Article review by Alexander Backus Distributed representations meeting article review.

Aim: Investigate whether an object-based system of scene recognition exists in LOC and/or PPA

Article review by Alexander Backus

Distributed representations meeting article review

Experimental design E1

104 scenes

208 object images (exemplars) associated with the scenes

Presented interleaved for 1s with2s ISI

Subjects silently named each scene/object

If scene representations in LOC (LO + pF) and/or PPA are built from the constituent exemplars, it is possible to classify scene-evoked patterns with combinations of object-evoked patterns.

Scene classification analysis

Results E1

Confound: for scenes, subjects might have attended serially to the single constituent objectsE2: faster stimulus presentation (150ms)

Results E2

Confound: relationship between scene- and object-evoked patterns reflect template development for object search?E3: first a block with only scenes, then a block with only objects

Results E3

Other analyses

Which common features of scenes and objects does LO encode? Visual/semantic?Classify each object category on the basis of patterns evoked by the other object category from the same context:

Train on: & , test on: &

LO primarily encodes visual features for short stimulus presentation times, but also semantic features for longer presentation times

Other analyses

Is there information outside of the ROIs?Whole-brain searchlight procedure

LO is unique among occipitotemporal visual areas in possessing an object-based scene recognition mechanism

Motor-related

LO

PPA

pF

Other analysesIs an object-based scene recognition mechanism actually used?4AFC task with briefly presented scenes (50ms)Zero, one or two of the constituent objects were obscured

Significant effect of number of objects removed on performance and RT

Scene identification falters when both of the exemplars are removed, but only if enough of the image is obscured to simultaneously affect image-based system

Conclusions

An object-based scene recognition mechanism in LO

PPA concerned with large-scale spatial featuresLO patterns are built from individual object patterns