Feeding "noise" to AI can also boost its performance scores, and this work enables the positive transfer of noise.
The "source data" in transfer learning does not necessarily have to be images, text, or audio.
A set of randomly sampled noise without any semantic meaning drawn from a Gaussian distribution can also help models learn better with limited annotations.
This work is named Semi-Supervised Noise Adaptation (SSNA), and its corresponding paper has been published at ICML 2026.
The SSNA team further proposes the Noise Adaptation Framework (NAF), which leverages randomly generated noise to construct discriminative category structures, and transfers this structure to real target data, so as to improve the model's learning performance in scenarios with limited annotations.
△ NAF projects the noise domain and the target domain into a shared representation space and aligns them at the category level. Image from the paper.
With only 4 annotated samples per class, based on the ResNet-18 backbone, NAF achieves accuracy improvements of 12.35, 7.61, 4.38, and 2.74 percentage points on CIFAR-10, CIFAR-100, DTD-47, and Caltech-101, respectively, compared to ERM (Empirical Risk Minimization), the standard supervised learning baseline trained solely on annotated samples. Currently, the source code of the paper has been open-sourced, with details provided at the end of the article.
Replace the Real Source Domain with Noise
Traditional transfer learning usually requires a source domain with abundant labels. However, real source data is not always easily accessible. Privacy, confidentiality, and copyright restrictions may all prevent source domain data from being shared. SSNA attempts to remove this prerequisite: the source domain no longer contains real samples, but is replaced by noise generated from a simple probability distribution.
The specific implementation is not complicated. Assuming the target task contains C classes, the research team first randomly samples a mean vector for each class in a 1024-dimensional space, uses the identity matrix as the covariance matrix, thereby constructs C Gaussian distributions, and then samples 50 noise vectors from each distribution. The noise domain and the target domain share the same set of category indices, with each noise class pre-assigned to a target category index, but this correspondence itself contains no semantic information.
The indices here carry no inherent semantic meaning. Noise class 0 does not inherently represent "cat"; it is simply fixed to correspond to a certain class in the target domain before training starts. What is truly transferred is not the visual knowledge of cats or dogs, but the discriminative structure formed in the noise domain.
Noise of the same class is clustered together, while noise of different classes is separated. As long as this structure can be aligned with the target domain, it can provide clearer classification boundaries for real target samples.
A Small Number of Labels Remain Necessary
Noise can replace real source data, but it cannot completely replace target domain labels. The reason is straightforward: noise comes from another space, and the category indices are artificially assigned. The model must rely on a small number of annotated target samples to figure out which real class noise class 0 should be aligned with.
When the number of annotated samples per class on CIFAR-100 is reduced to 0 in the paper, the accuracy of ERM and NAF is only 0.97% and 1.34%, respectively, both close to random level. Once a small number of annotations are provided, NAF consistently outperforms ERM.
Therefore, the conclusion of this work is: In the semi-supervised classification setting, a real source domain is not necessarily a prerequisite for achieving positive transfer.
NAF Mainly Does Three Things
Based on the generalization upper bound derived in the paper, NAF splits the training objective into three parts.
First, Learn the Limited Real Labels Well
The target domain encoder maps real samples to the shared representation space, and the classifier calculates the cross-entropy loss using a small number of annotated samples. This part is consistent with ordinary supervised learning, serving as a bridge between noise classes and real categories.
Then Make Noise Form a Clear Category Structure
The noise projector is responsible for mapping random vectors to the same representation space, and the classifier identifies these noise samples according to the pre-defined category indices. After training, noise of the same class gradually converges, while noise of different classes becomes separated from each other.
Finally, Align the Two Sides
NAF also calculates the distribution discrepancy between the noise domain and the target domain. The distribution alignment module does not restrict specific implementations; the paper compares multiple schemes and empirically adopts Negative Domain Similarity (NDS) as the default mechanism in experiments. NDS compares both the global mean and the per-class mean of the two domains, and uses cosine similarity to push them closer. The class means of unannotated target samples are iteratively estimated using pseudo-labels generated by the model.
The entire objective can be written as
, where
handles the classification of the limited annotated target samples,
handles noise classification,
handles distribution alignment between the target domain and the noise domain. The generalization upper bound derived in the paper also corresponds to these three terms: the target domain empirical error, the noise domain empirical error, and the distribution discrepancy between the two domains in the shared representation space.
