HAL Id: hal-02004242
https://hal.archives-ouvertes.fr/hal-02004242
Submitted on 1 Feb 2019
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
BRAID: a new Bayesian word Recognition model with Attention, Interference and Dynamics
Julien Diard, Thierry Phénix, Sylviane Valdois
To cite this version:
Julien Diard, Thierry Phénix, Sylviane Valdois. BRAID: a new Bayesian word Recognition model with Attention, Interference and Dynamics. Annual Meeting of the Psychonomic Society, Nov 2018, New-Orleans, United States. �hal-02004242�
A
The structure of the BRAID model contains three classical levels of processing.
1. The letter sensory submodel implements low-level visual mechanisms involved in letter identification and letter position coding. Feature extraction is parallel over the input string, an acuity gradient is implemented symmetrically around fixation and location is distributed over adjacent letter positions, implementing lateral
interference between adjacent letters.
2. The letter perceptual submodel implements how information extracted from the sensory input accumulates gradually over time to create a percept, i.e. an internal representation of input letters.
3. The lexical knowledge submodel implements knowledge about the spelling of 40.481 English words (British Lexicon Project). A prior probability distribution represents word frequency. The lexical membership submodel implements a mechanism to decide whether or not the input letter-string is a known word, by observing how predicted spellings compare with perceived letters.
One major originality of BRAID is to assume the existence of a fourth level:
4. The visual attentional submodel implements a filtering mechanism between the letter sensory submodel and the letter perceptual submodel. Transfer of
information is modulated by the amount of attention allocated to each letter position.
References
Ginestet, E. (2016). Modélisation probabiliste de reconnaissance visuelle des mots : simulation d’effets d’amorçage et de compétition lexicale. Master’s thesis, Univ. Grenoble Alpes Ginestet, E., Phenix, T., Diard, J., & Valdois, S. (submitted). Modelling length effect for words in lexical decision: the role of visual attention
Phénix, T. (2018). Modélisation bayésienne algorithmique de la reconnaissance visuelle de mots et de l’attention visuelle. Université Grenoble Alpes. PhD thesis
Phénix, T., Valdois, S., and Diard, J. (2018). Reconciling opposite neighborhood frequency effects in lexical decision: Evidence from a novel probabilistic model of visual word recognition. In Proceedings of the 40th Annual Conference of the Cognitive Science Society, pages 2238–2243
Phénix, T., Valdois, S., & Diard, J. (submitted). Bayesian word recognition with attention, interference and dynamics
BRAID: a new Bayesian word Recognition model with Attention, Interference and Dynamics
Julien DIARD, Thierry PHÉNIX and Sylviane VALDOIS
Laboratoire de Psychologie et NeuroCognition
Univ. Grenoble Alpes, CNRS, LPNC UMR 5105, F-38000 Grenoble, France Contact: julien.diard@univ-grenoble-alpes.fr
Simulating behavioral effects Some examples
BRAID model definition
Simulating cognitive tasks by Bayesian inference BRAID: Bayesian Word Recognition with A ttention,
Interference and Dynamics
Lexical Knowledge
Submodel
Acuity Interference Final
Sensory Letter Submodel
C L O N E
Visual Attention Submodel
clone close alone
0 200 400 600 800 1000
0.0 0.2 0.4 0.6 0.8 1.0
Iteration
Recognitionprobability
Perceptual Letter Submodel 100 iterations 250 iterations
350 iterations 500 iterations
Lexical Membership
Submodel Word Is word?
Uniform Standard Narrow
C L O N E
0.0 0.2 0.4 0.6 0.8
Probability
C L O N E
0.0 0.2 0.4 0.6 0.8
Probability
C L O N E
0.0 0.2 0.4 0.6 0.8
Probability
BRAID is a probabilistic, hierarchical model:
Nodes represent probabilistic variables.
Arrows represent the dependency structure.
BRAID contains:
• Dynamic variables (over perceived letters Pt, over word identity Wt, over lexical
membership Dt), with accumulation of
perceptual evidence over time and gradual decay of information in the absence of
stimulation.
• Coupled dynamical chains over P, W and D:
as in coupled Hidden Markov Models,
Bayesian inference yields recurrent transfer of information: bottom-up perceptual
processing and top-down predictions and feedback.
• Submodels are linked by coherence variables (
l
variables, in white). Theycontrol information transfer throughout the model, acting as informational filters.
The visual-attentional submodel spatially controls where sensory processing is.
• Word frequency effect in WR and LD: more frequent words are recognized faster
QWT = P(WT | s1:T1:N g1:T µ1:TA 1:TA [ L1:T1:N = 1] [ P1:T1:N = 1]) /
266666 66666 66666 4
X
wT 1
hQwT 1T 1 P(WT | wT 1)i
YN
n=1
DP(LTn | wT), QPTn E
377777 77777 77777 5
Q0PTn = P(PTn | s1:T1:N g1:T µ1:TA 1:TA [ L1:T1:N = 1] [ P1:T1:N = 1]) QDT = P(DT | s1:T1:N g1:T µ1:TA 1:TA [ D1:T1:N = 1] [ P1:T1:N = 1])
Cognitive tasks are modeled by questions asked to BRAID, and solved by Bayesian inference.
