Locally-connected Hierarchical Neural Networks for GPU-accelerated Object Recognition

Convolutional neural networks have achieved good recognition results on image datasets while being computationally efficient, i.e., scaling well with the number of training patterns and the resolution of the patterns. Here we investigate a neu- ral network model that has a similar hierarchical structure, but does not employ weight sharing. Instead, each neuron has a fixed receptive field with unique con- nection weights. To deal with the enormous number of weights resulting from this architecture, we implemented a parallel version of the model using Nvidia's CUDA framework. This implementation is up to 82 times faster than a serial CPU implementation. Our model achieves state-of-the-art recognition performance on the NORB normalized-uniform dataset (2.87% error rate) and good results on the MNIST dataset (0.76% error rate). This suggests that large networks with local, non-shared connections might be an interesting architecture for object recognition tasks. To further evaluate the model, we created a large, publicly available train- ing and testing set, which consists of objects extracted from the LabelMe natural image dataset.