Poziv na predavanje na engleskom jeziku Computationally feasible greedy algorithms for neural nets

Odjel za matematiku Sveučilišta Josipa Jurja Strossmayera u Osijeku organizira predavanje Computationally feasible greedy algorithms for neural nets, na engleskom jeziku

---

NAJAVA – Predavanje će se održati u srijedu, 16. ožujka 2016. godine s početkom u 12 sati u predavaonici broj 2 na Odjelu za matematiku Sveučilišta Josipa Jurja Strossmayera u Osijeku (Trg Ljudevita Gaja 6).

Predavanje će održati:

• prof. Andrew Barron, Yale University, Department of Statistics, USA

[br] S obzirom da je riječ o predavanju na engleskom jeziku, na tom jeziku objavljujemo i kratak opis predavanja:

Previously, greedy algorithms have been shown to have good approximation and estimation properties for superpositions of a sigmoidal function or other ridge activation functions. In each step the parameters of a new sigmoid are fit to the residuals of the previous sigmoids. However, it remains a challenging task to produce guarantees for the parameter search for each sigmoid. Here we discuss developments of two algorithms for this task. One is an implementation of adaptive annealing, in which internal modifi- cations of the sigmoid is made, allowing stability of the adaptive annealing solution. The other algorithm is a convergent nonlinear modification of the tensor methods of Anandkumar and her colleagues which allows optimization of the inner product of the residuals with certain sigmoids and related activation functions. This work is joint with Jason Klusowski.