IPAB Workshop - 21/5/26 Title: Some Properties of High Dimensional Spaces and How They Affect Classifiers Robert B. FisherAbstract: Many classifiers, including deep nets, are based on feature vectors that lie in a high-dimensional space, eg. D=512, 1000, 2000, etc.Classifications are often based on distances, eg. from a boundary, a prototype, or a nearest neighbors. Unfortunately, distances in high dimensions are somewhat unintuitive, in part because of the law of large numbers. For example, for large enough dimension D, any two randomly chosen points in [0,1]^D will be approximately an identical distancesqrt(D/6) apart. This talk will present some observations on geometry and classifiers in high dimensions. It will also present an alternative to one-hot encoding that could improve deep-net based classifiers. May 21 2026 13.00 - 14.00 IPAB Workshop - 21/5/26 Bob Fisher MF2 This article was published on Monday 13 July 2026
IPAB Workshop - 21/5/26 Title: Some Properties of High Dimensional Spaces and How They Affect Classifiers Robert B. FisherAbstract: Many classifiers, including deep nets, are based on feature vectors that lie in a high-dimensional space, eg. D=512, 1000, 2000, etc.Classifications are often based on distances, eg. from a boundary, a prototype, or a nearest neighbors. Unfortunately, distances in high dimensions are somewhat unintuitive, in part because of the law of large numbers. For example, for large enough dimension D, any two randomly chosen points in [0,1]^D will be approximately an identical distancesqrt(D/6) apart. This talk will present some observations on geometry and classifiers in high dimensions. It will also present an alternative to one-hot encoding that could improve deep-net based classifiers. May 21 2026 13.00 - 14.00 IPAB Workshop - 21/5/26 Bob Fisher MF2 This article was published on Monday 13 July 2026
