Fig. 2.

Multidimensional scaling (MDS) illustrates the benefit of denoising RDMs.

We apply classical MDS to visualize the similarity structure of the RDMs obtained for Participant 18. Each point is color-coded with regards to stimulus category (red: places, orange: objects, blue: bodies, cyan: faces) and shows the actual stimulus presented to the participant. To quantify the replicability of MDS results across Test and Re-test, we co-registered the two sets of results by first normalizing the scale of each MDS result and then rotating the Re-test MDS result to minimize the error with respect to the Test MDS result. The colored lines depict the distance between Test and Re-test results. Short lines indicate that an item's position in the representational space is stable across data halves. Long lines suggest that the underlying activity patterns contain noise that is corrupting the RDM results. Denoising substantially improves test-retest replicability of MDS results and increases clustering of similar points. Note that this is just an illustrative example suggesting the potential impact of denoising on neuroscientific results; systematic evaluation of denoising performance for all participants and experiments is provided in later figures.