On the saddle point problem for non-convex optimization
Comments 11 pages, 8 figures
作者
Yoshua Bengio
Machine Learning
Deep Generative Stochastic Networks Trainable by Backprop
Comments arXiv admin note: text overlap with arXiv:1305.0445, Also published in ICML'2014
Bounding the Test Log-Likelihood of Generative Models
Comments 10 pages, 1 figure, 2 tables. International Conference on Learning Representations (ICLR'2014, conference track)
How to Construct Deep Recurrent Neural Networks
Comments Accepted at ICLR 2014 (Conference Track). 10-page text + 3-page references
Representation Learning: A Review and New Perspectives
On the number of response regions of deep feed forward networks with piece-wise linear activations
Comments 17 pages, 9 figures
An empirical analysis of dropout in piecewise linear networks
Comments Extensive updates; 8 pages plus acknowledgements/references
Generalized Denoising Auto-Encoders as Generative Models
Maxout Networks
Comments This is the version of the paper that appears in ICML 2013
Journal ref JMLR WCP 28 (3): 1319-1327, 2013
Pylearn2: a machine learning research library
Comments 9 pages
Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
Comments arXiv admin note: substantial text overlap with arXiv:1305.2982
Knowledge Matters: Importance of Prior Information for Optimization
Comments 37 Pages, 5 figures, 5 tables JMLR Special Topics on Representation Learning Submission
Challenges in Representation Learning: A report on three machine learning contests
Comments 8 pages, 2 figures
Deep Learning of Representations: Looking Forward
Estimating or Propagating Gradients Through Stochastic Neurons
Joint Training Deep Boltzmann Machines for Classification
Comments Major revision with new techniques and experiments. This version includes new material put on the poster for the ICLR workshop
A Semantic Matching Energy Function for Learning with Multi-relational Data
Metric-Free Natural Gradient for Joint-Training of Boltzmann Machines
Big Neural Networks Waste Capacity
On the difficulty of training Recurrent Neural Networks
Comments Improved description of the exploding gradient problem and description and analysis of the vanishing gradient problem
Advances in Optimizing Recurrent Networks
Joint Training of Deep Boltzmann Machines
Comments 4 pages
High-dimensional sequence transduction
Evolving Culture vs Local Minima
Texture Modeling with Convolutional Spike-and-Slab RBMs and Deep Extensions
Theano: new features and speed improvements
Comments Presented at the Deep Learning Workshop, NIPS 2012
