I’m seeking guidance on improving a deep learning model I’ve developed using the Boston Housing Dataset. Despite completing Course 3, I’m finding it challenging to apply the concepts to regression problems.
Currently, say my model’s performance is as follows:
Training MSE: 38.44
Development MSE: 2.34
Unfortunately, I don’t have access to expert predictions for house prices, which limits my ability to further assess for Bayes error and refine the model.
what I can see is the training MSE is too large, meaning I need to bigger model?
I can only assume Bayes error should be less than 2.34 (as in dev MSE)
Could you provide more advice or resources that might help me bridge the gap between the course content and applying these techniques to a regression problem like this? Any insights or suggestions would be greatly appreciated.
Thank you for your time and assistance.
Not necessarily a “bigger” model, but maybe a more complex one.
Fundamental things to try:
- add units to the hidden layer
- change the activation in the hidden layer
- add another hidden layer
However, I see you are overfitting the dev set (compared to the training results). That’s unusual.
How were the training and dev sets selected?
Another check you can try.
Create a totally artificial simulated training set, for a simple scenario that that your model should have no trouble fitting.
Then use it to test your model.
Absolutely, here’s a more concise version of your message:
I’ve been using
train_test_split for splitting training and development sets, and I’m now exploring the use of a simulated training set.
However, I find myself no idea on how to proceed when encountering error outputs. I need advice on practical steps for addressing such issues in real-world scenarios. Thanks.