Bayesian Learning of Neural Network Architectures
This work addresses the challenge of optimizing neural network architectures efficiently for researchers and practitioners, though it is incremental as it builds on existing variational learning techniques.
The paper tackles the problem of estimating neural network architectural parameters like layer size and depth using a Bayesian method, resulting in improved generalization on small datasets and increased robustness to parameter initialization.
In this paper we propose a Bayesian method for estimating architectural parameters of neural networks, namely layer size and network depth. We do this by learning concrete distributions over these parameters. Our results show that regular networks with a learnt structure can generalise better on small datasets, while fully stochastic networks can be more robust to parameter initialisation. The proposed method relies on standard neural variational learning and, unlike randomised architecture search, does not require a retraining of the model, thus keeping the computational overhead at minimum.