ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards) (05/14/91)
Archive-name: ai/neural-nets/zemel-unsup-recog/1991-05-13 Archive: cheops.cis.ohio-state.edu:/pub/neuroprose/zemel.unsup-recog.ps.Z [128.146.8.62] Original-posting-by: ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards) Original-subject: Re: NNs in 2D shape recognition Reposted-by: emv@msen.com (Edward Vielmetti, MSEN) In article <1991May12.115515.7741@comp.vuw.ac.nz> Conrad.Bullock@comp.vuw.ac.nz (conrad Bullock) writes: >Greetings. >I am working on an honours project, aiming to apply neural networks to >recognising simple shapes in two-dimensional space, independent of >position, noise, rotation, magnification, and other transforms. >Does anyone have any good references in relevant work, particularly in >rotation-invariant recognition? Get Zemel and Hinton "Discovering Viewpoint-Invariant Relationships That Characterize Objects" from the /pub/neuroprose dir of cheops.cis.ohio-state.edu via anon ftp. Two networks are trained on images of the same object with various orientations, positions, and sizes. The networksd are trained to have high mutual information between their four outputs, which if properly trained must represent a coding of the orientation, position, and size of the of the object. While extracting that recoding might not be easy, we can use this property to have a network trained in this way reject other shapes it is exposed to since the outputs will no longer agree on the position, orientation, and size of the object. Thus multiple nets trained in this way can compete and see which pairs have the highest mutual information. The pair which has the highest mutual information will be the pair trained on the shape of the test object. -Thomas Edwards -- comp.archives file verification cheops.cis.ohio-state.edu -rw-r--r-- 1 3169 274 59069 Feb 23 21:43 /pub/neuroprose/zemel.unsup-recog.ps.Z found zemel-unsup-recog ok cheops.cis.ohio-state.edu:/pub/neuroprose/zemel.unsup-recog.ps.Z