Students from the MISIS University and Moscow Institute of Physics and Technology, in collaboration with researchers from the T-Bank AI Research Lab, have introduced a new ensemble neural network called SDDE (Saliency Diversified Deep Ensembles) that can identify objects in previously unregistered images with greater accuracy than similar networks. This innovation has promising applications in fields like autonomous vehicles and medical diagnostics, where the ability to distinguish unidentified elements and graphic artifacts is crucial. The researchers presented their work in October at the ICIP 2024 International Conference on Image Processing in Abu Dhabi, UAE.
With the growing volume of data, there is a need for more reliable neural networks that can not only classify new objects but also recognize technical artifacts that inevitably arise during image acquisition. This unknown information is collectively referred to as out-of-distribution (OOD) data. Human factors in detecting OOD data can lead to undesirable consequences. The creators of this algorithm addressed this challenge by diversifying the ensemble model, which reduced correlations among inputs and improved the system’s overall accuracy.
The SDDE ensemble neural network consists of several models, each trained on subsets of distinct datasets. This approach allows each model to focus on unique image features by diversifying the attention maps of each model—a concept that helps interpret where the neural network “looks.” As a result, the ensemble becomes more diverse, enabling the network to identify objects in images with minimal error. To assess the network’s effectiveness, researchers tested it on various datasets: CIFAR10, CIFAR100, and ImageNet-1K. The SDDE ensemble neural network demonstrated superior results compared to similar algorithms, such as Negative Correlation Learning and Adaptive Diversity Promoting.
“One of the most critical tasks in developing machine learning models is aligning real probabilities with those predicted by the neural network. In other words, the network should be confident in its predictions to the extent that it accurately reflects the ease of predicting the target for a given sample. Typically, networks show no hesitation in their predictions. In this study, we proposed a new ensemble diversification method based on logits—the values the network outputs before converting them into probabilities. This innovation has improved the network’s ‘confidence’ when detecting out-of-distribution data, which is crucial for real-world applications. For instance, autonomous driving requires the network to identify objects on the road accurately to prevent accidents. In medical diagnostics, extensive databases are essential for accurate diagnoses. Uncalibrated models can be overly confident in their incorrect assumptions, but our network lacks this excessive confidence, enabling it to assess its calculations more realistically,” Maxim Zhdanov, a third-year student at the MISIS Institute of Computer Science.
To improve the detection of noise artifacts, the researchers used an Outlier Exposure approach, training the model on specialized datasets containing examples of OOD data.
Previously, scientists from MISIS University and HSE University introduced a new neural network, LAPUSKA, which enhances image quality twice as fast as comparable products.
