Test-time adaptation (TT A) aims to address distributional shifts between training and testing data using only unlabeled test data streams for continual model adaptation.

Test-time adaptation (TTA) aims to address distribution shifts between source and target data by relying solely on target data during testing. In open-world scenarios, models often encounter noisy samples, i.e., samples outside the in-distribution (ID) label space. Leveraging the zero-shot capability of pre-trained vision-language models (VLMs), this paper introduces Zero-Shot Noisy TTA (ZS-NTTA), focusing on adapting the model to target data with noisy samples during test-time in a zero-shot manner. We find existing TTA methods underperform under ZS-NTTA, often lagging behind even the frozen model. We conduct comprehensive experiments to analyze this phenomenon, revealing that the negative impact of unfiltered noisy data outweighs the benefits of clean data during model updating. Also, adapting a classifier for ID classification and noise detection hampers both sub-tasks. Built on this, we propose a framework that decouples the classifier and detector, focusing on developing an individual detector while keeping the classifier frozen. Technically, we introduce the Adaptive Noise Detector (AdaND), which utilizes the frozen model's outputs as pseudo-labels to train a noise detector. To handle clean data streams, we further inject Gaussian noise during adaptation, preventing the detector from misclassifying clean samples as noisy. Beyond the ZS-NTTA, AdaND can also improve the zero-shot out-of-distribution (ZS-OOD) detection ability of VLMs. Experiments show that AdaND outperforms in both ZS-NTTA and ZS-OOD detection. On ImageNet, AdaND achieves a notable improvement of $8.32\%$ in harmonic mean accuracy ($\text{Acc}_\text{H}$) for ZS-NTTA and $9.40\%$ in FPR95 for ZS-OOD detection, compared to SOTA methods. Importantly, AdaND is computationally efficient and comparable to the model-frozen method. The code is publicly available at: https://github.com/tmlr-group/ZS-NTTA.

Sputnik / Evgenya Novozhenina The contribution of artificial intelligence (AI) to the global GDP will increase 16-fold in the next 12 years, according to the head of Russia's largest bank Sberbank Herman Gref. "Expansion of artificial Intelligence in the coming years is likely to only grow. According to forecasts of a number of companies, if today AI contributes $1 trillion to global GDP, then according to forecasts of consulting companies, this figure will increase 16-fold over the next 12 years, until 2030," he said. The number of specialists in demand in the area will also increase significantly, he added, explaining that in 10 years the need will reach ten million people. A recent study by McKinsey Global Institute suggested that AI could boost annual GDP growth by 1.2 percent for at least the next decade.

Apparently, teaching artificial intelligence to read our innermost thoughts or turning them into terrifying psychopaths isn't enough - now researchers are teaching AI systems to predict what humans will do in the future (and how long you'll be doing it) "minutes or even hours" before we decide to do it. It's fine when Google finishes your sentences when typing into a search bar, but this new technology might be able to recognize patterns in human behavior and perform tasks before you've even thought about asking. Like most tasks performed by artificial intelligence, this ability is tied to machine learning and neural networks. In the course of their research, a team from the University of Bonn in Germany tried out two models for their networks: one that made predictions and "reflected" before making new more, and one that was based on a matrix structure. Both networks were shown videos of people making relatively simple food dishes (especially breakfasts and salad) with the goal of teaching them to predict what the chef was going to do next.