ClimateBench Neural Process
The project was done as a further extension upon my Senior Capstone Project. Which I undertook the challenge due to my interest in the Deep Learning and hybrid models that were used in the project. In my Senior Project I worked on a Deep Kernel Learning model which combined the capabilities of a Gaussian Process with a Neural Network forming a more complex and improved model. However, there are many limitations to the model, such as it’s limited ability to incorporate the spatial and temporal dependencies of the climate data. This is where a Neural Process comes in, which is a model instead takes and blends ideas from neural networks and Gaussian processes.
Neural Process still in development
Conclusion
Although as of writing this conclusion the project is still incomplete, and following the incomplete report there is much that still needs to be completed. For example, I am trying to implement 3 variations of the Neural Process, which are the Base Neural Process which takes in a sparse representation of the climate data, the Convolutional Neural Process which takes in a spatial representation of the climate data using a CNN to extract features from the data and finally a Spherical Convolutional Neural Process which takes in a spatial representation of the climate data using a spherical CNN with the hopes to have the model understand the spherical nature of the Earth and the climate data. All of these models are still in development and I hope to have them completed soon.
