At NITU MISIS, a neural network has been introduced that accurately identifies objects not included in databases.

Students from the National University of Science and Technology MISIS and the Moscow Institute of Physics and Technology (MIPT), in collaboration with researchers from the non-profit laboratory T-Bank AI Research, have developed a new ensemble neural network called SDDE (Saliency Diversified Deep Ensembles). This innovative network enhances the accuracy of object recognition in images that are not part of existing databases. In the future, this algorithm is expected to advance the fields of autonomous vehicles and medical diagnostics, where distinguishing between unidentified elements and graphical artifacts is crucial.

2024-10-31 10:00:41https://naked-science.ru/article/column/obekty-ne-zagruzhennye-v

Researchers presented their work in October at the international image processing conference ICIP 2024 in Abu Dhabi (UAE).

With the growing volume of data, there is a need for more robust neural networks capable of not only classifying new objects but also recognizing technical noise that inevitably arises during image acquisition. The aggregate of all unknown information is referred to as out-of-distribution data (OOD). Human error in detecting OOD can lead to undesirable consequences. The creators of the algorithm addressed this issue through the diversity of the ensemble model, which reduced the correlation between inputs and improved the overall accuracy of the system.

The ensemble neural network SDDE consists of several models trained on subsets of specific databases, allowing each to focus on unique characteristics of the images. This is achieved through the diversification of the attention maps of each model—a concept that helps understand where the neural network is looking. As a result, the diversity of the ensemble increases, enabling the neural network to identify objects in images with minimal error. To assess the effectiveness of the neural network, the researchers conducted tests on several databases: CIFAR10, CIFAR100, and ImageNet-1K. The ensemble neural network SDDE demonstrated the best results compared to similar algorithms, such as Negative Correlation Learning and Adaptive Diversity Promoting.

“One of the most important tasks in developing machine learning models is ensuring that the real probability aligns with what the neural network outputs. In other words, the neural network is as confident as it is easy for it to predict the target for a given sample. Typically, networks have no doubt about their predictions at all. In this study, we proposed a new method of ensemble diversification based on logits—that is, the values that the neural network outputs before converting them into probabilities.

This innovation has improved the accuracy of the neural network's 'opinion' when detecting out-of-distribution data, which is critical for applying models in real-world conditions. For example, in autonomous driving, it is essential to accurately identify objects on the road to prevent accidents. In medical diagnostics, a comprehensive database is required for accurate diagnosis. Uncalibrated models can be overly confident in their incorrect assumptions. Our neural network lacks excessive confidence, allowing it to assess its calculations more appropriately,” said Maxim Zhdanov, a third-year student at the Institute of Computer Science at NITU MISIS.

To better detect noise artifacts, the researchers employed the Outlier Exposure approach, which involves training the model on specialized datasets containing examples of OOD. Previously, scientists from MISIS University and HSE introduced a new neural network LAPUSKA, which improves image quality twice as fast compared to similar products.