Skip to content

Add backprop training support for GMM and KMeans #50

@MechaCritter

Description

@MechaCritter

Is your feature request related to a problem? Please describe.
Inspired by this paper:

P. Wieschollek, F. Groh, and H. P. A. Lensch. “Backpropagation Training for Fisher Vectors
within Neural Networks”. In: CoRR abs/1702.02549 (2017). arXiv: 1702.02549. url:
http://arxiv.org/abs/1702.02549.

I thought this would be an interesting experiment: the GaussianMixtureModel used used for Fisher Vector can actually be used as a neural network layer and hence trained via Backprop, which might improve the model's performance significantly. The same could be done for KMeans for VLAD.

Describe the solution you'd like
Under pyvisim/neural_networks, we can try to add a FisherNet or something similar, that uses this trick. If you also have time. in the demo notebooks, you can add a notebook demonstrating how to train a neural net and benchmark it against e.g. the OxfordFlowerDataset.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

Metadata

Metadata

Assignees

No one assigned

    Labels

    P3Priority 3contributor neededNeeds a contributor to take over the issue/feature

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions