Building and analysing single hidden layer neural network on data with sampling & replacement

AI Discussions
1

Hello all
Embarking on my first ML (single hidden layer neural network) task.
Below is the workflow I chalked out.
Any seasoned inputs/suggestions/feedback much appreciated.

Actual data 1000 images., 2 labels.
From original data set, Repeat 10 times with replacement:
1.Segregate randomly 250 images & dump in each of training & CV/dev dataset.
2.Train the model on 1 above.
3.Check for bias(b)& variance(v), readjust relevant parameters to mitigate b&v.

(Note: each model may differ in regularization, number of hidden nodes parameters)

  1. Simulate predictions on 500 of test set.
    Finally get all prediction accuracy, cost function values and truth tables for the 10 models in the iterative process.

  2. 3d Plot with regularization parameter on x axis, number of nodes on y axis and accuracy on z axis.

  3. Similar 3d plot like 5 above but with cost function values on z axis.

  4. Compute precision and recall from the 10 models and generate 3d plots separately.

  5. Cross check with outputs from human decision and compare performance.

Thanks for your time

2

As far as I read seem quite thorough and analytical but make sure you dont contaminate each the training of the model with validation and test data.

3

Thanks much for the caution! :+1:t5:

1 Like