Given: stimulus , gaze position , visual attention position , and spread
• Letter recognition (without lexical influence)
s1:T1:N g1:T µ1:TA 1:TA
Dynamical evolution from time t-1 Perceptual accumulation of evidence:
• Sensory decoding with lateral interference
• Attentional filtering
QPTn = P(PTn | s1:T1:N g1:T µ1:TA 1:TA [ P1:Tn = 1])
/
266666 66666 66666 66666 66666 4
X
pTn 1
hP(PTn | pTn 1) QPTn 1i
266666 66666 66664
↵n X
iTn
"
P( iTn )
P([iTn = pTn ] | sT1:N iTn gT)
# + (1 ↵n) 1
|DL|
377777 77777 77775
377777 77777 77777 77777 77777 5
↵n = P([CATn = 1] | [AT = n])
Dynamical evolution from time t-1 Perceptual accumulation of evidence:
Matching between predicted spelling and letter recognition
• Word superiority effect (WSE): letters are recognized faster in words than in nonwords
Frequencies 1
2 5 10 20 50 100 200
500 1000 2000 5000 10 000
0 200 400 600 800 1000
0.0 0.2 0.4 0.6 0.8 1.0
Iteration
Recognitionprobability
0 1 2 3 4
0 100 200 300 400 500
Log Frequency
numberofiteration
• Word recognition (WR)
• Letter recognition (with lexical influence): top-down feedback from lexical knowledge
• Lexical decision (LD)
• The word superiority effect is modulated if letter position is pre-cued, differently for word contexts (cueing slows down) and nonword contexts (cueing facilitates)
Context Word Nonword Isolated letter
0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
iteration
Letteridentificationprobability Data
BRAID
Word Nonword Isolated letter
0.0 0.2 0.4 0.6 0.8 1.0
Context
Letteridentificationprobability
Data BRAID
whole string cued letter
0.0 0.2 0.4 0.6 0.8 1.0
Context
Letteridentificationprobability
Context whole string cued letter
0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
iteration
Letteridentificationprobability
Data BRAID
whole string cued letter
0.0 0.2 0.4 0.6 0.8 1.0
Context
Letteridentificationprobability
Context whole string cued letter
0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
iteration
Letteridentificationprobability
context is a nonword context is a word
Other simulated effects:
• Variants on letter perceptibility: consonant-strings, duration of context presentation, length of context, context letter spacing (Phénix, 2018)
• Effects in
LD
: faster YES than NO responses, orthographic legality (Phénix, 2018)• Transposition effects in primed-WR, primed-LD and same-different tasks (Ginestet, 2016)
• Optimal viewing position in normal and impaired readers (Psychonomics 2018 posters)
• Length effects in LD (Ginestet et al. submitted)
• Opposite
neighborhood frequency effects
in LD (Phénix et al., 2018)Material and methods:
Replicates Norris (2006)
130 5-letter words from CELEX Randomly attributed to 13 frequency classes (1, 2, 5, …, 10,000)
Experiment repeated 13 times with frequencies permuted in a round-robin manner
Decision threshold set at .85
Material and methods:
Replicates Johnston (1978) 72 4-letter word pairs differing by a letter (e.g., LAST / LOST) Isolated letters in the same position: _A__ / _O__
Non-words built by scrambling letters of the context (e.g., LAST / SATL) and switching contexts (e.g., LAST / OASW)
Iteration found where the match is best between human data and simulation on the Word condition (arbitrarily chosen): The other conditions are model predictions.
Results: The WSE is simulated. Note that, whatever the chosen iteration, the predicted effect is in the correct direction.
Material and methods:
Replicates Johnston &
McClelland (1974)
Same 72 4-letter word pairs as above
To simulate cueing, attention mean µand gaze g are set to the cued position
Reference iteration chosen as above: best match between human data and simulation on the “whole string” condition.
Results:
Effect direction holds for other iterations; but see the initial predicted reversal in the word condition
Results: The simulation of the dynamics of perceptual evidence accumulation shows that the effect holds for the chosen and other decision threshold values. Effect also holds for lexical decision (simulations not shown here).
Letter perceptual submodel
λL2T
λTL3
λTL5
λTL1
P5T-1
P2T-1 P3T-1 P4T-1
P1T-1
P2T P3T P4T P5T
P1T
Visual attentional submodel
CA1T CA2T CA3T CA4T CA5T
AT
λP2T
λP3T
λP4T
λP5T
λP1T
Letter sensory submodel
S1T S2T S3T S4T S5T
I1T I2T I3T I4T I5T
GT
DI1T DI2T DI3T DI4T DI5T
Lexical knowledge
submodel
W T W T-1
L2T
L1T
L3T
Lexical membership submodel
D T CD1T
D T-1
CD2T
CD3T
CD4T
CD5T
λD2T
λD3T
λD4T
λD5T
λD1T
µTA σAT
L4T
L5T
λL4T