APLIKO TANI
E-Shërbime
E-Shërbime
Apliko online

A new artificial intelligence architecture that learns from few examples.

Authors: Oriod Malo, Alexandre Xavier Falcão dhe Laurent Najman — IEEE Access

Can an artificial intelligence system learn to recognize new objects after seeing only one or five examples? For humans, this is often a natural ability. For neural networks, which are typically trained on thousands of images, it is an entirely different challenge.

This challenge is known as Few-Shot Learning (FSL). It becomes particularly important in medicine, where labeled images may be scarce, expensive to annotate, or difficult to obtain. Instead of always seeking more data, we can ask: can we improve the way the network itself perceives images?

This is precisely the question addressed by our paper “A Novel Lightweight Multi-Scale Branching Encoder for Few-Shot Image Classification”, accepted for publication in the international journal IEEE Access. It proposes Sum12, a novel convolutional architecture that learns visual features at multiple scales without relying on an excessively complex model.

How Does It Work?

Imagine that you are analyzing an image: a small detail may be decisive, but its meaning also depends on the broader structure surrounding it. Traditional convolutional neural networks do not always combine these levels of information directly.

Sum12 uses seven parallel convolutional branches. Each branch processes the same information through a different sequence of filters with sizes of 3×3, 6×6, or 9×9. The rule that connects all branches is simple: the sum of the filter sizes in each sequence is 12. In this way, the network analyzes structures at different scales while maintaining a compact architecture.

Një arkitekturë e re e inteligjencës artificiale që mëson nga pak shembuj
Finally, the outputs of the branches are aligned to the same dimensions and averaged. This information is then used by the Prototypical Network to compare new images with the representatives (“prototypes”) of the classes. Therefore, the innovation does not lie in an entirely new decision-making method, but rather in a more expressive approach to feature extraction.

Why Does It Matter?

Accuracy improvements are often achieved by building increasingly larger models. Sum12 follows a different approach: using seven parallel branches to combine features at different scales. The architecture was evaluated on 16 datasets: Omniglot, CIFAR-FS, FC100, and miniImageNet for Few-Shot Learning, as well as 12 MedMNIST 2D collections. The comparison with Prototypical Networks was conducted under the same experimental conditions. Here are some of the results:
Krahasimi me Prototypical Networks
On Omniglot, with 60 classes and one example per class, accuracy increased from 82.93% to 87.75%. Sum12 does not outperform every existing architecture: deeper networks can achieve higher accuracy. However, the results demonstrate the value of a compact encoder for few-shot learning.

What About Medical Imaging?

Across the 12 MedMNIST 2D datasets, which include images from different modalities, Sum12 was also evaluated on supervised classification (not Few-Shot Learning). At a resolution of 224×224, it achieved an average accuracy of 83.4% and an AUC of 91.9%. Google AutoML Vision achieved an AUC of 92.7%, but an average accuracy of 80.9%. Sum12 ranked first in accuracy on four of the 12 datasets. This makes the architecture an interesting approach for medical imaging research, although it has not yet been validated for clinical decision-making. The study also identified variability in results on very small and imbalanced datasets, such as BreastMNIST.

A Step Towards Resource-Constrained Devices

The standard encoder has approximately 1.88 million parameters, while the dilated convolution variant has approximately 0.74 million. After INT8 quantization, the model was reduced to approximately 799 KB and deployed on the STM32 NUCLEO-H753ZI microcontroller. Discrepancies between the results obtained on a computer and those obtained on the microcontroller indicate that this implementation has not yet been functionally validated for Few-Shot classification. Addressing these discrepancies remains a direction for future work.

Where Can It Be Applied?

  • In visual recognition tasks with only a few labeled images per class.
  • In medical image analysis when only limited data is available.
  • In the development of compact models for resource-constrained devices.
This paper brings together researchers from UniversitetY Metropolitan Tirana, University of Campinas, Khalifa University dhe Université Gustave Eiffel / CNRS. Publikimi: Malo, O., Falcão, A. X., & Najman, L. “A Novel Lightweight Multi-Scale Branching Encoder for Few-Shot Image Classification.” IEEE Access (versioni i pranuar). DOI: https://doi.org/10.1109/ACCESS.2026.3745693