Reference no: EM133559683
Themed Project: Model Selection and Transferability Estimation in the Presence of the Distribution Shift
The problem of transferability estimation and model selection in deep learning iis an early problem in computer vision. The problem becomes more important in presence of the domain shift where the train and test data don't follow similar distributions. In the first part of this project, we will go throgh the current state of the art methods of model selection and transferability estimation techniques in computer vision applications. We consider two problems: First is choosing the deep learning model to use on an unseen dataset for which no labels are available. Specifically, given a set of deep learning models for object recognition and an unseen test dataset, we aim to identify which deep learning model performs best in the target dataset, without the need to evaluation labels on this dataset. Second problem is estimating the transferability of a deep learning model on a new dataset, for which the labels are available. Note that these two settings are different in a sense that: in the first one we have a set of models and a target dataset, and we want to select the optimum model for the given dataset, while in the second problem, we have a set of target datasets and a model for which we want to estimate the transferability on each of these datasets.
Task description
The final report reports on the results from the student's project. It should be written as to be understandable by persons other than the supervisor, and should comprehensively include material on the problems and goals of the project, applicable methods, the approach taken, major decisions and the reasons for the selection of goals and methods, results, the extent to which the goals have been achieved, the relevance, importance and context of achievements and the reasons for any shortcomings.
There is no set format other than it should have an executive summary and conclusion section. The format should be discussed with the project's supervisor. Students are reminded that they should follow the guidelines they have learnt in DATA7901 when writing the final report. The maximum length of the report is 40 pages (as long as necessary and as short as possible).
The report will marked according to the criteria below. The supervisor will assess the technical components of the report (Thesis definition & scope, Background, Approach & execution, Conclusion). The "Writing & presentation" component will be marked by Lisa Kelly.
Final Report
The final report reports on the results from the student's project. It should be written as to be understandable by persons other than the supervisor, and should comprehensively include material on the problems and goals of the project, applicable methods, the approach taken, major decisions and the reasons for the selection of goals and methods, results, the extent to which the goals have been achieved, the relevance, importance and context of achievements and the reasons for any shortcomings. There is no set format other than it should have an executive summary and conclusion section. The format should be discussed with the project's supervisor.