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

OW-DETR: Open-world Detection Transformer

Open-world object detection (OWOD) is a challenging computer vision issue in which it is necessary to simultaneously detect a set of recognized object categories and identify unfamiliar objects. In addition, the model must gradually acquire new classes when introduced in subsequent training episodes. Unlike conventional object detection, the OWOD environment provides major hurdles for generating high-quality candidate proposals on potentially unknown things, isolating unidentified objects from the background, and recognizing various unknown objects.

The researchers present OW-DETR, a novel transformer-based end-to-end framework for open-world object identification. The proposed OW-DETR consists of three dedicated components, namely attention-driven pseudo-labelling, novelty classification, and objectness score, to specifically solve the aforementioned OWOD issues. Their OW-DETR store's multi-scale contextual information exhibits less inductive bias, promotes knowledge transfer from available classes to the unknown type, and can more effectively distinguish between unknown objects and backgrounds. Two benchmarks, MS-COCO and PASCAL VOC, are subjected to exhaustive trials. Extensive ablations demonstrate the value of their promised contributions.

In terms of unknown recall on MS-COCO, their model surpasses the previously introduced OWOD technique, ORE, with increases ranging from 1.8% to 3.3%. On PASCAL VOC, OW-DETR outperforms the state-of-the-art for all incremental object detection settings.

Proto2Proto: Can you recognize the car the way I do?

Prototypical approaches have recently received much attention because of their inherent interpretability, provided through prototypes. In addition, with the increased use of model reuse and distillation, there is an increased need to investigate the transfer of interpretability from one model to another.

The researchers introduce Proto2Proto, a unique knowledge-distillation-based strategy. Their strategy seeks to provide interpretability to the "black" knowledge conveyed from the teacher to the student's model, which is more superficial. To facilitate such a transfer, the researchers suggest two novel losses: "Global Explanation" loss and "Patch-Prototype Correspondence" loss. Global Explanation Loss compels student prototypes to be near to teacher prototypes, but Patch-Prototype Correspondence Loss causes local student representations to resemble those of the teacher.

In addition, the researchers suggest three additional metrics to quantify the student's proximity to the teacher as measures of interpretability transfer in their contexts. Finally, the researchers use CUB-200-2011 and Stanford Cars datasets to demonstrate our strategy's efficacy qualitatively and statistically. Experiments show that the suggested approach successfully transfers interpretability from teacher to student while maintaining competitive performance.

Robust outlier detection by de-biasing VAE likelihood

Deep networks frequently produce confident but wrong predictions when put to the test with data that is an extreme departure from their training distributions. Deep generative model (DGM) likelihoods are a potential statistic for outlier detection in unlabeled data. However, prior research has demonstrated that DGM likelihoods are inaccurate and are easily skewed by trivial changes to the input data.

Here, the researchers suggest innovative algorithmic and analytical techniques to reduce significant biases in VAE likelihoods. We may easily calculate their bias corrections for a variety of observable decoder distributions, are sample-specific, and are computationally affordable. The researchers next demonstrate how contrast stretching, a well-known picture pre-processing technique, increases the usefulness of bias correction to enhance outlier detection further.

With nine grayscale and natural picture datasets, their method achieves state-of-the-art accuracies and outperforms four rival approaches in speed and performance. In conclusion, simple solutions can provide reliable outlier detection with VAEs.