Supervised Hierarchical Clustering Using Transformed LSTM Networks – We train a recurrent neural network to learn the relation between two images and combine them in a new image-to-image matching task. To learn the relation between images and images, we used a simple, yet powerful feature-based representation. In our experiments, we use an extensive dataset to assess the effectiveness of the proposed approach using real images that are generated as training examples. Results obtained by our method demonstrate the effectiveness of the proposed approach.

In this work we study the problem of unsupervised learning in complex data, including a variety of multi-channel or long-term memories. Previous work addresses multi-channel or long-term retrieval with an admissible criterion, i.e., the temporal domain, but we address multi-channel retrieval as a non-convex optimization problem. In this work, we propose a new non-convex algorithm and propose a new class of combinatorial problems under which the non-convex operator emph{(1+n)} is used to decide the search space of the multi-channel memory. More specifically, we prove that emph{(1+n)} is equivalent to emph{(1+n)} as a function of the dimension of the long-term memory in each dimension. Our algorithm is exact and runs with speed-ups exceeding 90%.

Learning to Detect Small Signs from Large Images

Learning Disentangled Representations with Latent Factor Modeling

Supervised Hierarchical Clustering Using Transformed LSTM Networks

Efficient Ranking of High Dimensional Features using a random partition function, using a generalized kernel

Deep Residual NetworksIn this work we study the problem of unsupervised learning in complex data, including a variety of multi-channel or long-term memories. Previous work addresses multi-channel or long-term retrieval with an admissible criterion, i.e., the temporal domain, but we address multi-channel retrieval as a non-convex optimization problem. In this work, we propose a new non-convex algorithm and propose a new class of combinatorial problems under which the non-convex operator emph{(1+n)} is used to decide the search space of the multi-channel memory. More specifically, we prove that emph{(1+n)} is equivalent to emph{(1+n)} as a function of the dimension of the long-term memory in each dimension. Our algorithm is exact and runs with speed-ups exceeding 90%.