Abstract
Successful engagement with the world requires an intuitive understanding of the physical properties of objects including their mass. Previous behavioral and neuroimaging studies have shown that people can infer the mass of an object by observing its interactions with other objects. But people also learn and remember the mass of familiar objects. Here we used magnetoencephalography (MEG) to test whether and when the mass of familiar objects can be decoded from neural responses to static images of familiar objects. Specifically, we showed participants (N = 20) sequences of images at the center of the screen and instructed them to make relative mass judgements on randomly interspersed cued trials (“is this object lighter or heavier than the object in the previous trial?”), which were discarded from further analysis. We collected MEG responses to 80 object images (20-32 trials/image in each subject) in which mass was not confounded with real-world size and could not be decoded from early layers of a CNN (conv_1 and pool_1 layers of an ImageNet-trained VGG-16). We trained a linear SVM classifier on the average MEG signal within 10ms non-overlapping time bins to distinguish between the 25 lightest and the 25 heaviest objects in the dataset using 5-fold cross-validation across stimuli. We found weak but significant decoding of object mass (accuracy = 54.6%) starting at 350 ms after stimulus onset, much later than decoding of object identity (onset latency = 65 ms, accuracy = 55.1%; accuracy reaches 78.8% at 115 ms), or animacy (onset latency = 110ms, accuracy = 56.9%; accuracy reaches 83.8% at 170 ms). The relatively weak and late decoding of mass, even when encouraged by the task, suggests that representation of object mass may not result automatically and robustly from the rapid processes underlying the extraction of object identity or animacy.