Evaluation of Feature Learning in a Biological ... - Semantic Scholar

2 downloads 51 Views 51KB Size Report
AG, Compaq/Digital Equipment Corporation, Eastman Kodak Company, Honda R&D Co., Ltd., ITRI,. Komatsu Ltd., Merrill-Lynch, Mitsubishi Corporation, NEC ...
Evaluation of Feature Learning in a Biological Model with Real World Tasks Thomas Serre, Maximilian Riesenhuber, and Jennifer Louie

The Problem: To evaluate a biological model of object recognition that uses feature learning to distinguish between classes of objects. Motivation: On the computational neuroscience side, applying a biologically plausible model of object recognition to a real world task, namely face detection, tests out whether the concepts behind the model hold. Current models account for learning only at the level of view-tuned units. This work explores learning at lower stages in the visual system. This system can also help improve current object recognition algorithms, which are mainly computer vision techniques. Computer vision algorithms use a centralized approach, where a global window scans over the image and searches for the target object. The biological model uses distributed processing through local receptive £elds, whose outputs are pooled together. Part of the project is to compare the performance of the two approaches. Previous Work: A biological model of the visual system is the HMAX model described in [1]. HMAX models the V1 area to the IT area of the brain where object recognition is thought to occur. HMAX’ s structure of alternating levels of (1) combining lower inputs to create more complex feature detectors and (2) taking the max over inputs from these detectors allows the system to increase speci£city while maintaining invariance. In the HMAX model, object speci£c learning only occurs at the view-tuned units level. There has been previous work done in object recognition with computer vision systems. A system that describes an object class in terms of wavelets and trains using a support vector machine(SVM) is described in [3]. A component-based face detection system that automatically learns components by growing image parts until error in detection is minimized is described in [2]. In this system, these image parts and their geometric arrangement are used to train a two-level SVM. Approach: Our system is an extension of the HMAX model described in the previous section. While the HMAX model uses the same features to feed into the view-tuned unit level for all object classes, the extension permits learning of features speci£c to the object class . Fig. 1 shows where the feature learning takes place in the HMAX hierarchy. Currently, we are working on face detection and later will extend the system to other objects. To learn features, patches (whose size and number are parameters) are extracted from the training faces. The location of the patches are randomized with each run. Using K-means, the patches from all the training faces are clustered to produce a vector of prototypes. An SVM classi£er is trained on these prototypes. We chose to use an SVM classi£er so that this representation can be easily compared with computer vision representation. Later on, we hope to replace the SVM with a simpler, more biologically plausible classi£er. The training faces used are synthetic faces [4]. The test sets consist of synthetic faces, synthetic faces with background, and real faces. Aspects of the feature learning that we are experimenting with are: in the training stage, training with both faces and non-faces, and testing the invariance of the model by varying face scales and running tests with different pooling bands. Dif£culty: Learning in a hierarchy. Impact: There are three main impacts from this work:(1) For neuroscience, a model of feature learning will give ideas on how feature learning occurs in the brain and what kind of features neurons respond to. (2) A study between computer vision and biological vision systems may lead to better techniques for maintaining invariance in vision systems. (3) The use of a simpler, yet as effective classi£er than SVM. Future Work: Learning throughout the hierarchy. Research Support: Reserach at CBCL is sponsored by grants from: Of£ce of Naval Research (DARPA) Contract No. N00014-00-1-0907, National Science Foundation (ITR/IM) Contract No. IIS-0085836, National Science Foundation (ITR) Contract No. IIS-0112991, National Science Foundation (KDI) Contract No. DMS9872936, and National Science Foundation Contract No. IIS-9800032. Additional support was provided by: AT&T, Central Research Institute of Electric Power Industry, Center for e-Business (MIT), DaimlerChrysler AG, Compaq/Digital Equipment Corporation, Eastman Kodak Company, Honda R&D Co., Ltd., ITRI, Komatsu Ltd., Merrill-Lynch, Mitsubishi Corporation, NEC Fund, Nippon Telegraph & Telephone, Oxygen,

Figure 1: The HMAX model with the addition of feature learning between the S2 and C2 levels. Siemens Corporate Research, Inc., Sumitomo Metal Industries, Toyota Motor Corporation, WatchVision Co., Ltd., and The Whitaker Foundation. M.R. is supported by a McDonnell-Pew Award in Cognitive Neuroscience. References: [1]

M. Riesenhuber and T. Poggio. Hierarchical models of object recognition in cortex. Nature Neuroscience, 2:1019-1025, 1999.

[2]

B. Heisele, T. Serre, M. Pontil, and T. Poggio. Component-based Face Detection. Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 1:657-62, 2001.

[3]

C. Papageorgiou and T. Poggio. A Trainable System for Object Detection International Journal of Computer Vision, 38(1), 15-33, 2000.

[4]

T. Vetter. Synthesis of novel views from a single face. International Journal of Computer Vision, 28(2): 103-116, 1998.

Suggest Documents