... Grundy$^\ddag$¹

Corresponding author: Department of Computer Science, Columbia University, 450 Computer Science Building, Mail Code 0401, 1214 Amsterdam Avenue, New York, NY 10027

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Cristianini$^\S$²

Work done while visiting UCSC.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... weights.³

The scaled dot product kernel gives better performance using a threshold that is optimized on the training set, so we report results for this threshold, rather than a threshold of 0.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... as⁴

An alternate formulation of support vector machines does not use an explicit bias b but makes the bias implicit by adding 1 to the kernel function. In this case, the hyperplane goes through the origin, and the optimization does not require the constraint $\sum \alpha_{i}y_{i}=0$. We use this implicit bias method in our experiments.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.