Spatial pooling for greyscale images
- PDF / 1,008,279 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 59 Downloads / 182 Views
ORIGINAL ARTICLE
Spatial pooling for greyscale images John Thornton • Andrew Srbic
Received: 28 October 2011 / Accepted: 23 February 2012 / Published online: 18 March 2012 Ó Springer-Verlag 2012
Abstract It is a widely held view in contemporary computational neuroscience that the brain responds to sensory input by producing sparse distributed representations. In this paper we investigate a brain-inspired spatial pooling algorithm that produces sparse distributed representations of spatial images by modelling the formation of proximal dendrites associated with neocortical minicolumns. In this approach, distributed representations are formed out of a competitive process of inter-column inhibition and subsequent learning. Specifically, we evaluate the performance of a recently proposed binary spatial pooling algorithm on a well-known benchmark of greyscale natural images. In the process, we augment the algorithm to handle greyscale images, and to produce better quality encodings of binary images. We also show that the augmented algorithm produces superior population and lifetime kurtosis measures in comparison to a number of well-known coding schemes and explain how the augmented coding scheme can be used to produce highfidelity reconstructions of greyscale input. Keywords Hierarchical temporal memory Sparse distributed representations Spatial clustering Sparse coding Self-organisation
This paper extends work previously published as conference proceedings [22]. J. Thornton (&) Institute for Integrated and Intelligent Systems, Griffith University, Queensland 4222, Australia e-mail: [email protected] A. Srbic School of Information and Communication Technology, Griffith University, Queensland 4222, Australia e-mail: [email protected]
1 Introduction Advances in computational neuroscience over the last 20 years have produced increasingly realistic and viable models of the functioning of the mammalian neocortex. These advances connect directly with the original inspiration of artificial intelligence: to build machines that capture and exhibit the principles on which natural intelligence operates. Neurological evidence concerning the hierarchical structure of bottom-up and top-down processing in sensory cortex has already produced a number of promising machine learning applications, including Hinton’s work on multilayer generative models [10], and Serre and Poggio’s neuromorphic approach to computer vision [20]. In the current paper, we investigate a more general model of neocortical function known as hierarchical temporal memory (HTM) [8]. HTMs were first proposed by Jeff Hawkins in 2004, but a practical computational description of their low-level functioning has only recently been developed [7]. Our task is to evaluate the spatial pooling component of this algorithm in terms of its ability to robustly and efficiently encode Willmore and Tolhurst’s well-known benchmark of greyscale natural scene images [23]. The HTM model is grounded in the view that the broadly uniform structure of the ne
Data Loading...
