New Workingpaper: "The Role of #Hyperparameters in #MachineLearning Models and How to Tune Them". We suggest: Handle HPs with the same loving care as parameter estimates---you could end up choosing the wrong model. https://tinyurl.com/mr2akrn3

'Beyond the Golden Ratio for Variational Inequality Algorithms', by Ahmet Alacaoglu, Axel Böhm, Yura Malitsky.

http://jmlr.org/papers/v24/22-1488.html

#ascent #constrained #hyperparameters

Computationally-efficient initialisation of GPs: The generalised variogram method

Felipe Tobar, Elsa Cazelles, Taco de Wolff

Action editor: Cédric Archambeau.

https://openreview.net/forum?id=slsAQHpS7n

#gps #geostatistics #hyperparameters

'Prior Specification for Bayesian Matrix Factorization via Prior Predictive Matching', by Eliezer de Souza da Silva, Tomasz Kuśmierczyk, Marcelo Hartmann, Arto Klami.

http://jmlr.org/papers/v24/21-0623.html

#factorization #hyperparameters #priors

Computationally-efficient initialisation of GPs: The generalised variogram method

https://openreview.net/forum?id=slsAQHpS7n

#gps #geostatistics #hyperparameters

No More Pesky Hyperparameters: Offline Hyperparameter Tuning for RL

Han Wang, Archit Sakhadeo, Adam M White et al.

https://openreview.net/forum?id=AiOUi3440V

#hyperparameters #hyperparameter #learns

What are you using to tune your #hyperparameters? As #AutoML researchers, it is very important for us to understand the needs and expectations of ML researchers, engineers and data scientists. Help us and yourself by being part of the following survey https://www.soscisurvey.de/hpo-method-validation/

Is there a #julialang equivalent of https://github.com/google/gin-config ?

I found using .gin files a really simple but useful way to store #hyperparameters during #deeplearning

If you do #machinelearning with #python and haven't heard of it, check it out!

The first one is the dual #benchmark - comparing all models both default and tuned #hyperparameters.
Sure, it doesn't make much difference for production deployment of the model, but good defaults are very convenient during #EDA and early experiments