Affiliations: CCRCB, Queen's University Belfast, Belfast, Northern Ireland, UK | Faculty of Mechanics and Mathematics, Moscow State University, Moscow, Russia | Department of Computer Science, King's College London, London, England, UK | T.E.I. of Western Macedonia, Kastoria, Greece
Note: [] Address for correspondence: Department of Computer Science, King's College London, London WC2R 2LS, England, UK
Abstract: The paper aims at tight upper bounds on the size of pattern classification circuits that can be used for a priori parameter settings in a machine learning context. The upper bounds relate the circuit size S(C) to nL := [log2mL], where mL is the number of training samples. In particular, we show that there exist unbounded fan-in threshold circuits with less than (a) SRcc := 2·√2nL + 3 gates for unbounded depth, (b) SLcc := 34.8 · √2nL + 14 · nL − 11 · log2nL + 2 gates for small bounded depth, where in both cases all mL samples are classified correctly. We note that the upper bounds do not depend on the length n of input (sample) vectors. Since nL << n in real-world problem settings, the upper bounds return values that are suitable for practical applications. We provide experimental evidence that the circuit size estimations work well on a number of pattern classification tasks. As a result, we formulate the conjecture that [1.25 · SRcc or [0.07 · SLcc] gates are sufficient to achieve a high generalization rate of bounded-depth classification circuits.
