View in #questions-forum on Slack
@Weston: Thank you @CireNeikual. My system is an Android edge device, where the problems are face authentication and malware detection. Both scenarios, I am in need of a class incremental learning for classification. Have found code for incremental SVM many places but it will not handle new classes. Hence, read about it and found that, we need class incremental learning for doing this. Exploring about this, Do we have any pointers to find the code for class incremental learning for classification problems? (Eg: SVM) . I will port it to Android(Java/C/C++)
@vlomonaco: I think you can find SVM implementations where you can also add new classes over time. Still if you are interested in a neural approach you can take a look at our recent paper: https://arxiv.org/abs/1912.01100
We will be releasing more details on the app with a follow up technical paper later this year
arXiv.org: Latent Replay for Real-Time Continual Learning
@Weston: Thank you for the support. I will analyse the paper.
@mattdl: In your use case, you might also want to take a look at few-shot learning. For example in face authentication you might want to use a few samples of a some users and quickly get an adapted version at runtime from these ‘few shots’.
@Alex_Torex: for face auth you need zero shot learning
[6:23 PM] to enroll new users
[6:23 PM] the net should predict a distance between embeddings of 2 faces
[6:24 PM] just search face verification state of art papers on paperswithcode
[6:24 PM] you need somthing like siamese nets or better
@MUHAMMED_NABEEL_K: In the present scenario there will be enough data sets available for face authentication. In that case can Few-shot learning be used or Zero-shot learning? Can we avoid retraining by using this?