DIPart

Domain Informed Partitioner

- DIPart is a domain informed feature extraction tool for low-dimensional data
- Generalized mixture model feature separator with automated self correction
- Use Cases: Compressed Sensing, Discrete Fourier Transform, Gaussian Mixture Model
- Significantly fewer parameters than competitive state-of-the-art AI methods
- Fully interpretable features with no post-hoc fine tuning

Practical Uses

Low Rank Extractor

Partition input data into user specified number of dominant interpretable features

Compressed Sensing

Jointly identify and select non-redundant dominant features

K+DFT

Extract user specified Discrete Fourier frequency basis with rank ordering

K+GMM

Extract user specified top rank-K Gaussian mixture modes including mean and variance

Automated Rejection

Self-corrects an inadvertently over-specified number of desired features

Resolving Interference

Extract interfering features. Coming Soon

Benchmarks

Coming Soon

So What's New?

- Avoid building costly N x N matrices for DFT and Inverse DFT
- Uses a N x K reduced discrete Fourier basis
- Select top-K information preserving basis for Compressed Sensing
- No post-hoc hyper-parameter tuning for feature selection
- Satisfies user-specified sparsity when possible
- HyperTune + DIPart = suggest sparsity alternatives