Variance Networks: When Expectation Does Not Meet Your Expectations

Authors: Kirill Neklyudov, Dmitry Molchanov, Arsenii Ashukha, Dmitry Vetrov

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that such layers can learn surprisingly well, can serve as an efficient exploration tool in reinforcement learning tasks and provide a decent defense against adversarial attacks. We also show that a number of conventional Bayesian neural networks naturally converge to such zero-mean posteriors. We observe that in these cases such zero-mean parameterization leads to a much better training objective than more flexible conventional parameterizations where the mean is being learned. 6 EXPERIMENTS We perform experimental evaluation of variance networks on classification and reinforcement learning problems.
Researcher Affiliation Collaboration Kirill Neklyudov Samsung-HSE Laboratory, National Research University Higher School of Economics Samsung AI Center Moscow k.necludov@gmail.com Dmitry Molchanov Samsung-HSE Laboratory, National Research University Higher School of Economics Samsung AI Center Moscow dmolch111@gmail.com Arsenii Ashukha Samsung AI Center Moscow ars.ashuha@gmail.com Dmitry Vetrov Samsung-HSE Laboratory, National Research University Higher School of Economics Samsung AI Center Moscow vetrovd@yandex.ru
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Experiments were implemented using Py Torch (Paszke et al. (2017)). The code is available at https://github.com/ da-molchanov/variance-networks.
Open Datasets Yes We consider three image classification tasks, the MNIST (Le Cun et al. (1998)), CIFAR-10 and CIFAR-100 (Krizhevsky & Hinton (2009)) datasets.
Dataset Splits No The paper mentions training and test sets but does not provide specific details on validation splits (percentages or sample counts), nor does it explicitly state the use of predefined standard validation splits for reproduction.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions 'Py Torch (Paszke et al. (2017))' and 'Caffe (2014)' but does not provide specific version numbers for the software dependencies needed for replication.
Experiment Setup Yes In all experiments the policy was approximated with a three layer fully-connected neural network containing 256 neurons on each hidden layer. Parameter noise and variance network policies had the second hidden layer to be parameter noise (Fortunato et al. (2017); Plappert et al. (2017)) and variance (Section 3) layer respectively. For both methods we made a gradient update for each episode with individual samples of noise per episode. Stochastic gradient learning is performed using Adam (Kingma & Ba (2014)).