Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It only takes a minute to sign up.
Sign up to join this community
I am very new at machine learning, and I'm building an artificial neural network that aims to classify inputs into 2 labels. I am training the network with randomly initialized weights and through the stochastic gradient descent. What would you say is an acceptable accuracy for training set and test set? Is around 80% for test set good enough or should I aim to get close to 100%?
Thank you so much in advance.
It’s impossible to say in general. It’s even impossible to say in general what makes for an acceptable $R^2$ value in linear regression! This will depend on the problem at hand and what has been done on the same or similar data sets. You may think that 95% accuracy is fantastic. For MNIST, for instance, that’s nothing special.
Finally, I would encourage you to read what Frank Harrell and others on this Stack have to say about accuracy (even AUC) as opposed to “proper scoring rules” (such as Brier score). Shamelessly, I will link a post of mine where I challenge this a bit, though it even contains an example where you can see an argument for why accuracy might not be the best performance metric.
Proper scoring rule when there is a decision to make (e.g. spam vs ham email)
