ATR International / JST Presto Program
A hierarchical statistical model of natural images explains tuning properties in V2
Thursday 20th of November 2014 at 12:00pm
Although previous theoretical and experimental studies suggested a tight relationship between natural image statistics and neural representations in V1, whether such relationship exists in higher areas has not been clarified. Here, we present a theoretical investigation in which certain statistical learning models can explain simultaneously several experimental findings on macaque V2. Concretely, we built a family of sparse coding models on top of a fixed V1-like model with varied network configurations, and trained them with image patches extracted from a huge natural image dataset. After learning, the units in the top layer, robustly across the family of models, represented detectors of contour, texture, and corner features in conjunction with certain four types of inhibition patterns. Moreover, these units, also robustly, exhibited response properties qualitatively and quantitatively compatible with three major neurophysiological results: 1) local orientation integration with a U-shaped distribution of orientation differences consistent with (Anzai et al., 2007), 2) selectivities to angles with response specificity to one componential orientation as in (Ito and Komatsu, 2004), and 3) length and width suppression properties with a bimodal joint distribution of length and width suppression indices as shown in (Schmid et al., 2014). These results indicate a close relationship between natural image statistics and neural representations in V2. A formal classification of the representation patterns of the model units also offers additional insights into the quantitative data in the existing experiments.
Joint work with Aapo Hyvarinen.
Join Email List
You can subscribe to our weekly seminar email list by sending an email to
email@example.com that contains the words
subscribe redwood in the body of the message.
(Note: The subject line can be arbitrary and will be ignored)