Supervised Knowledge May Hurt Novel Class Discovery Performance
Ben Dai (CUHK)
(Joint work with Li, Otholt, Hu, Meinel, and Yang)
IMS China 2024
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689446/pasted-from-clipboard.png)
NCD Background
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.
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/11457526/open-class.png)
Liu, Ziwei, et al. "Large-scale long-tailed recognition in an open world." CVPR. 2019.
NCD Background
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689703/pasted-from-clipboard.png)
NCD Background
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689708/pasted-from-clipboard.png)
How can we borrow supervised knowledge and break the category constrain?
NCD Background
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689754/pasted-from-clipboard.png)
NCD: Existing Methods
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690358/pasted-from-clipboard.png)
Vaze et al (CVPR 2022) Generalized Category Discovery
NCD: Existing Methods
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690372/pasted-from-clipboard.png)
NCD: Existing Methods
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690394/pasted-from-clipboard.png)
NCD: Existing Methods
What makes the implementation of NCD possible?
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689754/pasted-from-clipboard.png)
NCD: Existing Methods
What makes the implementation of NCD possible?
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689754/pasted-from-clipboard.png)
Supervised info \( \mathbf{X} | Y \)
NCD: Existing Methods
What makes the implementation of NCD possible?
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689754/pasted-from-clipboard.png)
Supervised info \( \mathbf{X} | Y \)
Unsupervised info \( \mathbf{X} \)
NCD: Existing Methods
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690358/pasted-from-clipboard.png)
Vaze et al (CVPR 2022) Generalized Category Discovery
NCD: Existing Methods
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690394/pasted-from-clipboard.png)
NCD: Existing Methods
DL:
DL typically assumes that more data is better, and the focus lies in designing different network structures to effectively utilize the available data.
STAT:
Statisticians always strive to clarify when and how to utilize data effectively in various situations/assumptions.
NCD: Existing Methods
DL:
DL typically assumes that more data is better, and the focus lies in designing different network structures to effectively utilize the available data.
STAT:
Statisticians always strive to clarify when and how to utilize data effectively in various situations/assumptions.
An interesting question:
Is more (supervised) data necessarily better?
From a practical perspective, we would like to propose a metric that can serve as a reference to guide us in determining which data to utilize, thereby avoiding the need to train a time-consuming, huge model unnecessarily.
NCD: Outline
DL:
- More data is better...
- design a DL architecture
STAT:
- under this kind of assumption you should ...
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690465/pasted-from-clipboard.png)
Step 1
Step 2
Step 3
NCD: Metric
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690480/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689754/pasted-from-clipboard.png)
NCD: Metric
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690480/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689754/pasted-from-clipboard.png)
Suppose we learn a mapping \(\mathbf{p}\) from training samples
How to measure the effectiveness of \(\mathbf{p}\)
NCD: Metric
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690500/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690515/pasted-from-clipboard.png)
Recall: MMD
Muandet et al (2020) Kernel Mean Embedding of Distributions: A Review and Beyond
NCD: Metric
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690500/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690515/pasted-from-clipboard.png)
Recall: MMD
Muandet et al (2020) Kernel Mean Embedding of Distributions: A Review and Beyond
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690527/pasted-from-clipboard.png)
Fini et al (ICCV 2021) A Unified Objective for Novel Class Discovery
NCD: Metric
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690500/pasted-from-clipboard.png)
Yet, in practice, \(Y_u\) is unknown...
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690540/pasted-from-clipboard.png)
NCD: Benchmark
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690465/pasted-from-clipboard.png)
Step 1
Step 2
Step 3
NCD: Benchmark
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690559/pasted-from-clipboard.png)
NCD: Benchmark
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690559/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690574/pasted-from-clipboard.png)
NCD: Benchmark
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690616/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690617/pasted-from-clipboard.png)
Conclusion: consistency between Semantic Similarity and Accuracy. The proposed benchmark is good...
NCD: Benchmark
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690623/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690624/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690634/pasted-from-clipboard.png)
Conclusion: consistency among Semantic Similarity, Accuracy, and (pseudo) transfer flow. The proposed metric is good...
NCD: Supervised Info May Hurt
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690465/pasted-from-clipboard.png)
Step 1
Step 2
Step 3
Step 4
NCD: Supervised Info May Hurt
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690644/pasted-from-clipboard.png)
NCD: Supervised Info May Hurt
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690644/pasted-from-clipboard.png)
Suboptimal
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690646/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690646/pasted-from-clipboard.png)
NCD: Supervised Info May Hurt
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690644/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690646/pasted-from-clipboard.png)
Conclusion: Supervision information with low semantic relevance may hurt NCD performance.
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690646/pasted-from-clipboard.png)
NCD: Supervised Info May Hurt
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690662/pasted-from-clipboard.png)
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
NCD: Supervised Info May Hurt
Application: Data Combining
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690687/pasted-from-clipboard.png)
Contribution
- We find that using supervised knowledge from the labeled set may lead to suboptimal performance in low semantic NCD datasets. Based on this finding, we propose two practical methods and achieve ∼3% and ∼5% improvement in both CIFAR100 and ImageNet compared to SOTA.
- We introduce a theoretically reliable metric to measure the semantic similarity between labeled and unlabeled sets. A mutual validation is conducted between the proposed metric and a benchmark, which suggests that the proposed metric strongly agrees with NCD performance.
- We establish a comprehensive benchmark with varying degrees of difficulty based on ImageNet by leveraging its hierarchical semantic similarity.
Thank you!
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10689446/pasted-from-clipboard.png)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/2070857/images/10690714/tmlr.png)
ncd
By statmlben
ncd
- 98