How many neurons and layers did you select for your neural network model? Why?
- 2 Layers
- First Layer => 36 Neurons
- Second Layer => 18 Neurons
This decision was made after I experimented with higher number of neurons. I realized that as I increased the number of neurons, the accuracy of the neural network model diminished. In addition, dropping the number of neurons below 36 and 18 for the first and second layers, respectively, adversely affected model accuracy.
I also added a third layer but this had little effect on the model accuracy.
Were you able to achieve the target model performance? What steps did you take to try and increase model performance?
- Model Accuracy => 73 percent
- Loss => 0.55
I was unable to attain the model accuracy of 75 percent.
Initially I used 'relu' but switched to 'sigmoid' for both layers as well as the output and this measure resulted in an improved model accuracy.
I experimented with increasing the epochs but this had a negative effect on accuracy. Adding layers and neurons had no positive effect as well. I also switched the loss function from 'binary_crossentropy' to 'mean_squared-error' and 'mean_absolute_error' and both changes yielded no improvement.
Based on this evaluation, I believe the steps that I can take to improve the model include:
- Experimenting with different optimizers
- Work on additional preprocessing of the dataset. Probably some other variables need to be binned into different categories.
If you were to implement a different model to solve this classification problem, which would you choose? Why?
Honestly, I need to fully explore more on each model to fully understand the appropiate situations to use them. But probably, I might try a 'tanh' or 'leaky relu'.