Research‎ > ‎

Non-Linear Embedding

Large-scale Non-linear Multimodal Semantic Embedding

This project aims to investigate efficient and effective methods that take as input visual data with other associated data modalities from a multimodal collection and produce a common semantic representation that models the relationships between these different modalities. Classical latent semantic models can boost the performance in tasks like content-based retrieval, exploration and automatic annotation and classification, among others, by modeling relations between the different modalities. Nevertheless, one important limitation of most approaches based on latent semantic analysis is that they assume linear dependencies between the data modalities. However, it is reasonable to expect that there is an inherent non-linearity in the nature of the data. Therefore, this kind of linear models can only give a limited and restricted approximation of the nature of the data and the complex relationships between the different modalities.