David Scott Krueger

Bayesian Hypernetworks

Chin-Wei Huang

Riashat Islam

Ryan Turner

Alexandre Lacoste

2017-10-13

ArXiv (prépublication)

Bayesian Hypernetworks

Chin-Wei Huang

Riashat Islam

Ryan Turner

Alexandre Lacoste

We propose Bayesian hypernetworks: a framework for approximate Bayesian inference in neural networks. A Bayesian hypernetwork, h, is a neura… (voir plus)l network which learns to transform a simple noise distribution, p(e) = N(0,I), to a distribution q(t) := q(h(e)) over the parameters t of another neural network (the ``primary network). We train q with variational inference, using an invertible h to enable efficient estimation of the variational lower bound on the posterior p(t | D) via sampling. In contrast to most methods for Bayesian deep learning, Bayesian hypernets can represent a complex multimodal approximate posterior with correlations between parameters, while enabling cheap iid sampling of q(t). In practice, Bayesian hypernets provide a better defense against adversarial examples than dropout, and also exhibit competitive performance on a suite of tasks which evaluate model uncertainty, including regularization, active learning, and anomaly detection.

2017-10-13

ArXiv (prépublication)

Bayesian Hypernetworks

Chin-Wei Huang

Riashat Islam

Ryan Turner

Alexandre Lacoste

2017-10-13

ArXiv (prépublication)

A Closer Look at Memorization in Deep Networks

Devansh Arpit

Stanisław Jastrzębski

Nicolas Ballas

Maxinder S. Kanwal

Asja Fischer

We examine the role of memorization in deep learning, drawing connections to capacity, generalization, and adversarial robustness. While dee… (voir plus)p networks are capable of memorizing noise data, our results suggest that they tend to prioritize learning simple patterns first. In our experiments, we expose qualitative differences in gradient-based optimization of deep neural networks (DNNs) on noise vs. real data. We also demonstrate that for appropriately tuned explicit regularization (e.g., dropout) we can degrade DNN training performance on noise datasets without compromising generalization on real data. Our analysis suggests that the notions of effective capacity which are dataset independent are unlikely to explain the generalization performance of deep networks when trained with gradient based methods because training data itself plays an important role in determining the degree of memorization.

2017-07-17

Proceedings of the 34th International Conference on Machine Learning (publié)

proceedings.mlr.press

A Closer Look at Memorization in Deep Networks

Devansh Arpit

Stanisław Jastrzębski

Nicolas Ballas

Maxinder S. Kanwal

Asja Fischer

We examine the role of memorization in deep learning, drawing connections to capacity, generalization, and adversarial robustness. While dee… (voir plus)p networks are capable of memorizing noise data, our results suggest that they tend to prioritize learning simple patterns first. In our experiments, we expose qualitative differences in gradient-based optimization of deep neural networks (DNNs) on noise vs. real data. We also demonstrate that for appropriately tuned explicit regularization (e.g., dropout) we can degrade DNN training performance on noise datasets without compromising generalization on real data. Our analysis suggests that the notions of effective capacity which are dataset independent are unlikely to explain the generalization performance of deep networks when trained with gradient based methods because training data itself plays an important role in determining the degree of memorization.

2017-06-16

ArXiv (prépublication)

Deep Nets Don't Learn via Memorization

Nicolas Ballas

Stanisław Jastrzębski

Devansh Arpit

Maxinder S. Kanwal

Tegan Maharaj

Emmanuel Bengio

Asja Fischer

We use empirical methods to argue that deep neural networks (DNNs) do not achieve their performance by memorizing training data in spite of … (voir plus)overlyexpressive model architectures. Instead, they learn a simple available hypothesis that fits the finite data samples. In support of this view, we establish that there are qualitative differences when learning noise vs. natural datasets, showing: (1) more capacity is needed to fit noise, (2) time to convergence is longer for random labels, but shorter for random inputs, and (3) that DNNs trained on real data examples learn simpler functions than when trained with noise data, as measured by the sharpness of the loss function at convergence. Finally, we demonstrate that for appropriately tuned explicit regularization, e.g. dropout, we can degrade DNN training performance on noise datasets without compromising generalization on real data.

2017-02-17

International Conference on Learning Representations (publié)

dblp.uni-trier.de

Facilitating Multimodality in Normalizing Flows

Chin-Wei Huang

The true Bayesian posterior of a model such as a neural network may be highly multimodal. In principle, normalizing flows can represent such… (voir plus) a distribution via compositions of invertible transformations of random noise. In practice, however, existing normalizing flows may fail to capture most of the modes of a distribution. We argue that the conditionally affine structure of the transformations used in [Dinh et al., 2014, 2016, Kingma et al., 2016] is inefficient, and show that flows which instead use (conditional) invertible non-linear transformations naturally enable multimodality in their output distributions. With just two layers of our proposed deep sigmoidal flow, we are able to model complicated 2d energy functions with much higher fidelity than six layers of deep affine flows.

Zoneout: Regularizing RNNs by Randomly Preserving Hidden Activations

Tegan Maharaj

J'anos Kram'ar

Mohammad Pezeshki

Nicolas Ballas

Nan Rosemary Ke

Anirudh Goyal

We propose zoneout, a novel method for regularizing RNNs. At each timestep, zoneout stochastically forces some hidden units to maintain thei… (voir plus)r previous values. Like dropout, zoneout uses random noise to train a pseudo-ensemble, improving generalization. But by preserving instead of dropping hidden units, gradient information and state information are more readily propagated through time, as in feedforward stochastic depth networks. We perform an empirical investigation of various RNN regularizers, and find that zoneout gives significant performance improvements across tasks. We achieve competitive results with relatively simple models in character- and word-level language modelling on the Penn Treebank and Text8 datasets, and combining with recurrent batch normalization yields state-of-the-art results on permuted sequential MNIST.

2016-06-03

ArXiv (prépublication)