Aggregate of HMAXs for image classification
Lau Kean Hong, Tay Yong Haur, Lo Fook Loong · 2016
Artificial neural networks (ANN) with deep learning using convolutional neural networks have recently achieved good results in various challenging problems. The HMAX is yet another deep architecture that could offer similar performance but with less training cycles required. In this paper, we extended the performance of HMAX by aggregating several HMAX networks together to achieve state-of-the-art results on Caltech-101 and Caltech-256 (excluding transfer learning methods). Our experiments showed that apart from having good performing individual networks, it is also critical to combine subsystems with a diverse set of parameter variations to achieve good aggregated output accuracy. We discovered that changes in image resolution and aspect ratio, both affecting the first layer of a multilayer neural network, produced the best combinations for good aggregation.