RW
R.T. Wiersma
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
We present a new approach for deep learning on surfaces, combining geometric convolutional networks with rotationally equivariant networks. Existing work either learns rotationally invariant filters, or learns filters in the tangent plane without correctly relating orientations between different tangent planes (orientation ambiguity). We propose a solution to both problems by applying Harmonic Networks on surfaces in the tangent plane: Harmonic Surface Networks (HSN).
Harmonic Networks constrain their filters to circular harmonics, which output complexvalued, rotatable feature maps. Considering these complex features as vectors inside the tangent plane, we can use parallel transport along shortest geodesics to transport them along the surface in a natural way. Additionally, Harmonic Networks can be configured so that the output is rotationally invariant, while containing rotationally equivariant filters in hidden layers. This property solves the orientation ambiguity problem, while learning directional filters. We evaluate HSN on three different problems: classification on Rotated MNIST in a plane and mapped to a sphere, correspondence on FAUST, and shape segmentation on FAUST. The results suggest that HSN could improve on state of the art approaches. ...
Harmonic Networks constrain their filters to circular harmonics, which output complexvalued, rotatable feature maps. Considering these complex features as vectors inside the tangent plane, we can use parallel transport along shortest geodesics to transport them along the surface in a natural way. Additionally, Harmonic Networks can be configured so that the output is rotationally invariant, while containing rotationally equivariant filters in hidden layers. This property solves the orientation ambiguity problem, while learning directional filters. We evaluate HSN on three different problems: classification on Rotated MNIST in a plane and mapped to a sphere, correspondence on FAUST, and shape segmentation on FAUST. The results suggest that HSN could improve on state of the art approaches. ...
We present a new approach for deep learning on surfaces, combining geometric convolutional networks with rotationally equivariant networks. Existing work either learns rotationally invariant filters, or learns filters in the tangent plane without correctly relating orientations between different tangent planes (orientation ambiguity). We propose a solution to both problems by applying Harmonic Networks on surfaces in the tangent plane: Harmonic Surface Networks (HSN).
Harmonic Networks constrain their filters to circular harmonics, which output complexvalued, rotatable feature maps. Considering these complex features as vectors inside the tangent plane, we can use parallel transport along shortest geodesics to transport them along the surface in a natural way. Additionally, Harmonic Networks can be configured so that the output is rotationally invariant, while containing rotationally equivariant filters in hidden layers. This property solves the orientation ambiguity problem, while learning directional filters. We evaluate HSN on three different problems: classification on Rotated MNIST in a plane and mapped to a sphere, correspondence on FAUST, and shape segmentation on FAUST. The results suggest that HSN could improve on state of the art approaches.
Harmonic Networks constrain their filters to circular harmonics, which output complexvalued, rotatable feature maps. Considering these complex features as vectors inside the tangent plane, we can use parallel transport along shortest geodesics to transport them along the surface in a natural way. Additionally, Harmonic Networks can be configured so that the output is rotationally invariant, while containing rotationally equivariant filters in hidden layers. This property solves the orientation ambiguity problem, while learning directional filters. We evaluate HSN on three different problems: classification on Rotated MNIST in a plane and mapped to a sphere, correspondence on FAUST, and shape segmentation on FAUST. The results suggest that HSN could improve on state of the art approaches.
Bachelor thesis
(2017)
-
Ruben Wiersma, Hung Nguyen, Alexander Geenen, Cynthia Liem, Otto Visser, S.M. Smulders
As GeoPhy is developing its business model and looking into the future of automated valu- ation models (AVM), this project delivers a proof of concept of a system that automates the training, maintaining, and delivery of machine learning models for automated valuations. In order to achieve this goal, the situation and problem were first analysed. This resulted in an outline of the desired product and requirements in the form of a MoSCoW analysis. An important goal for this project was to incorporate streams of data from a stream processing platform (Apache Kafka) into a service that would train and update models automatically. The second goal for this project was to keep track of the changes in the data in order to detect significant changes in distribution (concept drift) of the target prediction value.
