CODEWITHSUNDEEP
machine-learningneural-network
milo204
milo204

Reputation: 125

Should ANN results change based on output?

Changing the number of output variables reducing accuracy of the model, although architecture of the NN remains the same (ie same no of input and hidden neurons, same activation func, same data used for learning).

Hi, Ive run a NN with shape (22,20,8) and received an rms value of 5.048 and R2 of 0.925, with one of the output variables predicting to 98% accuracy on new data. However Ive deceided that i dont need all 8 outputs (only 4 are of interest) so ive run the model again with shape (22,20,4) and received an rms value of 8.20, R2 of 0.62 and accuarcy of a variable at 75%.

If im only changing the output, how does this effect the accuracy of the model ive already created? Im using the same data, so surely the relationship learned from the model will be the same and therefore predictions on the same remaining 4 variables will be the same? How come changing the output changes the model so much? Thanks

Upvotes: 1

Views: 780

Answers (1)

skillsmuggler
skillsmuggler

Reputation: 1902

If I'm only changing the output, how does this effect the accuracy of the model I've already created?

As form your question, you must have realized that the model accuracy changes. This behavior is as expected, because you are changing the model.

When you retrain the model with 4 outputs instead of 8 outputs, the relationships learnt are different from the model with 8 outputs.

Changing even small factors like the initial weights and bias, the random states of various initializes and any other tune-able parameters, will for sure result in a change in accuracy.

In you case, a change in the architecture (big change) will and has created a big impact on the results.

Upvotes: 1

Related Questions