opac, catalogo, online,servizi, biblioteca, sebina

On variants of online linear-threshold algorithms for classification : dottorato di ricerca in informatica : tesi di dottorato

Cavallanti, Giovanni <1979- >

Tesi o dissertazioni - 2009

Online linear-threshold algorithms for classification are easy toimplement, are efficient to run, and usually lead to simple analyses.However, as effective as they are, existing online linear-thresholdalgorithms have significant drawbacks. Here, we introduce and empiricallyevaluate various approaches to overcoming these limitations. Our newalgorithm operates within the framework of the selective sampling model inwhich labels are disclosed on demand. We state an upper bound on bothcumulative regret and number of stored examples. We show that ouralgorithm achieves state-of-the-art performance on real-world problems.We study the problem of simultaneously learning K different tasks and showwe can effectively capitalize on task relatedness. Theoretical guaranteesfor different notions of task relatedness and different multitask learningprotocols are provided. In relation to this problem, we discuss theproblem of learning from multiview examples where, at each time step, weare given a set of K encodings of the same underlying instance [...]
  • Lo trovi in
  • Scheda
  • Unimarc
USMS7@Archivio Tesi di dottorato

Archivio Tesi di dottorato

Documento per sola consultazione interna

D. 2009 0240
Autorizzazione: si.