VSSPoster_JaniniKonklepdf
VSSPoster_JaniniKonklepdf
Page 1 of 1
Konk Lab shared this file. Want to do more with it?
  1. Do number representations in the neural network match human perception?Connectedness decreases number representations Spatial grouping decreases number representations ConclusionsDaniel Janini and Talia KonkleHarvard University, Psychology Department, Vision Sciences LaboratoryWhat input analyzers yield number representations?Nasr et al. (2019): Number representations can emerge as a byproduct of exposure to objectsNeurons in parietal and prefrontal cortex“Neurons” in a convolutional neural networkDo units in object-trained neural networks represent number?Example Stimuli, replicating and extending Nasr et al. (2019)Selecting units with candidate number representations1. Get tuning curve for each stimulus set in each unit of MaxPool5 in AlexNetUnits in an object-trained neural network represent number according to Weber’s Law2. Select units with positive correlations between all stimulus sets4. Obtain tuning curves using new imagesTuning curves for units sorted by preferred number (PN)Normalized responseApproximate number representations emerge in object-trained convolutional neural networks and show human-like signatures of number discriminationStimuli: 14 dots with 0-7 connectionsNumber representations in CNN are susceptible to grouping and connectedness like human perceptionStimuli: 16 or 20 dots grouped in pairsHuman Behavioral Experiment:++Compare 1s500ms300ms500msParticipants (n=29) compared spatially grouped dots to randomly arranged dotsWhich image had more dots?CNNs trained on object classification have units that represent number like neurons in parieto-frontal cortexPreferred Number: 4Preferred Number: 8UncontrolledConvex HullCircumferenceDensityRadiusSurface TextureAreaThe same input analyzers that untangle object categories may also yield number representationsNumber presentedNumber presentedResponseResponse (Hz)PN: 1 (n=88)PN: 2 (n=15)PN: 3 (n=43)PN: 4 (n=81)PN: 5 (n=87)PN: 6 (n=101)3. Sort units by preferred numberKutter, et al. (2018)Agrillo and Bisazza (2017)Xu and Spelke (2005)Leaky and Sereno (2007)McCrink and Wynn (2004)Results replicate in VGG-1600.51P(Select grouped dots)Numerosity of randomly arranged dots5101520253016 dots20 dots130101301013010301013010130101Model ResultsBehavioral ResultsPN: 7 (n=93)PN: 8 (n=83)PN: 9 (n=63)PN: 10 (n=38)PN: 11 (n=25)PN: 12 (n=18)301030110301103011030110301101Number presented3010130101301013010130101301013010130101301013010130101301013010130101301013010130101PN: 25 (n=46)PN: 26 (n=40)PN: 27 (n=49)PN: 28 (n=107)PN: 29 (n=96)PN: 30 (n=10)PN: 19 (n=7)PN: 20 (n=5)PN: 21 (n=13)PN: 22 (n=15)PN: 23 (n=13)PN: 24 (n=27)PN: 13 (n=16)PN: 14 (n=11)PN: 15 (n=9)PN: 16 (n=5)PN: 17 (n=9)PN: 18 (n=2)30101Closest number representation201612820 dots16 dotsPreferred Number: 2Nieder (2016)
We use cookies to provide, improve, protect and promote our services. Visit our Privacy Policy and Privacy Policy FAQs to learn more. You can manage your personal preferences, including your ‘Do not sell or share my personal data to third parties’ setting using the “Customize cookies” button below.