Dynamically-Scaled Deep Canonical Correlation Analysis

Canonical Correlation Analysis (CCA) is a method for feature extraction of two views by finding maximally correlated linear projections them. Several variants CCA have been introduced in the literature, particular, based on deep neural networks learning highly nonlinear transformations views. As these models are parameterized conventionally, their learnable parameters remain independent inputs after training process, which limits capacity representations. We introduce novel dynamic scaling an input-dependent canonical correlation model. In our deep-CCA models, last layer scaled second network that conditioned model’s input, resulting parameterization dependent input samples. evaluate model multiple datasets and demonstrate learned representations more comparison to conventionally-parameterized CCA-based also obtain preferable retrieval results.