Master 2 Informatique, Sorbonne Université
Week 1 : Bayesian Models (January, 12th 2022)
Week 2 : Bayesian Neural Networks (January, 26th 2022)
Week 3 : Bayesian Deep Learning and Robustness (February, 2nd 2022)
The deadline for sending the reports is February, 14th 2022 (12:00am)
The reports on practical session should be sent to: nicolas.thome@lecnam.net - charles.corbiere@valeo.com - remy.sun@isir.upmc.fr (please only send to the supervisor in your group)
Please send a single archive file containing:
- A report answering the questions below with experimental results and their analyses. PDF format requested.
- The source codes (.pynb) of your work
Here are the specific elements to be developed in the reports:
- week 1
- Predictive distribution with Bayesian linear regression => [Question 1.3]
- Results of the distribution predictive distribution on the synthetic dataset
- Theoretical analysis to explain the form of the distribution (simplified case \alpha=0, \beta=1)
- Comment and discuss dataset with a "hole" (bonus question)
- week 2
- Part I.3 "Variational inference" : comment the class LinearVariational
- Results and analysis of VI on the non-linear dataset (moons)
- Results of MC on the non-linear dataset, difference with VI [Question 2.1]
- week 3
- Comment results for investigating most uncertain samples [Question 1.1]
- Explain the goal of failure prediction
- Why evaluating performances with AUPR and not AUC ? [Question 2.1]
- Comments on MCP, MCDropout and ConfidNet performances