I did the "Deep Learning with Flux" course provided by Julia Academy (https://juliaacademy.com/courses/). I express my sincere gratitude to Dr. Matt Bauman for presenting this course and sharing the associated Jupyter Notebook. This notebook is associated with second lecture "Introduction to Neural Networks" and third lecture "Deep Learning with Flux" of the mentioned course.
As I was attending the lecture, I realized I need to code myself to understand and record my inferences and in the process came up with this working notebook. Intent here is to have this for my future reference as well as be useful to someone new to the field learning ML/DL.
This notebook provides a quick introduction to Neural Networks by having a Non Binary Classification Model implemented to classify Apples, Bananas and Grapes Images.
Details of what this notebook covers is listed below:
- How to setup the Environment
- Where to get the datasets from
- Which packages to install and how
- Basic Model
- Understanding Multiple Output Model
- Quick recap on Matrix Multiplication
- Prepare input data for the model
- Quick Introduction to One Hot Vectors
- Prepare output labels for the input data
- Define Model
- Train the model and visualize the results
- Updated Model
- Quick Introduction to Linearity and Non-Linearity
- Define Updated Model
- Improve Training Time by using 'Flux.batch'
- Train model multiple times using 'Iterators.repeated'
- Check Loss Value
- Visualization of Decision Boundaries
- Improvement Opportunity - Better Loss Function
- Improvement Opportunity - Normalization of Output Data
- Update Model with Improvement Opportunities
Basic Model Summary
Model used here is quite simple to help understand the concept of multiple outputs. However, note it is not able to encapsulate the classes uniquely (refer below decision boundaries... red/green/blue do not have only one respective class on one side). The model needs to be more complicated to handle this complexity.
Updated Model Summary
Model is able to separate grapes very well and to large extent the apples and bananas. However, there is overlap seen in grapes and bananas ... we can improve this by adding more features to enable the model to separate them. e.g. we used mean value of red and blue of the fruits... probably using the mean value of green as well might help.
