| Topic: | Combining parametric and non-parametric classifiers for visual object recognition |
|---|
| Department: | Katedra kybernetiky |
|---|
| Supervisor: | doc. Georgios Tolias, Ph.D. |
|---|
| Announce as: | Diplomová práce, Bakalářská práce, Semestrální projekt |
|---|
| Description: | The impressive generalization performance of modern neural networks is attributed in part to their ability to implicitly memorize complex training patterns. Inspired by this, we will explore a mechanism to improve model generalization via explicit memorization. We will rely on the residual-memorization (ResMem) algorithm, a new method that augments an existing prediction model (e.g. a neural network) by fitting the model’s residuals with a k-nearest neighbor based regressor. The final prediction is then the sum of the original model and the fitted residual regressor. By construction, ResMem can explicitly memorize the training labels. Empirically, we show that ResMem consistently improves the test set generalization of the original prediction model across various standard vision and natural language processing benchmarks. The project will explore ways to improve ResMem and find the domains and tasks where it is the most effective. |
|---|
| Bibliography: | Yang et al, 2023, ResMem: Learn what you can and memorize the rest
Radenovic Tolias Chum, PAMI 2019, Fine-tuning CNN Image Retrieval with No Human Annotation
|
|---|
