Online Learning of Invariant Object Recognition in a Hierarchical Neural Network

We propose the Temporal Correlation Net (TCN) as an object recognition system implementing three basic principles: forming temporal groups of features, learning in a hierarchical structure and using feedback for predicting future input. It is an improvement of the Temporal Correlation Graph and shows improved performance on standard datasets like ETH80 and COIL100. In contrast to its predecessor it can be trained online on all levels at once instead of having to be trained in batch mode level per level. Training images are presented in a temporal order showing objects undergoing specific transformations in viewing conditions the system is supposed to become invariant to. Computation time and memory demands are low because of a sparse connectivity structure resulting from the learning algorithm and efficient handling of neural activities.