Continual Learning for Open Set Recognition with Deep Generative Models

Assignment:

Open set recognition (OSR) aims to detect the novel classes for trained machine learning models during inference. Deep generative models, e.g., variational autoencoders (VAEs), have been widely applied for OSR problems. However, in most works for OSR, the models are static, i.e., the models will not be updated even though the novel classes are detected. One more realistic scenario in practice is that the novel classes will not be novel anymore after being detected. And therefore, the OSR models should be updated in order to incorporate the newly detected classes. Continual learning is the paradigm to update the machine learning models for novel samples. The objective of this work is to investigate how to update the VAE models for OSR with continual learning algorithms.

Tasks:

In this thesis, the student is expected to finish the following tasks:

  1. Review the related work on VAEs for OSR and continual learning for VAEs.
  2. Test several methods on VAE on OSR on open source datasets and expand it to continual learning fashion through exploring the continual learning methods studied in and propose necessary adaptations for VAEs.
  3. Test OSR accuracy on open datasets and analyze the results.

Requirements:

Valid knowledge of deep learning and programming skills in Python and related machine learning frameworks, such as PyTorch.

Contact Person:

Jiawen Xu, M.Sc. (jiawen.xu ∂ campus.tu-berlin.de)

Start: Immediately

References:

  1. Cao et al., “Open set recognition with Gaussian mixture variational autoencoders,” AAAI’20.

  2. Ran et al., “Detecting out-of-distribution samples via variational auto-encoder with reliable uncertainty estimation,” Neural Computing, Elsevier, 2022.

  3. Wiewel et al., “Continual learning for anomaly detection with variational autoencoder,” ICASSP’19.

Notes:

The supervision will be remote and online. The student is still expected to work closely with the supervisor, and the final work would be submitted to a machine learning conference.