Hi, I’m working on a task to predict breast cancer from mammograms in a maximum timelapse of 2 years.
I have tried multiple models and hyperparameters but my validation loss does not improve, it remains almost the same since the beginning, always in the range of 0.69…
Both validation and training losses start at 0.69…, but at the moment that the training loss decreases to 0.68, the validation loss starts increasing, and the overfitting starts. When the loss is 0.69, the accuracy is at 0.50, which is the baseline because the data is equally balanced (50% cases, 50% controls).
That led me to think that the model was in a plateau or local minima since the beginning, and I tried to schedule the learning rate to be bigger in the first epochs, but it didn’t work.
Any suggestions or ideas about how to solve it? Maybe making the predictions using only the mammograms cannot solve the task?