Determining Selectivities in the Neocognitron

In Fukushima's neocognitron, the S-cell selectivity parameters have a profound eeect on the classiication performance of the network. To date, no one has presented a successful approach to determining these parameters so as to maximize the neocognitron's correct recognition rate. This paper presents two new algorithms for adjusting S-cell selectivities. The rst method | Sub-Optimal Feature Training | adopts a similar philosophy to Hildebrandt's closed-form training algorithm, but avoids the problem of training feature rejection experienced by Hildebrandt's algorithm. The second method | Selectivity Hunting to Optimize Performance | is reminiscent of a two-factor experiment and makes use of real world data to estimate a good choice of second and third layer selectivities. The latter method is the most eeective of the two algorithms, achieving a correct classiication rate of around 75% on ZIP code digits.