Using Generative Models for Handwritten Digit Recognition

By Michael Revow, Christopher K. I. Williams, Geoffrey E. Hinton. 1996

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Image Parsing: Unifying Segmentation, Detection, and Recognition (2003, unknown): “Similar models were proposed for text ([14]) but cannot be used here because they generate image features and not intensities.”
Training Products of Experts by Minimizing Contrastive Divergence (2002, unknown): “The error rate is 1.1% which compares very favorably with the 5.1% error rate of a simple nearest-neighbor classifier on the same training and test sets and is about the same as the very best classifier based on elastic models of the digits ( Revow, Williams, & Hinton, 1996 ).”
Toward automatic phenotyping of developing embryos from videos (2005, background): “Another important task is to automatically extract quantitative measurements such as the number of cells, the relative positions of the cell nuclei, the time between each cell division, etc... Ultimately, one may want an automated system for classifying the movies into a number of known scenarios…”
Human-level concept learning through probabilistic program induction (2015, methodology): “6B, new concepts (from type)], which is not significantly different from chance [t(34) = 0.”
Continuous latent variable models for dimensionality reduction and sequential data reconstruction (2001, background): “There are other applications where models based on continuous variables have performed well with discrete data, such as self-organising maps for text categorisation (Kohonen, 1995), or some generative models for handwritten character recognition (Revow et al., 1996; Hinton et al., 1997).”
One shot learning of simple visual concepts (2011, methodology): “…widely in length so they were reduced to a common dimensionality by fitting a cubic B-spline with 10 control points and clustering was done in this new space (Revow et al., 1996; Branson, 2004).3 Strokes are direction specific, meaning a left to right line and a right to left line are different.”
One-shot learning by inverting a compositional causal process (2013, methodology): “But HBPL also learns a significantly more complex representation than earlier models, which used only one stroke (unless a second was added manually) [24, 10] or received on-line input data [9], sidestepping the challenging parsing problem needed to interpret complex characters.”
Data driven image models through continuous joint alignment (2006, unknown): “[14] developed a model for characters by hand-specifying and then adapting a set of “control points,” defining the probability of an observed character as a function of the deviation of the character from these predefined control points.”
Point Set Registration: Coherent Point Drift (2009, methodology): “To overcome the ICP limitations, many probabilistic methods were developed [5], [6].”