アイテム詳細

要旨

In this study, we used three computational algorithms to compute basis sets for natural image patches, such that each patch could be synthesized as a linear combination of basis functions. The two biologically plausible algorithms non-negative matrix factorization (NMF) and sparsenet (SPN) were compared to standard principal component analysis (PCA). We assessed human psychophysical performance at identifying natural image patches synthesized using different basis set sizes in each of the algorithms. We also computed the reconstruction error, which represents a simple objective measure of synthesis performance. We found that the reconstruction error was a good predictor of human psychophysical performance. Performance was best for PCA, followed by NMF and SPN despite large differences in basis function characteristics. All algorithms were well able to generalize to represent novel natural image patches. When applied to white noise patches instead of natural images, PCA and SPN outperformed NMF. This shows that of the three algorithms the one that is least biologically plausible (PCA) actually supported best psychophysical performance, suggesting that in the present study it is low-level quality of reconstruction that is the main determinant of psychophysical performance.