T. Sche#er and R. Herbrich. Unbiased assessment of learning algorithms. In IJCAI97, pages 798--803, 1997. 7  Home/Search   Document Details and Download   Summary   Related Articles   Check

....true error rates of the resulting hypotheses exactly. In order to obtain almost unbiased error estimates of the resulting decision tree algorithm, it is therefore necessary to wrap a second layer of 10 fold cross validation around the cross validation wrapper that selects the number of leaves; see Sche er and Herbrich (1997) and Kohavi and John (1997) for more detailed discussions on almost unbiased error estimation. All error rates which we will refer to in the following are measured in an outer cross validation wrapper and are only subject to a very small pessimistic bias. Note that disagreements between prediction ....

Scheer, T., & Herbrich, R. (1997). Unbiased assessment of learning algorithms. Proceedings of the Fifteenth International Joint Conference on Articial Intelligence (pp. 798-803).

.... size grows (Kearns et al. 1997) Most practical learning algorithm incorporate some form of complexity penalization (like tree pruning, weight decay) and account for this problem by providing a parameter that can be adapted for each learning problem e.g. by cross validation (Ng, 1997; Scheffer Herbrich, 1997). Bayesian learners (Berger, 1985) focus on the posterior probability of a function f having generated the sample S: P (f jS) Under certain ideal conditions, one can, under high computational effort, derive the Bayes hypothesis from P (f jS) which is guaranteed to have the least generalization ....

Scheffer, T., & Herbrich, R. (1997). Unbiased assessment of learning algorithms. In IJCAI-97, pp. 798--803.

.... practical relevance (the resulting equation reads like a sample size which is enormous plus linear in the number of parameters is required to make sure that the bias is reasonably small ) The following consideration, however, leads to an accurate quantification of the parameter adaptation bias [5]: Suppose that the entropy of the holdout data is zero (i.e. all class labels are equal) Even if our hypothesis is really arbitrarily inaccurate, we can guess all holdout labels if we try out just two distinct hypotheses. If the entropy of the holdout data is one, then we need to try out two ....

T. Scheffer and R. Herbrich. Unbiased assessment of learning algorithms. In IJCAI-97, pages 798--803, 1997.

.... size grows (Kearns et al. 1997) Most practical learning algorithms incorporate some form of complexity penalization (like tree pruning, weight decay) and account for this problem by providing a parameter that can be adapted for each learning problem e.g. by cross validation (Ng, 1997; Scheffer Herbrich, 1997). Bayesian learners (Bayes, 1763; Berger, 1985) focus on the posterior probability of a function f having generated the sample S: P (f jS) Under certain ideal conditions, one can, under high computational effort, derive the Bayes hypothesis from P (f jS) which is guaranteed to have the least ....

Scheffer, T., & Herbrich, R. (1997). Unbiased assessment of learning algorithms. In IJCAI-97, pp.

No context found.

T. Sche#er and R. Herbrich. Unbiased assessment of learning algorithms. In IJCAI97, pages 798--803, 1997. 7

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC