Task 1
In this worksheet, you will explore and construct decision regions of feedforward neural networks (FNNs) with Heaviside activations.
Exercise 1
Use the tool and configure the decision region so that the neuron outputs class 1 for all points satisfying
Exercise 2
Use the following settings:
Set the classification mode
Select 2 classes and Single-Label mode.
Configure the graph display
In the graph display, enable Show weights, Show bias values, and Show activations. Set the position to 70%.
Configure the model
Use one hidden layer with a Heaviside activation and select the Heaviside activation for the output layer as well.
Click the weights and bias values to change them. The FNN should create a red decision region bounded by the following separator lines, as shown in the figure:
Exercise 3
When the FNN editor opens, the dataset containing bill length, flipper length, and sex is loaded automatically and the decision-region editor is shown. You can also download the CSV file. The column names and class labels in this prepared dataset are in German.
- Draw four decision regions.
- Combine two regions at a time into an OR group. Hold Ctrl while selecting the regions.
- Let the FNN realize the decision regions.
- Investigate: How are non-convex polygons handled?
Data source and license
Horst, A. M., Hill, A. P., & Gorman, K. B. (2020). palmerpenguins: Palmer Archipelago (Antarctica) penguin data (R package version 0.1.0) Dataset. Zenodo. https://doi.org/10.5281/zenodo.3960218
The original dataset is licensed under CC0 1.0 Universal. For this worksheet, the variables bill length (bill_length_mm), flipper length (flipper_length_mm), and sex (sex) were selected. Incomplete observations were removed, and the column and class labels were translated into German. The derived dataset contains 333 observations.