Implementation of Chris Bishop's Mixture Density Networks based on an inspiring blog post and RStudio's recently released TensorFlow API.
Using a relatively simple test case, estimating x = 7sin(3y/4) + y/2 (with Gaussian noise), TensorFlow successfully converges and fits the model. An example of sample (B=1000) drawn from the predicted distribution of y given a random test sample is shown:
- Employ magrittr %>% paradigm to make construction of TensorFlow model easier to read
- Move code for constructing TensorFlow model into functions
- Create a Jupyter notebook with illustrative code
- Find a more interesting example!
- Explore extensions of model
- Higher-dimension covariates
- More complex neural architecture
- Multivariate Gaussian (non-diagonal covariance)