Biologically plausible saliency networks for object recognition

Abstract

Event Abstract Back to Event Biologically plausible saliency networks for object recognition Various biologically inspired neural network architectures have been proposed for object recognition. These networks are usually hierarchical, solving the recognition problem through a sequence of processing stages that implement different trade-offs between selectivity and invariance. Recently, HMAX networks (which differ from earlier models mostly through the choice of pooling function) have been shown to achieve state-of-the-art performance across a number of recognition tasks [4, 3]. In this work, we introduce an alternative family of biologically plausible networks for visual recognition, based on the principle of discriminant saliency (DiscSal). This principle has been shown to provide a unified formulation for bottom-up and top-down processing in visual cortex. In the bottom-up realm, extensive evidence is now available on the ability of optimally discriminant center-surround detectors to predict the psychophysics of human saliency. These predictions are quantitative, with precision far superior to those of other popular saliency models [2]. With respect to top-down processing, discriminant saliency has been shown to detect regions of the visual field which are more informative for object recognition than those provided by bottom-up interest point detectors. This work extends the DiscSal principle to the design of object recognition networks with a 1-to-1 mapping to the standard model of cortical networks [1]. A comparison to their HMAX counterparts reveals that DiscSal networks have identical structure but several advantages. First, it is shown that all their computations are widely accepted as biologically plausible. Second, it is shown that, under DiscSal, all network units have a precise computational interpretation: simple cells implement an optimal decision rule (classification), while complex cells compute the discriminant power of visual features (more precisely the mutual information between the responses of each feature and its class label). Third, it is shown that DiscSal networks can be tuned to the solution of multiple vision problems. We consider two examples, by deriving the optimal S (simple) and C (complex) units for bottom-up and top-down saliency. This is particularly useful for object recognition, where it is shown that bundles of S and C units can be tuned to detect locations of the visual field that are informative for the presence of the objects of interest. Finally, an extensive experimental comparison of recognition performance shows that a single-layer DiscSal network matches the previous best results obtained with a highly optimized two-layer HMAX network, on standard image classification benchmarks. References 1. M. Carandini, J. Demb, V. Mante, D. Tolhurst, Y. Dan, B. Olshausen, J. Gallant, and N Rust. Do we know what the early visual system does? J. Neurosci., 25, 2005.2. D. Gao, V. Mahadevan, and N. Vasconcelos. On the plausibility of the discriminant center-surround hypothesis for visual saliency. Journal of Vision, 8(7):1-18, 6 2008.3. J. Mutch and D. Lowe. Object class recognition and localization using sparse features with limited receptive fields, IJCV, 80:45-57, 2008.4. T. Serre, L. Wolf, S. Bileschi, M. Riesenhuber, and T. Poggio. Robust object recognition with cortex-like mechanisms. IEEE Trans. on PAMI, 2007 Conference: Computational and systems neuroscience 2009, Salt Lake City, UT, United States, 26 Feb - 3 Mar, 2009. Presentation Type: Poster Presentation Topic: Poster Presentations Citation: (2009). Biologically plausible saliency networks for object recognition. Front. Syst. Neurosci. Conference Abstract: Computational and systems neuroscience 2009. doi: 10.3389/conf.neuro.06.2009.03.099 Copyright: The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers. They are made available through the Frontiers publishing platform as a service to conference organizers and presenters. The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated. Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed. For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions. Received: 02 Feb 2009; Published Online: 02 Feb 2009. Login Required This action requires you to be registered with Frontiers and logged in. To register or login click here. Abstract Info Abstract The Authors in Frontiers Google Google Scholar PubMed Related Article in Frontiers Google Scholar PubMed Abstract Close Back to top Javascript is disabled. Please enable Javascript in your browser settings in order to see all the content on this page.