Validation of input data for trained neural-nets

✍ Scribed by Georg Stimpfl-Abele

• Book ID: 103048814
• Publisher: Elsevier Science
• Year: 1995
• Tongue: English
• Weight: 803 KB
• Volume: 85
• Category: Article
• ISSN: 0010-4655
• DOI: 10.1016/0010-4655(94)00142-o

✦ Synopsis

The validation of input-data samples is a very important issue for neural-net applications if high precision is required. The aim is to make sure that a neural net properly trained on one data sample gives reliable results for another. Using the search for the standard Higgs boson at LEP-200 as an example, the effect of systematic changes in the input data is studied in detail for feed-forward nets. Several methods for recognizing significant differences between the training-data sample and a test-data sample are presented.