RoNFA: Robust Neural Field-based Approach for Few-Shot Image Classification with Noisy Labels

arXiv:2506.03461 · cs.CV · Submitted 2025-06-03 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "RoNFA: Robust Neural Field-based Approach for Few-Shot Image Classification with Noisy Labels".

Jane: Here's the gist of the paper. The gist: The proposed model achieves state-of-the-art accuracy in few-shot image classification with noisy labels by employing a robust neural field-based approach called RoNFA, which utilizes two neural fields for feature and category representation, soft K-means clustering for class representation, and scale adaptation to ensure prediction accuracy despite label errors.

Tom: Few-shot learning (FSL) methods aim to train classifiers for new categories using only a few labeled samples Most discussions on FSL, often assume that the support set samples are accurately labeled to represent their categories However, this assumption rarely holds in real-world scenarios since error labels can happen in sample collecting, labeling, or their transition, almost inevitable due to weakly supervised annotation methods, ambiguities, or human errors

30, 24, 16: The performance of typical FSL models significantly relies on the accurately labeled sample When trained on the samples with noisy labels, many FSL methods may dramatically lose their accuracy and struggle in practical applications

31: Most current learning methods for handling mislabeled samples are large-sample methods

5, 9, 11: These methods assume that the majority of the dataset is correctly labeled, enabling the model to statistically identify and mitigate the impact of a small number of mislabeled samples by leveraging correctly labeled samples to estimate the noise distribution or correct erroneous labels

23: For instance, noise-tolerant loss functions such as the mean absolute error or robust cross-entropy focus less on outliers, reducing their influence during optimization

27, 9: Alternatively, label correction approaches may iteratively identify and relabel noisy samples by comparing their predictions with model confidence scores

1: However, extensive noise may lead to unreliable data distribution when the sample is small.

Jane: Representative for Category The support set often contains very few samples and practical scenarios frequently introduce noisy labels Label noises pose a significant challenge since traditional classifiers rely on accurate labels and assume that they reflect the correct class distribution, but noisy labels violate this To address these issues, a feasible way is correcting the label errors with sample distribution, as shown in Figure 1 Following this idea, we cluster the support samples to generate the representatives for their categories according to the clustering results Since we have known the number of categories, we leverage K-means clustering and let K be category number m The K-means clustering is an iterative process.

Tom: It usually randomly selects K initial cluster centers before it starts.

Title and authors: Tom: So, let’s talk about the title and who wrote this. The paper is called RoNFA: Robust Neural Field-based Approach for Few-Shot Image Classification with Noisy Labels, written by Nan Xiang, Lifeng Xing, and Dequan Jina.

Jane: The authors are from the School of Mathematics and Information Science at Guangxi University in China; they’re pushing the boundaries here by focusing on how to make few-shot learning work even when the labels are noisy.

Lu: The core idea is moving beyond standard methods that assume clean labels, instead proposing this neural field approach as a robust way to classify images under those tricky conditions.

Meng: A neural field model sounds complex; I need to know if this is something we can actually deploy without needing massive GPU clusters just to set up the initial fields.

Tom: The authors are focused on building these two specific fields, one for features and one for categories, which they link together using their classification framework.

Jane: They’re essentially creating a structured way for the model to represent both what an image looks like and what category it belongs to simultaneously.

Lalam: This structure is key because it allows the category representation neurons to have a receptive field that shifts based on the feature representation, which is something new.

The paper's summary: Tom: The summary explains that RoNFA uses these two neural fields—the Feature Field for features and the Category Field for categories—to perform the classification task.

Jane: It describes how the category neurons get their input from the feature field, and this connection is what allows them to adjust their sensitivity dynamically based on where they are in space.

Lu: They introduce a specific mathematical formulation for this interaction, showing that it’s like a difference between two Gaussian functions determining the shape of that receptive field.

Meng: So they are using these Gaussian shapes to define exactly how much input from the feature field matters to each category neuron, which sounds like it gives more control over what information gets through.

Tom: They then detail the process for generating those category representatives using soft K-means clustering, which as we discussed, creates weighted averages instead of hard points.

Jane: This clustering part is crucial because it handles the randomness in sample selection by allowing samples to have a probability of belonging to multiple classes during that iterative process.

Lalam: They use this soft assignment weight w i,c and then update the cluster center mu c,k using a weighted sum of the support samples' features.

The paper's improvements: Tom: The paper points out several key improvements over existing few-shot methods. One is this soft representation via K-means clustering, which they say significantly outperforms the hard K-means based approach when testing under symmetric label swap noise.

Jane: That means if you have a set of support samples and you randomly swap two labels around, the soft method handles that confusion much better than just assigning every sample to one class rigidly.

Lu: Another big contribution is this adaptive receptive field control, where the system adjusts its classification sensitivity according to the neuronal activation in the Category Field.

Meng: That sounds like it could be a way to dynamically manage model complexity; if a category neuron isn't strongly activated by the input features, it can scale back its influence.

Tom: And they also mention a non-backpropagation learning mechanism based on a local learning rule derived from the Hebbian rule, which lets them learn feature representations without relying on standard error backpropagation through layers.

