2015

Variance Reduction in SGD by Distributed Importance Sampling

Alain, Guillaume, Lamb, Alex, Sankar, Chinnadhurai et al.

Understand

Humans are able to accelerate their learning by selecting training materials that are the most informative and at the appropriate level of difficulty.

  • We propose a framework for distributing deep learning in which one set of workers search for the most informative examples in parallel while a single worker updates the model on examples selected by importance sampling.
  • This leads the model to update using an unbiased estimate of the gradient which also has minimum variance when the sampling proposal is proportional to the L2-norm of the gradient.
  • We show experimentally that this method reduces gradient variance even in a context where the cost of synchronization across machines cannot be ignored, and where the factors for importance sampling are not updated instantly across the training set.

Built on

  • Importance sampling: a review

    Tokdar, Surya T. and Kass, Robert E · 1939

    Earlier work this paper cites.

  • A neural probabilistic language model

    Bengio, Y., Ducharme, R., Vincent, P., and Jauvin, C · 2003

    Earlier work this paper cites.

  • Curriculum learning

    Bengio, Yoshua, Louradour, Jerome, Collobert, Ronan, and Weston, Jason · 2009

    Earlier work this paper cites.

  • Theano: a CPU and GPU math expression compiler

    Bergstra, James, Breuleux, Olivier, Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Desjardins, Guillaume, Turian, Joseph, Warde-Farley, David, and Bengio, Yoshua · 2010

    Earlier work this paper cites.

  • Distributed delayed stochastic optimization

    Agarwal, Alekh and Duchi, John C · 2011

    Earlier work this paper cites.

  • Reading digits in natural images with unsupervised feature learning

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and Ng, A. Y · 2011

    Earlier work this paper cites.

Similar

  • Hogwild: A lock-free approach to parallelizing stochastic gradient descent

    Recht, Benjamin, Re, Christopher, Wright, Stephen, and Niu, Feng · 2011

    Cited alongside, same era.

  • Theano: new features and speed improvements

    Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Bergstra, James, Goodfellow, Ian J., Bergeron, Arnaud, Bouchard, Nicolas, and Bengio, Yoshua · 2012

    Cited alongside, same era.

  • Large scale distributed deep networks

    Dean, J., Corrado, G.S, Monga, R., Chen, K., Devin, M., Le, Q.V., Mao, M.Z., Ranzato, M.A., Senior, A., Tucker, P., Yang, K., and Ng, A. Y · 2012

    Cited alongside, same era.

  • Low-Rank Approximations for Conditional Feedforward Computation in Deep Neural Networks

    Davis, A. and Arel, I · 2013

    Cited alongside, same era.

  • Maxout networks

    Original

    Goodfellow, Ian J., Warde-Farley, David, Mirza, Mehdi, Courville, Aaron, and Bengio, Yoshua · 2013

    Cited alongside, same era.

Then

  • Compete to compute

    Srivastava, Rupesh K, Masci, Jonathan, Kazerounian, Sohrob, Gomez, Faustino, and Schmidhuber, Jürgen · 2013

    Later among the works it cites.

  • Stochastic Optimization with Importance Sampling

    Zhao, P. and Zhang, T · 2014

    Later among the works it cites.

  • Accelerating stochastic gradient descent via online learning to sample

    Original

    Bouchard, Guillaume, Trouillon, Théo, Perez, Julien, and Gaidon, Adrien · 2015

    Closest in time.

  • Efficient Per-Example Gradient Computations

    Goodfellow, I · 2015

    Closest in time.

  • Asynchronous parallel stochastic gradient for nonconvex optimization

    Lian, Xiangru, Huang, Yijun, Li, Yuncheng, and Liu, Ji · 2015

    Closest in time.

Beyond the bibliography

alphaXiv searches the wider corpus for related work and actual follow-ups.

Open on alphaXiv

alphaXiv is searching for related work…