Assuming that there always exists a function/NN that can perfectly model the data, I apply a neural network/random forest or ... etc. to data. If my model has a training and validation MSE/loss that are very similar, it means that my model generalises 'well'.

However if both this training and val loss are unacceptably small, and converge to these values pretty quickly, does it mean that more complex models are required to model this data? I believe this is true as large differences between validation and training loss would suggest that my model has not been able to learn properly with the amount of data and I would need more training data. Therefore I have enough training data, and assuming that a function/NN exists to be almost a perfect approximation, (and my MSE's are too high but therefore can be made much lower) am I correct in reasoning that my next step is to increase model complexity?




Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.