Jane: That’s interesting because if you skip the deep backprop, you change how the model learns those core visual features, which is a big design choice for representation learning.

Conclusion: Tom: So to wrap up RoNFA, it’s a neural field-based approach that uses two fields and soft K-means to create robust category prototypes that handle noisy labels effectively.

Jane: It shows that by being adaptive with the receptive fields and using weighted averages for those prototypes, you can achieve very high accuracy on real datasets even when the input labels are imperfect.

Lu: The paper suggests that this combination of soft clustering and adaptive field control is what gives the model its superiority over current FSL algorithms trained on clean support sets.

Meng: From a deployment standpoint, if we can make those fields manageable in size, it means we might be able to deploy few-shot classification systems that are much more reliable in messy, real-world scenarios.

Lalam: I’m confident that this framework will help us build more resilient systems because it directly addresses the issue of label noise without needing to clean the entire dataset beforehand.

Nan Xiang, Lifeng Xing, Dequan Jina

School of Mathematics and Information Science · Center for applied mathematics of Guangxi, Guangxi University

cs.CV

Submitted: 2025-06-03

Updated: 2025-06-03

Comments: 7 pages, 1 figure

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 87/100

The gist: The gist: The proposed model achieves state-of-the-art accuracy in few-shot image classification with noisy labels by employing a robust neural field-based approach called RoNFA, which utilizes two

Key concepts

RoNFA: A novel neural field-based approach designed for few-shot image classification that specifically addresses the challenge of noisy labels. It integrates feature and category representations using two distinct neural fields and soft K-means clustering to generate robust class prototypes, leading to superior performance under label errors.
Soft K-Means Clustering: An iterative clustering technique used to generate category representatives. Unlike standard K-means, this method uses a 'soft' strategy where samples can probabilistically belong to multiple clusters. This results in weighted average cluster centers that better reflect the true, noisy sample distribution.
Neural Fields (FFR/FCR): The model employs two neural fields: a Feature Field for representing input image data and a Category Field for representing class categories. These fields interact, and their receptive fields dynamically adjust based on the support samples, allowing the system to adapt its sensitivity during classification.
Adaptive Receptive Field Control: A mechanism where the sensitivity of category neurons is dynamically adjusted based on the activation patterns within the Category Field. This adaptation ensures that when classifying a new image, the model focuses its attention appropriately, leading to more accurate predictions despite label inaccuracies.

Terminology

Accuracy — 3× in this paper · explained in 1 episode
Robustness — 3× in this paper · explained in 6 episodes
Few-Shot Learning — 2× in this paper · explained in 5 episodes
Clustering — 1× in this paper · explained in 2 episodes

Summary

The gist: The proposed model achieves state-of-the-art accuracy in few-shot image classification with noisy labels by employing a robust neural field-based approach called RoNFA, which utilizes two neural fields for feature and category representation, soft K-means clustering for class representation, and scale adaptation to ensure prediction accuracy despite label errors.

Introduction

Few-shot learning (FSL) methods aim to train classifiers for new categories using only a few labeled samples Most discussions on FSL, often assume that the support set samples are accurately labeled to represent their categories However, this assumption rarely holds in real-world scenarios since error labels can happen in sample collecting, labeling, or their transition, almost inevitable due to weakly supervised annotation methods, ambiguities, or human errors [30, 24, 16] The performance of typical FSL models significantly relies on the accurately labeled sample When trained on the samples with noisy labels, many FSL methods may dramatically lose their accuracy and struggle in practical applications [31] Most current learning methods for handling mislabeled samples are large-sample methods [5, 9, 11] These methods assume that the majority of the dataset is correctly labeled, enabling the model to statistically identify and mitigate the impact of a small number of mislabeled samples by leveraging correctly labeled samples to estimate the noise distribution or correct erroneous labels [23] For instance, noise-tolerant loss functions such as the mean absolute error or robust cross-entropy focus less on outliers, reducing their influence during optimization [27, 9] Alternatively, label correction approaches may iteratively identify and relabel noisy samples by comparing their predictions with model confidence scores [1] However, extensive noise may lead to unreliable data distribution when the sample is small.

