Thesis proposal – Enhancing Weather Classification Machine Learning Models with DRLSH and BPLSH

Title:

Enhancing Weather Classification Machine Learning Models with DRLSH and BPLSH

Background:

When predicting the weather, you want to be consistent with reality in both timeline and accuracy. If there is a storm brewing you want to be able to predict when, where and how it will affect us. Modern weather predictions use multiple highly complex physical & chemical models that, when combined, can be used to describe the oncoming weather. These models are pretty good and are always improving, there is, however, a limiting factor, time.

Due to the complexity of our models, there are hundreds if not thousands of factors and variables that drastically affects the computational time of the predictions. Nowadays these are primarily carried out by supercomputers and even then, a few hours’ worth of predictions can take days to compute.

One possible solution to this problem is machine learning. By creating a model that learns from real-life weather data, employing a neural network, we achieved a model that is good enough to predict the weather with reasonable accuracy and certainty. However, this approach is not without its limitations. Machine learning models, while efficient, can sometimes struggle to capture the intricate dynamics of weather systems, leading to occasional inaccuracies.

This leads us to consider whether we can further enhance the trained models by incorporating advanced algorithms like DRLSH (Deep Randomized Locality Sensitive Hashing) and BPLSH (Binary Partitioning Locality Sensitive Hashing). These algorithms are known for their efficiency in data retrieval tasks and have been used in various domains to accelerate complex computations. However, their application in weather prediction is relatively unexplored.

This thesis proposal is a continuation of a previous thesis project done at Syntronic. You can find an article about the work here: Classification of Weather Conditions Based on Supervised Learning for Swedish Cities 

Link to BPLSH algorithm article: Efficient and decision boundary aware instance selection for support vector machines

Link to DRLSH algorithm article: A fast instance selection method for support vector machines in building extraction

An overview of this thesis project, could include: 

• To develop a weather model, including training the layers of the neural network using data provided by, for example, SMHI, NOAA (National centers for environmental information) or any equivalently sophisticated API.
• Building BPLSH and DRLSH algorithms.
• Testing the models with BPLSH and DRLSH algorithms.
• Validation on real-world weather scenarios and the possibility to compare your results.
• A report about your theory, methodology, results and discussion of your results.

For a successful thesis work, we believe that the applicants should have:

• Experience in C++, Python and/or MATLAB as well as respective machine learning libraries (TensorFlow and/or PyTorch etc…).
• Experience in machine learning is recommended.

Application:

We look forward to receiving your resume, and preferably, a personal letter in which you explain why you want to write your thesis with Syntronic.

We screen and evaluate applications on an ongoing basis.