Sampling Enhanced Reverse Correlation for the Reconstruction of Latent Representations

Roop, Benjamin

Etd

Sampling Enhanced Reverse Correlation for the Reconstruction of Latent Representations

Public

Internal representations are widely characterized using reverse correlation, a technique capable of producing unconstrained estimates of the representation, all on the basis of subjects’ responses to random stimuli. However, employing reverse correlation often entails collecting thousands of stimulus-response pairs, which severely limits the breadth of studies that are feasible using the method. Current techniques to improve efficiency bias the outcome and ultimately limit the truth to the reconstruction. Here, three techniques are used to increase the efficiency of reverse correlation. Stimulus whitening, a statistical procedure that decorrelates stimuli, provides greater than 85% improvement in efficiency for a given estimation accuracy and a two- to five-fold increase in accuracy for a given sample size. Compressive sensing, an advanced signal processing technique designed to improve sampling efficiency based on sparsity, improves the accuracy of reconstructed cognitive representations and dramatically reduces the required number of stimulus-response pairs in both simulations and on human subject response data. Autoencoders, a type of artificial neural network, increase the reconstruction quality to a level that necessitates only collecting 10% of the previously needed samples. Improving the efficiency of reverse correlation in such ways may enable a broader scope of investigations of perceptual mechanisms and could improve representation reconstruction throughout the field of neuroscience and beyond.

Creator