Hello,
I've build a Conv1D model to classify items into 6 categories (from 0 to 5). (simplified example)
I use therefore a CrossEntropy loss function that penalizes the errors with high probabilities and also the low probabilities associated to right answers. And that is exactly what I was looking for.
But CrossEntropy considers the classes as independent, while I'd like reduce the loss for an error where model predicts 0 (cat) while it is 1 (dog) and increase the loss when the model predicts 0 (cat) when the true answer is class 5 (fish).
Is there a way to embed the 'class proximity' in the loss computation of a classification problem, so that a error between classes 0 and 1 will be less penalized than an error between classes 2 and 5 ?
Many thanks in advance for any help !
You might want to implement a custom loss function based on the true and predicted category. You will need to design your own loss function since it depends on the classes of your problem:
https://towardsdatascience.com/how-to-create-a-custom-loss-function-keras-3a89156ec69b
In addition, maybe these more general papers will help you out:
Article A Comprehensive Survey of Loss Functions in Machine Learning
https://machinelearningmastery.com/how-to-choose-loss-functions-when-training-deep-learning-neural-networks/
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification