These are the most exciting AI research articles published this year. It mixes artificial intelligence (AI) innovations with data science. It is ordered chronologically and contains a link to a longer article.

A Train-Time Regularizing Loss for Improved Neural Network Calibration

Deep Neural Networks (DNNs) are prone to produce overconfident errors, making their usage in safety-critical applications troublesome. SOTA calibration strategies boost the confidence of projected labels alone while leaving non-max classes (e.g., top-2, top-5) uncalibrated. Such calibration is incompatible with label refinement via post-processing. Furthermore, most SOTA systems learn a few hyper-parameters after the fact, omitting the possibility of an image or pixel-specific calibration. As a result, they are unsuited for calibration under domain shift or intensive prediction tasks such as semantic segmentation.

The researchers propose in this work to intervene during the train time itself to directly build calibrated DNN models. They present a novel auxiliary loss function: Multi-class Difference in Confidence and Accuracy (MDCA), which can be used in conjunction with other application/task-specific loss functions to obtain the same results. The researchers demonstrate that MDCA training results in better-calibrated models in terms of Expected Calibration Error (ECE) and Static Calibration Error (SCE) on picture classification and segmentation tasks. Finally, the researchers demonstrate that MDCA training enhances calibration even when the data is uneven and for natural language classification tasks.

The researchers have released the code here.

Energy-based Latent Aligner for Incremental Learning

Deep learning models tend to forget their prior knowledge as they learn new tasks incrementally. This behaviour occurs because parameter updates optimized for new tasks may not be compatible with adjustments acceptable for older tasks. Because of the resultant latent representation mismatch, forgetfulness occurs. The researchers suggest ELI: An energy-based Latent Aligner for Incremental Learning, which first learns an energy manifold for the latent representations so that prior task latents have low energy values and current task latents have high energy values. This acquired manifold compensates for the representational shift during incremental learning. Their proposed methodology's implicit regularisation can be employed as a plug-and-play module in existing incremental learning approaches.

Furthermore, the researchers validated this using CIFAR-100, ImageNet subset, ImageNet 1k, and Pascal VOC datasets. When ELI is applied to three critical approaches in class-incremental learning, they see consistent progress across different incremental settings. Furthermore, when combined with a state-of-the-art incremental object detector, ELI improves detection accuracy by more than 5%, demonstrating its effectiveness and complementing advantage over the existing art.

Factorized Learning of Object Attributes for Improved Multi-object Multi-part Scene Parsing

Detecting several object classes in a scene and segmenting semantic portions within each object is difficult in multi-object multi-part scene parsing. The researchers offer FLOAT, a factorized label space framework for scalable multi-object multi-part parsing, in this publication. Compared to the monolithic label space equivalent, their architecture incorporates a dense independent prediction of object category and part attributes, which boosts scalability and minimizes task complexity. Furthermore, the researchers offer an inference-time 'zoom' refinement strategy that enhances segmentation quality dramatically, particularly for smaller objects/parts.

On the Pascal-Part-58 dataset, FLOAT achieves an absolute improvement of 2.0% for mean IOU (mIOU) and 4.8% for segmentation quality IOU (sqIOU) compared to the state-of-the-art. The more extensive Pascal-Part-108 dataset upgrades are 2.1% for mIOU and 3.9% for sqIOU. Finally, the researchers combine previously rejected components to generate the most comprehensive and complex version, Pascal-Part-201. The new dataset improves FLOAT by 8.6% for mIOU and 7.5% for sqIOU, demonstrating its parsing effectiveness across a challenging range of objects and parts.

The code and datasets are available here.