These subjects were studied in literature, reviewing existing and upcoming valuation prac- tices in real-estate, steps needed to perform machine learning tasks, architecture to support big data processing, and concept drift. This resulted in a design made up of four different components: An ETL and data processing component, a modelling component, a Kafka con- nector, and a client-facing API. An important part to ensure efficiency and scalability of the system is the implementation of concept drift: models are only retrained when the distribu- tion of the target training value has changed significantly.
These components use storage in the form of a Postgres database, disk storage and Elastic Search logs. The logs (on model performance and concept drift usage) can be interpreted through a Grafana dashboard, which is editable through its own GUI.
Finally, to test the success of the project, a testing plan was set up and the code was reviewed by an external group (SIG). The code achieved all the testing milestones and received a 4.5/5 in a mid-development review on maintainability. With this project, the concept of automated valuation models inside GeoPhy’s new architecture has been tested and proved and the project is ready to be further developed and used in practice. ...
These subjects were studied in literature, reviewing existing and upcoming valuation prac- tices in real-estate, steps needed to perform machine learning tasks, architecture to support big data processing, and concept drift. This resulted in a design made up of four different components: An ETL and data processing component, a modelling component, a Kafka con- nector, and a client-facing API. An important part to ensure efficiency and scalability of the system is the implementation of concept drift: models are only retrained when the distribu- tion of the target training value has changed significantly.
These components use storage in the form of a Postgres database, disk storage and Elastic Search logs. The logs (on model performance and concept drift usage) can be interpreted through a Grafana dashboard, which is editable through its own GUI.
Finally, to test the success of the project, a testing plan was set up and the code was reviewed by an external group (SIG). The code achieved all the testing milestones and received a 4.5/5 in a mid-development review on maintainability. With this project, the concept of automated valuation models inside GeoPhy’s new architecture has been tested and proved and the project is ready to be further developed and used in practice. ...
As GeoPhy is developing its business model and looking into the future of automated valu- ation models (AVM), this project delivers a proof of concept of a system that automates the training, maintaining, and delivery of machine learning models for automated valuations. In order to achieve this goal, the situation and problem were first analysed. This resulted in an outline of the desired product and requirements in the form of a MoSCoW analysis. An important goal for this project was to incorporate streams of data from a stream processing platform (Apache Kafka) into a service that would train and update models automatically. The second goal for this project was to keep track of the changes in the data in order to detect significant changes in distribution (concept drift) of the target prediction value.
These subjects were studied in literature, reviewing existing and upcoming valuation prac- tices in real-estate, steps needed to perform machine learning tasks, architecture to support big data processing, and concept drift. This resulted in a design made up of four different components: An ETL and data processing component, a modelling component, a Kafka con- nector, and a client-facing API. An important part to ensure efficiency and scalability of the system is the implementation of concept drift: models are only retrained when the distribu- tion of the target training value has changed significantly.
These components use storage in the form of a Postgres database, disk storage and Elastic Search logs. The logs (on model performance and concept drift usage) can be interpreted through a Grafana dashboard, which is editable through its own GUI.
Finally, to test the success of the project, a testing plan was set up and the code was reviewed by an external group (SIG). The code achieved all the testing milestones and received a 4.5/5 in a mid-development review on maintainability. With this project, the concept of automated valuation models inside GeoPhy’s new architecture has been tested and proved and the project is ready to be further developed and used in practice.
These subjects were studied in literature, reviewing existing and upcoming valuation prac- tices in real-estate, steps needed to perform machine learning tasks, architecture to support big data processing, and concept drift. This resulted in a design made up of four different components: An ETL and data processing component, a modelling component, a Kafka con- nector, and a client-facing API. An important part to ensure efficiency and scalability of the system is the implementation of concept drift: models are only retrained when the distribu- tion of the target training value has changed significantly.
These components use storage in the form of a Postgres database, disk storage and Elastic Search logs. The logs (on model performance and concept drift usage) can be interpreted through a Grafana dashboard, which is editable through its own GUI.
Finally, to test the success of the project, a testing plan was set up and the code was reviewed by an external group (SIG). The code achieved all the testing milestones and received a 4.5/5 in a mid-development review on maintainability. With this project, the concept of automated valuation models inside GeoPhy’s new architecture has been tested and proved and the project is ready to be further developed and used in practice.