Bayesian filtering unifies adaptive and non-adaptive neural network optimization methods

By Laurence Aitchison. 2018

23 citations recorded by the source corpus.

We formulate the problem of neural network optimization as Bayesian filtering, where the observations are the backpropagated gradients. While neural network optimization has previously been studied using natural gradient methods which are closely related to Bayesian inference, they were unable to recover standard optimizers such as Adam and RMSprop with a root-mean-square gradient normalizer, instead getting a mean-square normalizer. To recover the root-mean-square normalizer, we find it necessary to account for the temporal dynamics of all the other parameters as they are geing optimized. The resulting optimizer, AdaBayes, adaptively transitions between SGD-like and Adam-like behaviour, automatically recovers AdamW, a state of the art variant of Adam with decoupled weight decay, and has generalisation performance competitive with SGD.

Subfields: Artificial Intelligence, Ocean Engineering, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Chris J. Maddison
2. Jaehoon Lee
3. Christopher J. Shallue
4. Zhen Hu
5. Sankaran Mahadevan
6. Byeng Dong Youn
7. Dami Choi
8. George E. Dahl

In Ocean Engineering

1. Zhen Hu
2. Sankaran Mahadevan
3. Christopher J. Shallue
4. George E. Dahl
5. Jaehoon Lee
6. Frederik Benzing

In Optimization & Training

1. Christopher J. Shallue
2. Byeng Dong Youn
3. Chris J. Maddison
4. Dami Choi
5. George E. Dahl
6. Chao Hu
7. Jaehoon Lee
8. Frederik Benzing

Citations in the corpus

• On Empirical Comparisons of Optimizers for Deep Learning (2019, background): “Others do not even report the value of (cid:15) used (Balles and Hennig, 2017; Zhang and Mitliagkas, 2017; Keskar and Socher, 2017; Chen et al., 2018; Zhou et al., 2018; Aitchison, 2018; Reddi et al., 2019; Luo et al., 2019).”
• A comprehensive review of digital twin — part 1: modeling and twinning enabling technologies (2022, background): “It is not uncommon to see these recursive filtering alternatives succeed in training nonlinear models like neural networks with fixed optimal parameters (Puskorius and Feldkamp, 1994; Aitchison, 2020), with a possible extension to time-varying parameters (Wagner et al.”
• Unifying Importance Based Regularisation Methods for Continual Learning (2022, methodology): “Note that v t is a common approximation of the diagonal of the Hessian, e.g. [Graves, 2011, Khan et al., 2018, Aitchison, 2018], and recall that the Hessian is approximately equal to the Fisher (see Appendix C for a more detailed discussion).”