A Comparison of Dynamic Reposing and Tangent Distance for Drug Activity Prediction
Thomas G. Dietterich, Ajay N. Jain, Richard H. Lathrop, Tomás Lozano‐Pérez · 1994
In drug activity prediction (as in handwritten character recognition), the features extracted to describe a training example depend on the pose (location, orientation, etc.) of the example. In handwritten character recognition, one of the best techniques for addressing this problem is the tangent distance method of Simard, LeCun and Denker (1993). Jain, et al. (1993a; 1993b) introduce a new technique---dynamic reposing---that also addresses this problem. Dynamic reposing iteratively learns a neural network and then reposes the examples in an effort to maximize the predicted output values. New models are trained and new poses computed until models and poses converge. This paper compares dynamic reposing to the tangent distance method on the task of predicting the biological activity of musk compounds. In a 20-fold cross-validation, dynamic reposing attains 91% correct compared to 79% for the tangent distance method, 75% for a neural network with standard poses, and 75% for the neares...