ICNN: INPUT-CONDITIONED FEATURE REPRESENTATION LEARNING FOR TRANSFORMATION-INVARIANT NEURAL NETWORK

We propose a novel framework, ICNN, which combines the input-conditioned filter generation module and a decoder based network to incorporate contextual information present in images into Convolutional Neural Networks (CNNs). In contrast to traditional CNNs, we do not employ the same set of learned convolution filters for all input image instances. And our proposed decoder network serves the purpose of reducing the transformation present in the input image by learning to construct a representative image of the input image class. Our proposed joint supervision of input-aware framework when combined with techniques inspired by Multi-instance learning and max-pooling, results in a transformation-invariant neural network. We investigated the performance of our proposed framework on three MNIST variations, which covers both rotation and scaling variance, and achieved 0.98% error on MNIST-rot-12k, 1.12% error on Half-rotated MNIST and 0.68% error on Scaling MNIST, which is significantly better than the state-of-the-art results. Our proposed model also showcased consistent improvement on the CIFAR dataset. We make use of visualization to further prove the effectiveness of our input-aware convolution filters. Our proposed convolution filter generation framework can also serve as a plugin for any CNN based architecture and enhance its modeling capacity.