Noise Can Bring Gains from CIFAR to ImageNet
The main experiments cover 8 visual datasets and 1 text classification dataset. Except for ImageNet-1K, all visual tasks use 4 annotated samples per class, with the remaining training samples treated as unannotated data.
On ResNet-18, the Top-1 accuracy improvements of NAF over ERM are: CIFAR-10 +12.35, CIFAR-100 +7.61, DTD-47 +4.38, and Caltech-101 +2.74 percentage points, respectively.
△
Fine-grained classification also benefits. Using ResNet-18, NAF outperforms ERM by 8.94, 5.51, and 7.74 percentage points on CUB-200, Oxford Flowers-102, and Stanford Cars-196, respectively.
On the larger-scale ImageNet-1K, the research team retains 100 annotated samples per class. NAF achieves an accuracy of 37.10%, which is 0.99 percentage points higher than ERM. On the text task AG News-4, NAF using BERT reaches 82.82%, 4.18 percentage points higher than ERM's 78.64%.
NAF can also be directly plugged into existing semi-supervised methods. The paper integrates it into UDA, FixMatch, FlexMatch, DebiasMatch, DST, LERM, and SA-FixMatch, and all of them obtain overall performance gains. Taking the 20-epoch training results on CIFAR-10 as an example, after adding NAF to UDA and FixMatch, their accuracy increases by 20.83 and 9.91 percentage points, respectively.
What Is Truly Useful Is Not "Randomness", But Structure
Why does noise help? The ablation experiments in the paper point to the same factor: whether the classes have a separable structure.
The team first collapses all noise of all classes into a single point. In this case, the noise domain completely loses its discriminative structure, and NAF not only brings no gains but also shows significant negative transfer. The CIFAR-10 accuracy drops from ERM's 58.15% to 33.34%, while CIFAR-100 drops from 42.24% to 6.79%.
On the contrary, when the distance between the noise class centers is gradually increased, the CIFAR-100 accuracy steadily rises from 43.80% all the way to 49.78%. This indicates that the more distinguishable the noise classes are, the stronger the structural guidance they can usually provide.
The quantity of noise is not that critical. When the number of noise samples per class increases from 10 to 100, the accuracy remains stable at around 50%; after increasing to 200, it even drops slightly. Based on this, the paper argues that as long as a separable pattern can be formed, a small amount of noise is sufficient to exert its effect.
△
The research team also simplifies the noise domain to one center point per class. Whether the centers are fixed or learnable, the results are higher than ERM; among them, learnable centers perform better than fixed centers, but still underperform the full NAF.
In the transfer experiment from Amazon to Caltech-10, the real source data still performs slightly better overall, but the noise source domain has already improved the accuracy from ERM's 83.51% to approximately 88% to 89%. This also gives a more realistic positioning: when real source data is unavailable, synthetic noise can become a very low-cost alternative solution.
It is also different from common data augmentation. Data augmentation usually performs rotation, cropping, interpolation, or generation around real samples; SSNA first constructs an independent noise domain, and then completes cross-domain alignment in the representation space.
Summary: Structure Can Still Be Transferred Without Semantics
This work provides a counterintuitive new perspective for transfer learning: The fact that source data can help target tasks does not necessarily fully depend on its real semantic meaning. The category structure formed in the representation space can also become transferable knowledge.
NAF leverages exactly this point, allowing random noise to form a discriminative structure in the shared representation space, and then transfers this structure to real data with the help of a small number of annotated samples. Experimental results show that even if the source domain contains no real images, text, or audio, as long as an appropriate category structure is preserved, it can still have a positive effect on the target task.
In other words, what transfer learning transfers may not only be "what the data says", but also "how the data is organized in the representation space". This also provides a new research direction for scenarios with limited privacy, copyright sensitivity, or difficulty in accessing real source data.
Paper Title: Semi-Supervised Noise Adaptation: Transferring Knowledge from Noise Domain
Paper Link: https://arxiv.org/pdf/2606.00558
Source Code Link: https://github.com/AIResearch-Group/SSNA
Video Explanation: https://www.bilibili.com/video/BV1UV7h61EvW/
This article is from the WeChat Official Account "QbitAI", authored by the SSNA team, and published by 36Kr with authorization.