# Lower loss always better for Probabilistic loss functions?

I am working on an neural net int Tensorflow that predicts percentages for win, draw, loss for given data of a game. The labels I provide are always {1, 0, 0}, {0, 1, 0} or {0, 0, 1}. After some epochs my accuracy doesn't increase any further, but the loss still decreases for a many epochs (also on the validation set, though very slowly). I am using a softmax activation in the last layer and the categorical crossentropy loss function provided by Keras. I was wondering if in this case, lower loss always corresponds to better probabilities (because I obviously wouldn't want the net to output only values like 1 or 0 for probabilities), or in other words, does this net output the "true" probabilites and if so, why does it do that?

If $$0.5$$ is the threshold for declaring a class (perhaps more sensible in a binary classification than your problem, yes), there is no incentive for accuracy to regard a $$1$$ as a $$0.95$$ instead of a $$0.51$$.

Meanwhile your cross-entropy loss function sees that the correct answer is $$1$$ and wants to get the probability as close to $$1$$ as it can. Accuracy, however, doesn't care if the predicted probability is $$0.51$$ or $$0.95$$, so accuracy does not change as you move the predicted probability closer and closer to the observed value, even though the loss function decreases by getting closer and closer to the observation (as you would expect loss to do...consider how square loss behaves in a linear regression).

• Yes that's what I thought, but I'm still not sure if the probabilities for a game are realistic when I use the logloss/ crossentropy loss function. For example, after converging, my net may output 0.9 0.05 0.05 for a sample. With different loss functions this output is going to differ (after training), even though the accuracy stays the same. So how come that the crossentropy represents actual probabilities, representative on the dataset? – Evator Jul 3 '20 at 0:57
• Log loss is an example of a proper scoring rule. You can read more on Cross Validated (stats.stackexchange.com), as that topic appears not to be popular on this Stack, but the gist is that a proper scoring rule wants to find the true probabilities. When you don't have the true probabilities, log loss will not penalize the difference between truth and prediction the same way as another proper scoring rule. Log loss is popular, however, because of its relationship to maximum likelihood estimation in logistic regression. (That warrants a separate question.) – Dave Jul 3 '20 at 1:32
• Alright, I think I got it, thanks! – Evator Jul 3 '20 at 11:40

Consider loss to be something akin to implied variability of your model. When loss is extremely high, your model's output could be just about anything. As the loss lowers, your model is becoming more confident in its output and will be able to give similar output classification/regressions even if the initial weights or input data is slightly different. Lower loss in the training set is always a good thing so long as you are seeing a subsequent decrease in the loss of the validation set. As soon as that validation loss starts to stagnate though, I like to save the model and quit. Usually at this point you could get a little better loss, but it tends to just be overfitting and doesn't help any on the test set.