paradigm Neumann & Klotz,94,Results,Positive compatibility results, compatible trials yield behavioural benefits (both errors and reaction times) incompatible trials yield behavioural costs,Signal Detection Blocks,Signal detection blocks follow reac
13、tion time blocks (rules out learning) subjects asked to state whether prime present on 5 point scale, e.g. “I am pretty sure that prime was present” signal detection gives d-prime statistically equivalent to zero, i.e. no phenomenological experience of prime,Implications,“ a stimulus can have access
14、 to the motor system and activate or even start an intended, planned response without being represented in consciousness.” Neumann & Klotz,94 Note further, this preconscious processing requires integration of form (diamonds vs squares) and position (left vs right) information. More than just a prese
15、nce / absence judgement.,what - where dissociation?,Issue 2 - Levels of Inhibition,Inhibitory mechanisms certainly ubiquitous in brain, e.g. GABAergic interneurons throughout cortex and subcortical regions. Typically, psychological theories situate inhibitory mechanisms at level of attentional (exec
16、utive) function, e.g. (working memory) Baddeleys central executive Shallices Supervisory Attentional System,Inhibition as a (pre)frontal lobe function?,Inhibition and task / set shifting,Neuropsychological example - frontal lobe damage yields perseveration errors (e.g. Wisconsin Card Sorting task) a
17、nd increased distractibility. Psychological example - negative priming (conscious inhibition of distractors) Inhibition argued to be central to conscious attentional control QUESTION: could inhibition arise at level of direct parameter specification?,Eimer and Schlagheckens Pattern Masked Priming Ta
18、sk,response buttons under left and right index fingers; basic stimuli -neutral stimuli - and and ,A subliminal priming paradigm, (right response) and (left response);,(other masks explored, e.g. classic pattern masks),16 ms,100 ms,100 ms,Time,PRIME,MASK,TARGET,NOTE: Prime is subliminal (verified by
19、forced choice blocks, which follow RT-blocks.),Conditions,Eimer 99 - Results,Implications,negative compatibility; behavioural costs on compatible trials and benefits on incompatible trials; candidate explanation: inhibitory processes at work; suppression of response activation (even before response
20、fires); supporting evidence from EEG study.,Eimer 99 - LRPs,Further Implications,LRP shows it is more than just sensory priming, i.e. residual perceptual activation. Prime induced activation propagated right through to response systems.,Candidate Explanation,low-level inhibitory mechanism “emergency
21、 brake” - suppress response once visual evidence for that response removed possibly part of a larger “clearing-up” mechanism - suppress activation traces of completed responses in order to enable sequences of co-ordinated actions.,Data to reproduce,short mask-target SOAs (0-32ms) yield positive comp
22、atibility (target and mask presented together) longer mask-target SOAs (64ms - 150ms (ish) yield negative compatibility low strength primes yield positive compatibility. Reduce strength by, presenting prime in periphery, or, presenting prime centrally, but overlaid with random-dot degradation,Furthe
23、r Data to Reproduce,forced choice - at chance forced choice blocks both with and without target forced choice blocks follow reaction time response blocks (thus, learning to detect prime not an explanation) QUESTION? - how can priming affect target response speeds but not forced choice judgements?,NO
24、TE: Signal-detection theory not used.,Further LRP Data,Schlaghecken and Eimer data,Mechanism (1) - competition and masking,masked and masking stimuli compete for shared neural resources, see Keysers & Perrett,02 neural trace of prime rapidly suppressed when mask presented implementation possibilitie
25、s, lateral inhibition gating mechanism,Mechanism (2) - response competition,responses compete in a winner take all fashion. only one response can be executed. sustains (in fact, accentuates) response separation, c.f. M00 condition. implemented through lateral inhibition between response nodes.,respo
26、nse 1,response 2,inhibitory,excitatory,Mechanism (3) - Opponent Processes,previously investigated in a number of models, e.g. negative priming and inhibition of return Houghton & Tipper,94 serial order in working memory and in motor action sequencing Houghton,90,An Opponent Circuit,Response Node,Exc
27、itatory Link,Inhibitory Link,(Opponent) OFF Node,Inhibitory feedback can be threshold gated,OFF node just an inhibitory interneuron,Mechanism (4) - S-R Binding,nodes in relevant stimulus-response pathways pre-activated and hence foregrounded from the set of possible S-R bindings called response-set
28、delineation in Bowman et al,02 implemented by giving backgrounded nodes a strongly negative bias,The Network,15,14,1,3,2,Perception Layer,Mask/ Neutral,Response Selection Layer,6,5,LEFT ON,RIGHT ON,7,8,OFF,OFF,Perceptual Pathways,Response Selection: difference between response node activation,Exampl
29、e (left compatible): 1 cycle of ; 6 cycles ofthe mask; 6 cycles of .,Perceptual Pathways: apply time averaging to perceptual input,Formal Parameters,Time averaging activation function:,Sigmoidal, (squashes activation into range 0 to +1),activation on cycle i+1,regulates time averaging,input to node
30、on cycle i,t set to 0.3,Basic Results,difference between response nodes gives separation; similar pattern to LRP.,Response Time Comparison,Assuming: (i) one cycle corresponds to 16.6666 ms; (ii) selection criteria -separation magnitude (absolute value) - 0.4; (iii) latency of 200 ms compared to Eime
31、r data.,NOTE: RTs currently very approximate!,(mean response times),Reduced Strength Prime,Prime induced response activation does not cross opponent circuit threshold; reproduces basic switch to positive compatibility,Observations (i),model RTs in right general ball park (parameter optimizations wou
32、ld improve these); RT difference between conditions is good; time course of separation close to LRP profile.,Observations (ii),explanation of forced choice results, selection criteria is the models analogue of super / subliminality threshold (simplification since just located at action end). while s
33、election criteria not satisfied, no evidence available for decision process, i.e. at chance residual activation from prime only affects outcome if it is built upon (since it influences speed with which threshold (selection criteria) is crossed) one reason for selection criteria is to ensure backgrou
34、nd fluctuations do not yield overt responses,Relationship to Houghton and Tipper Model,inhibition modulated by high level attentional processes in HT94; selection of target from distractor in negative priming; orienting system in inhibition of return; our model - a direct (low level) link from perce
35、ption to action - inhibition is a “dumb” mechanism (not directed by high-level attention).,Further Work,Biological plausibility - fMRI studies suggest basal ganglia as locus of inhibition broaden scope of model - locate pattern and metacontrast masked priming in same modelling framework modelling er
36、ror rates,Conclusions,Masked priming data fits with a theory of consciousness in which, “when we apperceive the stimulus, we have usually already started responding to it; our motor apparatus does not wait for consciousness, but does restlessly its duty, and our consciousness watches it and is not e
37、ntitled to order it about.” Munsterberg,1889 Two largely independent effects of a stimulus, determines a motor response has later effect in consciousness,Also see Libets work and implicit memory literature,Further Conclusions,Effects not restricted to specific stimulus-response connections Neumann &
38、 Klotz,94 rapidly modified through instruction cues cannot be explained through automaticity,Discussion Points,Why this model? Clearly (even at psychological level) there are many models which could satisfy the data (even infinitely many); In favour of model,+ Simple and canonical; + Opponent networks used to model a number ofinhibitory phenomena; + Increasing body of empirical data explained.,
