Current inspire-classifier is using very outdated packages and technologies so maintaining this service will be a huge pain. Also the accuracy seems to be not very satisfying for some categories like quantum-ph. We can try to write a POC of inspire-classifier (we can use fastai as the current classifier or play with different algorithms and libraries) and experiment with hyperparameters tuning with kubeflow.
Current inspire-classifier is using very outdated packages and technologies so maintaining this service will be a huge pain. Also the accuracy seems to be not very satisfying for some categories like quantum-ph. We can try to write a POC of inspire-classifier (we can use fastai as the current classifier or play with different algorithms and libraries) and experiment with hyperparameters tuning with kubeflow.