|Topic:||Opravování pravděpodobnostních predikcích neuronových sítí|
|Supervisor:||Mgr. Oleksandr Shekhovtsov, Ph.D.|
|Announce as:||Diplomová práce, Semestrální projekt|
|Description:||Neural networks are widely applied in recognition problems. Commonly, they are trained on a typically very large training set by optimizing a probabilistic learning criterion, e.g. likelihood. The architecture and the learning scheme are tuned by researchers towards good generalization with respect to recognition accuracy. When the output of the trained classifier includes predictive probability, which is needed in a number of tasks, accurate values of predictive probabilities become necessary. However, the probabilistic predictions are typically not reliable and even misleading, e.g. overconfident w.r.t. the real error rate. It is thus required to correct the model when knowing the test task and test conditions and using a small (labeled) realistic data set. In the thesis:
Analyze the issue of misleading probabilistic predictions: in which scenarios it may be detrimental (change of the loss matrix, prior shift, domain shift, out-of-domain data)? We can possibly adjust the model or predictions at different stages: training, calibration, test. Discuss these options. Relate to the forecasting literature [5,6].
Assume that the network will be used for statistical decision making with a known loss matrix. Based on the results of the semestral project, experimentally compare, analyze and develop calibration methods for the known loss matrix and labeled data.
Assume that at the test time the distribution of classes is different from the training and is known or can be estimated by existing methods. If the model was calibrated, the Bayesian correction of the predictive distribution is expected to perform well. Calibrate the predictive distribution for this purpose by extending the techniques from 2).
Propose and test approaches in the following directions:
Improve calibration and correct for the prior shift (part 2) when the test class distribution is not known but we have an unlabeled data sample.
If the model is well-calibrated but there is a domain shift at test time (or even out-of-domain examples are presented to it) the probabilistic predictions can become misleading again. Can the domain shift adaptation and calibration be addressed in a unified framework?
Test the methods on realistic datasets (e.g., mushrooms, snakes, skin cancer , semantic segmentation of road scenes).
|Bibliography:|| Sipka T., Sulc M and Matas J. (2022) The Hitchhiker’s Guide to Prior-Shift Adaptation.
 Zhao S., et al. (2021) Calibrating predictions to decisions: A novel approach to multi-class calibration.
 Pampari A…. Ermon S. (2020) Unsupervised Calibration under Covariate Shift
 Sahoo R… Ermon S. (2021) Reliable Decisions with Threshold Calibration
 DeGroot and Fienberg (1982). Assessing probability assessors: calibration and refinement
 Murphy and Winkler (1987). A general framework for forecast verification.