Positive-Unlabeled (PU) learning aims to achieve high-accuracy binary classification with limited labeled positive examples and numerous unlabeled ones. Existing cost-sensitive-based methods often rely on strong assumptions that examples with an observed positive label were selected entirely at random. In fact, the uneven distribution of labels is prevalent in real-world PU problems, indicating that most actual positive and unlabeled data are subject to selection bias. In this paper, we propose a PU learning enhancement (PUe) algorithm based on causal inference theory, which employs normalized propensity scores and normalized inverse probability weighting (NIPW) techniques to reconstruct the loss function, thus obtaining a consistent, unbiased estimate of the classifier and enhancing the model's performance. Moreover, we investigate and propose a method for estimating propensity scores in deep learning using regularization techniques when the labeling mechanism is unknown. Our experiments on three benchmark datasets demonstrate the proposed PUe algorithm significantly improves the accuracy of classifiers on non-uniform label distribution datasets compared to advanced cost-sensitive PU methods.

Positive-Unlabeled (PU) learning aims to achieve high-accuracy binary classification with limited labeled positive examples and numerous unlabeled ones. Existing cost-sensitive-based methods often rely on strong assumptions that examples with an observed positive label were selected entirely at random. In fact, the uneven distribution of labels is prevalent in real-world PU problems, indicating that most actual positive and unlabeled data are subject to selection bias. In this paper, we propose a PU learning enhancement (PUe) algorithm based on causal inference theory, which employs normalized propensity scores and normalized inverse probability weighting (NIPW) techniques to reconstruct the loss function, thus obtaining a consistent, unbiased estimate of the classifier and enhancing the model's performance. Moreover, we investigate and propose a method for estimating propensity scores in deep learning using regularization techniques when the labeling mechanism is unknown. Our experiments on three benchmark datasets demonstrate the proposed PUe algorithm significantly improves the accuracy of classifiers on non-uniform label distribution datasets compared to advanced cost-sensitive PU methods.

NP hard optimization problems like the Traveling Salesman Problem (TSP) defy efficient solutions in the worst case, yet real-world instances often exhibit exploitable patterns. We propose a novel patternaware complexity framework that quantifies and leverages structural regularities e.g., clustering, symmetry to reduce effective computational complexity across domains, including financial forecasting and LLM optimization. With rigorous definitions, theorems, and a meta learning driven solver pipeline, we introduce metrics like Pattern Utilization Efficiency (PUE) and achieve up to 79 percent solution quality gains in TSP benchmarks (22 to 2392 cities). Distinct from theoretical NP hardness, our approach offers a unified, practical lens for pattern-driven efficiency.

Positive-Unlabeled (PU) learning aims to achieve high-accuracy binary classification with limited labeled positive examples and numerous unlabeled ones. Existing cost-sensitive-based methods often rely on strong assumptions that examples with an observed positive label were selected entirely at random. In fact, the uneven distribution of labels is prevalent in real-world PU problems, indicating that most actual positive and unlabeled data are subject to selection bias. In this paper, we propose a PU learning enhancement (PUe) algorithm based on causal inference theory, which employs normalized propensity scores and normalized inverse probability weighting (NIPW) techniques to reconstruct the loss function, thus obtaining a consistent, unbiased estimate of the classifier and enhancing the model's performance. Moreover, we investigate and propose a method for estimating propensity scores in deep learning using regularization techniques when the labeling mechanism is unknown.

Federated learning (FL) is a novel learning paradigm that addresses the privacy leakage challenge of centralized learning. However, in FL, users with non-independent and identically distributed (non-IID) characteristics can deteriorate the performance of the global model. Specifically, the global model suffers from the weight divergence challenge owing to non-IID data. To address the aforementioned challenge, we propose a novel diffusion strategy of the machine learning (ML) model (FedDif) to maximize the FL performance with non-IID data. In FedDif, users spread local models to neighboring users over D2D communications. FedDif enables the local model to experience different distributions before parameter aggregation. Furthermore, we theoretically demonstrate that FedDif can circumvent the weight divergence challenge. On the theoretical basis, we propose the communication-efficient diffusion strategy of the ML model, which can determine the trade-off between the learning performance and communication cost based on auction theory. The performance evaluation results show that FedDif improves the test accuracy of the global model by 10.37% compared to the baseline FL with non-IID settings. Moreover, FedDif improves the number of consumed sub-frames by 1.28 to 2.85 folds to the latest methods except for the model compression scheme. FedDif also improves the number of transmitted models by 1.43 to 2.67 folds to the latest methods.

As artificial intelligence (AI) steadily grows, so do concerns about its environmental footprint. Today's emerging natural language processing (NLP) models, such as GPT-3 can consume as much energy as five cars, according to a 2019 study. To reduce their environmental and climate impact, researchers in the United Arab Emirates are proposing a new development approach for these models that takes energy consumption into account at every stage, aiming to boost energy efficiency wherever possible. Last April, Abu Dhabi's Technology Innovation Institute (TII) launched NOOR, the largest Arabic-language NLP model to date. NOOR – Arabic for "light" – is trained on 10 billion parameters including books, poetry, news, and technical information, reinforcing the model's broad applicability, according to its creators.

The power and energy monitoring in carbontracker is limited to a few main components of computational systems. Additional power consumed by the supporting infrastructure, such as that used for cooling or power delivery, is accounted for by multiplying the measured power by the pue of the data center hosting the compute, as suggested by Strubell2019. Previous research has examined pue and its shortcomings (Yuventi2013). These shortcomings may largely be resolved by data centers reporting an average pue instead of a minimum observed value. In our work, we use a pue of 1.58, the global average for data centers in 2018 as reported by Ascierto2018.222Early

Atlas 900 stood out with its world-leading AI computing power, ultimate heat dissipation system, and best-in-class cluster network. Atlas 900 accelerates global basic AI research and quickly brings AI applications to industries to advance the AI era with unparalleled AI computing power. Innovative technology has propelled the mobile industry far beyond the wildest expectations of early tech pioneers. GSMA awards the GLOMO Award – Tech of the Future Award to recognize technology that is ahead of its time and reshapes the world. Atlas 900 is the world's fastest AI training cluster.

Advances in technology can allow you to order food by voice or unlock your phone with your face, but those new capabilities could take a toll on the environment. Enhanced tech capabilities are being developed through the use of artificial-intelligence approaches like neural networks, which detect patterns in speech and images by training programs across countless data points. That process constantly crunches reams of information on power-hungry servers in data centers that use a substantial amount of energy to power, cool and monitor the servers. The result: Training a neural network can emit 17 times more carbon dioxide than an average American does in a year, and five times the lifetime emissions of an average car. Those are the findings of a recent paper by researchers at the University of Massachusetts, Amherst, which highlighted the substantial power generated by AI technologies.

The 40% energy saving on cooling helped one of Google's data centres to achieve a 15% reduction in power usage efficiency, or PUE. PUE is defined as the ratio of the total building energy usage (pumps, chillers, cooling towers) to the IT energy usage (Google's servers). The lower the PUE, the better.