Choose timezone
Your profile timezone:
pyBAMBI developer meeting
→
Europe/Zurich
We discussed the major tasks required for pyBAMBI:
1) (Top priority): Implement the structure from the BAMBI paper in python, using the dumper function to train the neural network. We need to think about the persistence of the keras model (should it be a global object? Should the dumper function be moved?), and whether to train it from scratch with randomised initial weights/biases, or start from the previous values (BAMBI did the latter). Roberto and Joaquin will work on this in January with Will, with the aim of getting some basic progress by the start of February when Martin (plus Pat) will be in Cambridge for coding.
2) We need to define some interesting likelihood functions for testing. Csaba and Melissa will liase on this.
3) We need to try and define a default neural network architecture, loss function, training scheme and activation function, that works reasonably well on a variety of likelihood functions. This can be tested in parallel with (1) using an external script which Martin can supply. It relies on samples from (2) being available. Csaba, Melissa, Martin, Joaquin and Bob can work on this.
4) Bob has previous experience with Bayesian neural nets + related uncertainty measures, and can test these in an external code once (2) is finished.
There are minutes attached to this event.
Show them.
