Response to Reviewer xGwt
* **About the novelty of the proposed HSSL**
Heterogeneous Semi-Supervised Learning (HSSL) distinguishes itself from other SSL paradigms like Imbalanced SSL, Open Set SSL, and Universal SSL, each addressing distinct challenges in learning from limited labeled data.
In **Imbalanced SSL**, the focus is on addressing disparities in class representation within a single domain; both labeled and unlabeled data **share the same domain**, implying equal conditional distributions ($P_l(\mathbf{x}|y) = P_u(\mathbf{x}|y)$). This contrasts with HSSL, which emphasizes learning under conditions of **domain differences**, where labeled and unlabeled data come from distinct domains with different label distributions and class feature distributions.
**Open Set SSL** deals with **unknown or additional classes** present in the unlabeled data but absent in the labeled set. Open Set SSL also has the **same feature distribution** over labeled and unlabeled sets. This is different from HSSL, which operates under the assumption that both domains share **the same set of classes**, and labeled and unlabeled data come from separate domains with **different class feature distributions**.
**Universal SSL** allows for a **mix of classes** in both labeled and unlabeled sets, some of which may not appear in the other. However, the **test set** maintains the same distribution as **the labeled set**. In contrast, HSSL maintains the assumption of **shared classes** across domains and more importantly, the **test set** includes instances sampled from **both labeled and unlabeled domains** which presents a more complex challenge.
Each of these SSL paradigms addresses specific scenarios and challenges in semi-supervised learning, highlighting the diversity and complexity of learning with limited labeled data in various contexts.
* **About novelty**
Our work introduces a novel framework specifically tailored for HSSL, a task characterized by distinct challenges. While individual components, like weighted moving average pseudo-labeling, cross-domain prototype alignment, and inter-domain mixup, have been used in simpler setups, our innovation lies in their strategic integration and adaptation to the specific challenges of HSSL.
Our approach introduces novel elements to prototype alignment by utilizing a $2C$ classifier, distinguishing it from conventional prototype alignment methods. Specifically, our innovation lies in the pseudo-labeling strategy where we consider the first $C$ classes in labeled data and the second $C$ classes in unlabeled data, seamlessly integrating with the $2C$ classifier framework. Additionally, our method redefines the mixup strategy by innovatively designing the updating strategy of the moving average decay parameter, which is designed to promote information sharing and efficient and steady knowledge transfer from the labeled domain to the unlabeled domain. As a result, our approach offers a novel perspective and a valuable contribution to addressing the intricacies of the proposed novel HSSL task.
* **About using CIFAR10/100**
Heterogeneous Semi-supervised Learning (HSSL) requires the labeled and unlabeled domains to have distinct feature distributions, necessitating the condition that $P_l(\mathbf{x}|y) \neq P_u(\mathbf{x}|y)$. However, conventional datasets like CIFAR typically represent a single domain and thus don't inherently meet this condition for HSSL, which requires distinct feature distributions across domains.
* **About adding more comparisons**
We have incorporated the comparison with [9] into our analysis. It is important to note that the PatchMix [10] utilizes the Transformer as its underlying backbone and it is not directly applicable to change its backbone therefore direct comparison is not suitable. Due to time constraints, we were unable to conduct experiments comparing our proposed method to [5] and [10] within the allotted rebuttal period. However, we commit to including these comparisons in the final version of the paper.