Deep Generative Models for 3D Point Clouds – With large object tracking systems, there is a growing interest in the learning of object tracking systems to be trained with hand-crafted object predictions. In this paper, we propose an online learning algorithm to automatically find the most probable object, based on the estimated performance of the predicted object. A common training approach is the targeted feature learning, where the target is the object of interest and the training data is trained from pre-trained image pairs. We evaluate our algorithm in both online and hand-crafted tasks and propose a new state-of-the-art prediction algorithm to address each of the performance trade-offs. We demonstrate the benefit of our algorithm on various datasets from the UCI 3D Object Tracking Challenge and illustrate that our algorithm outperforms state-of-the-art object prediction algorithms.

We propose a new method of learning a representation from music. It is based on a notion of the melody and the rhythm, which provides a direct interpretation of the music. We formulate the algorithm as a neural-network learning and we prove the relevance for a music classification task. We show that we can learn the melody of a song by learning the rhythm of the song as the melody of the song. We describe the algorithm and the experiments it demonstrates on a challenging music classification task.

Efficient Parallel Training for Deep Neural Networks with Simultaneous Optimization of Latent Embeddings and Tasks

Generative model of 2D-array homography based on autoencoder in fMRI

Deep Generative Models for 3D Point Clouds

A hybrid algorithm for learning sparse and linear discriminant sequences

A comparative study of different types of recurrent neural networks for music classificationWe propose a new method of learning a representation from music. It is based on a notion of the melody and the rhythm, which provides a direct interpretation of the music. We formulate the algorithm as a neural-network learning and we prove the relevance for a music classification task. We show that we can learn the melody of a song by learning the rhythm of the song as the melody of the song. We describe the algorithm and the experiments it demonstrates on a challenging music classification task.