Representative for Category

The support set often contains very few samples and practical scenarios frequently introduce noisy labels Label noises pose a significant challenge since traditional classifiers rely on accurate labels and assume that they reflect the correct class distribution, but noisy labels violate this To address these issues, a feasible way is correcting the label errors with sample distribution, as shown in Figure 1 Following this idea, we cluster the support samples to generate the representatives for their categories according to the clustering results Since we have known the number of categories, we leverage K-means clustering and let K be category number m The K-means clustering is an iterative process. It usually randomly selects K initial cluster centers before it starts. However, this eliminates the connection between the obtained cluster and the sample category For this issue, we calculate the center of the support samples in the cth category μc 0=1/Nc∑j=1x cj and employ it as the initial center for the cth cluster instead to keep their correspondence After that, we calculate the Euclidean distance from the i-th support sample to the cth class center by d 0i,c=x i - μc 0 squared. Since the number of support samples is limited, the clustering results are easily affected by randomness in sample selection. To reduce its impact, we use a soft strategy that allows samples to belong to multiple clusters rather than rigidly assigning each data point to a single cluster. To find the new center of the cth cluster at the k-th iterative step, k=0, 1, 2,…, we first calculate the soft-assignment weight with a Gaussian kernel by the following equation: w i,c = e-d k i,c∑m k=1 e-d k i,k where i = 1, 2,… N and c = 1, 2,… m. The weight w i,c indicates the probability for the sample x i belonging to the class c. Then we update the cth cluster center μc k by the following equation: μc k=∑N i=1 w i,ckx i/∑N i=1 w i,c, c = 1, 2,… m. The obtained cluster centers are weighted averages, which better describe the realistic sample distribution and are less impacted by the randomness in sample selection. We further calculate the Euclidean distance from the i-th support sample to the obtain cth cluster center by d k i,c = x i - μc k squared. Repeat the clustering process. When the μc k, c = 1, 2,… m become stable, that is, ∑m c=1 μc k - μc k-1 < ε, where ε is a small positive constant, or k reaches a given upper bound kup, we end the clustering process and let the primary representative sample of the cth categorȳx c = μc k, c = 1, 2,… m. The K-means clustering is an unsupervised process insensitive to label noises. The soft strategy reduces the impact of the significant randomness in sample selection in fewshot learning. They together ensure the representatives of ̄x c for its category. Since there is only one representative for each category, in these procedure, we may not relabel some support samples because of their low weights. In this case, we have to abandon them to reduce the impact of label error. The fewer support samples requires the model more powerful few-shot learning capability.

Framework of Classifier

The proposed classifier utilize two neural fields, one for feature representation and the other one for category representation The field for feature representation (FFR) consists of m neurons v c located at ̄x c, c = 1, 2,… m and altering according to support samples. The field for category representation (FCR) also consists of m neurons u c, c = 1, 2,… m. They receive stimuli from their receptive fields in FFR. Recent research on neuroscience discovered that the receptive fields drift during learning, changing the sensitivity to its inputs [21, 17]. Inspired from this, we suppose the receptive field of an FCR neuron centered at corresponding v c whose position relies on the support samples. For a new input image sample I, let x = Net(I). Then its impact on the FCR neurons are formulated by the following equation: Φsigma(x, ̄x c) = Ae-1/2x-̄x c 2/sigma squared - Be-1/2x-̄x c squared (3σ) squared, (3σ) squared. whose right-hand side is the difference of Gaussian functions determining the shape of the receptive field, homogeneous with a Mexican hat shape. The constant σ > 0 determines the excitatory and inhibitory ranges of Φsigma<ref:2506.

Improvements for AI systems

Sample Representation via Soft K-Means Clustering for Robust Prototypes: The system can generate category representatives that are weighted averages rather than rigid single points, as described in Section 4.1, which better describe the realistic sample distribution and are less impacted by the randomness in sample selection. This allows the model to handle support sets where samples might belong to multiple classes probabilistically.
Adaptive Receptive Field Control: The system can dynamically adjust its classification sensitivity based on input features by modifying the receptive field parameter according to neuronal activation in FCR, as stated in the Abstract. This mechanism ensures that the range of these receptive fields adapts according to the neuronal activation in FCR to ensure prediction accuracy, addressing scenarios where a fixed scale might result in no neuron activated, single neuron activated, or multiple neurons activated.
Non-Backpropagation Learning Mechanism: The system can learn feature representations without relying on traditional error backpropagation by employing a local learning mechanism based on the Hebbian rule, enabling the model to operate without relying on multilayer backpropagation of error signals, as noted in Section 1's contributions.
Noise Robustness Against Label Errors: The system can achieve superior performance in few-shot learning with noisy labels, specifically demonstrating that its accuracy surpasses the results obtained by state-of-the-art FSL methods trained on clean support sets, as highlighted in the Abstract and verified across experiments on three types of label noise.
Improved Prototype Selection: The system can utilize a soft K-means algorithm to generate prototypes, which is shown to significantly outperform the hard Kmeans-based one in classification tests under symmetric label swap noise, as demonstrated in Table 5 ablation study.

Abstract

In few-shot learning (FSL), the labeled samples are scarce. Thus, label errors can significantly reduce classification accuracy. Since label errors are inevitable in realistic learning tasks, improving the robustness of the model in the presence of label errors is critical. This paper proposes a new robust neural field-based image approach (RoNFA) for few-shot image classification with noisy labels. RoNFA consists of two neural fields for feature and category representation. They correspond to the feature space and category set. Each neuron in the field for category representation (FCR) has a receptive field (RF) on the field for feature representation (FFR) centered at the representative neuron for its category generated by soft clustering. In the prediction stage, the range of these receptive fields adapts according to the neuronal activation in FCR to ensure prediction accuracy. These learning strategies provide the proposed model with excellent few-shot learning capability and strong robustness against label noises. The experimental results on real-world FSL datasets with three different types of label noise demonstrate that the proposed method significantly outperforms state-of-the-art FSL methods. Its accuracy obtained in the presence of noisy labels even surpasses the results obtained by state-of-the-art FSL methods trained on clean support sets, indicating its strong robustness against noisy labels.