1 Hasso Plattner Institute, 2 Chinese University of Hong Kong, 3 Renmin University of China
Novel class discovery (NCD) is a machine learning task focused on finding new classes in the data that weren't available during the training period.
How can we borrow supervised knowledge and break the category constrain?
Vaze et al (CVPR 2022) Generalized Category Discovery
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
What makes the implementation of NCD possible?
What makes the implementation of NCD possible?
Supervised info \( \mathbf{X} | Y \)
What makes the implementation of NCD possible?
Supervised info \( \mathbf{X} | Y \)
Unsupervised info \( \mathbf{X} \)
Vaze et al (CVPR 2022) Generalized Category Discovery
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
DL:
STAT:
DL:
STAT:
Motivated question:
Is supervised knowledge always helpful?
DL:
STAT:
Step 1
Step 2
Step 3
Suppose we learn a mapping \(\mathbf{p}\) from training samples
How to measure the effectiveness of \(\mathbf{p}\)
Recall: MMD
Muandet et al (2020) Kernel Mean Embedding of Distributions: A Review and Beyond
Recall: MMD
Muandet et al (2020) Kernel Mean Embedding of Distributions: A Review and Beyond
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
Yet, in practice, \(Y_u\) is unknown...
Step 1
Step 2
Step 3
Conclusion: consistency between Semantic Similarity and Accuracy. The proposed benchmark is good...
Conclusion: consistency among Semantic Similarity, Accuracy, and (pseudo) transfer flow. The proposed metric is good...
Step 1
Step 2
Step 3
Step 4
Suboptimal
Conclusion: Supervision information with low semantic relevance may hurt NCD performance.
Conclusion: pseudo transfer flow can be used as a practical reference to infer what sort of data we want to use in NCD.
Application: Data selection
Application: Data Combining