Category: World

The field of machine learning has long relied on Kolmogorov-Arnold Networks (KANs) to approximate complex multivariate functions. Traditionally, these networks used B-spline bases, which are effective but computationally intensive. Researchers needed a solution that balanced performance with efficiency.

Recent developments introduced the FastKAN, an improved version of KANs that utilized Gaussian radial basis functions (RBFs). However, FastKAN’s reliance on a fixed kernel and shape parameter limited its adaptability. This restriction created a demand for more flexible approaches that could better accommodate varying data characteristics.

The latest study presents the adaptive RBF-KAN, which integrates a diverse range of radial basis kernels, including Matérn and Wendland functions. By implementing leave-one-out cross-validation for kernel shape initialization, this framework allows for data-driven kernel scaling that enhances performance during network training. Evaluations on benchmark functions showed significant improvements in handling smooth and complex function behaviors.

The introduction of adaptive kernel selection marks a significant advancement in RBF-KAN models. Researchers found that different kernels exhibited distinct advantages, proving that adaptability is key to effectively addressing various function types. This innovation could reshape how neural networks process data, paving the way for more efficient and accurate machine learning applications.

In the world of Reinforcement Learning (RL), the prevailing belief has been that large-batch training often leads to diminishing returns. Researchers typically avoided large batches, especially after a certain point in training, due to the instability it could introduce. This norm shaped how RL algorithms were developed and fine-tuned.

However, recent findings challenge this long-held view. A study introduced Adaptive Batch Scaling (ABS), which alters batch sizes based on the stability of the learning policy. This approach hinges on a new metric called Behavioral Divergence, allowing for a more responsive adjustment that considers non-stationarity in policy behavior throughout training.

The researchers integrated ABS with the Parallelised Q-Network (PQN) algorithm, testing it against the Atari Learning Environment (ALE). Their results indicate a significant breakthrough: larger networks paired with larger batch sizes can indeed enhance performance. This counters the traditional perspective that associates larger batches exclusively with negative outcomes in RL.

The implications of these findings are profound. By reconciling rapid behavioral shifts and stable convergence, ABS opens up new avenues for RL applications. This could ultimately lead to more efficient training methods, allowing for quicker and more reliable learning across various complex tasks.

Traditionally, researchers generate hypotheses and conduct experiments to validate ideas. This process, while essential, can be time-consuming and resource-intensive. As scientific inquiry grows more complex, the need for efficient evaluation methods has never been more critical.

A shift is occurring as language models increasingly automate hypothesis generation. However, a new challenge has emerged: rapidly assessing the viability of numerous AI-generated concepts without exhaustive experiments. Recent research explores whether these models can predict which ideas are more likely to succeed based on comparative evaluations.

The study analyzed a dataset of 11,488 idea pairs from PapersWithCode. Initial performance from standard models was disappointing, achieving only 30% accuracy. However, fine-tuning with Reinforcement Learning allowed a significant boost to 77.1%, surpassing the previously leading GPT-5. This innovative approach provides interpretable reasoning for decision-making, enhancing the models’ reliability.

The implications of this research are profound. By offering a scalable method for evaluating scientific ideas, researchers can focus on the most promising hypotheses. This development paves the way for expedited discovery in various fields, potentially transforming how scientific challenges are addressed globally.

Traditional methods in feature ranking have relied on the belief that they can provide stable, faithful, and complete outputs. However, new research reveals a critical flaw: in the presence of collinear features, no ranking can maintain these three qualities simultaneously. This discovery challenges long-held assumptions in the field.

The study, published on arXiv, demonstrates that when features become collinear, ranking reduces to arbitrary outcomes, akin to a coin flip. Researchers quantified this impossibility across four model classes and proposed a solution—DASH, or Diversified Aggregation of SHAP. This ensemble method promises stability, even in cases where features are symmetrically relevant.

Analyzing 77 public datasets, the researchers found that 68% displayed attribution instability, raising serious concerns about existing SHAP-based methods. The findings suggest that switching to alternative approaches, like conditional SHAP, fails to address the core issue of collinearity. This necessitates a reevaluation of how feature importance is assessed in machine learning contexts.

The implications of this research extend to fairness auditing, revealing that SHAP-based discrimination audits may be fundamentally flawed under collinearity conditions. The work culminates in a set of practical diagnostic tools and a formally verified impossibility theorem, marking a significant milestone in explainable AI. This pivotal breakthrough redefines our understanding of feature ranking and its limitations.

The landscape of financial crime detection has remained relatively stable, relying heavily on traditional feature engineering methods. However, a new approach called the Temporal Contrastive Transformer (TCT) has emerged, promising a shift in how behavioral patterns in transactions are analyzed. Researchers have developed TCT to learn contextual temporal dynamics effectively.

This model introduces a self-supervised contrastive objective to generate embeddings that represent financial transaction sequences. Initial evaluations show TCT’s embeddings yield significant predictive performance, achieving an area under the curve (AUC) score of 0.8644. These results suggest that the model can capture essential temporal structures, raising expectations for its application in fraud detection.

Upon further testing, TCT’s performance was compared to established feature-engineered methods. Although the embeddings performed well on their own, no substantial improvement was observed when combined with existing features. The AUC scores revealed a minimal difference, indicating that TCT’s learned representations closely resemble traditional abstractions.

The findings highlight both the potential and limitations of TCT in the fight against financial crime. While it does not yet surpass manual feature engineering, the model’s ability to mirror domain-specific features marks a meaningful step forward. As research continues, there is hope that TCT will pave the way for more innovative techniques, reducing dependence on extensive feature engineering in the future.

The current landscape of industrial design is heavily reliant on the integration of Computer-Aided Design (CAD) and Computer-Aided Engineering (CAE). However, a significant challenge persists: the semantic gap between these two domains complicates the iterative optimization process. As designers grapple with ensuring that simulation feedback translates into viable geometric modifications, inefficiencies have become increasingly evident.

The introduction of COSMO-Agent seeks to address this issue by leveraging a tool-augmented reinforcement learning framework. This innovative approach enables Large Language Models (LLMs) to navigate the complexities of the CAD-CAE workflow autonomously. By framing the entire process as an interactive RL environment, COSMO-Agent effectively orchestrates the manipulation of external tools while revising geometries based on simulation results.

In extensive experiments, COSMO-Agent has demonstrated considerable advancements in design efficiency. It outperformed both large open-source and proprietary models in terms of feasibility, efficiency, and stability. The framework employs a multi-constraint reward system, promoting robust toolchain interaction and ensuring that outputs remain valid under varying constraints.

The implications for the industrial design community are profound. By streamlining the closed-loop process, COSMO-Agent not only enhances the feasibility of complex designs but also fosters innovation across multiple component categories. As a result, this development could transform standards in design optimization, pushing the boundaries of what’s achievable in manufacturing and engineering.

Researchers have long relied on complex models for estimating causal effects in high-dimensional data. Traditional approaches often depend on assumptions like causal sufficiency and the pretreatment condition. These requirements can limit applicability and hinder accurate analyses.

A recent study introduces a novel local covariate selection method that circumvents these assumptions. This approach focuses on a local boundary for effective identification, allowing researchers to search for valid adjustment sets without exhaustive computation. The technique is designed to be efficient, particularly in scenarios with numerous variables.

Following rigorous testing, the authors demonstrated that their method not only adheres to foundational soundness but also achieves completeness. Experiments across synthetic and real datasets yielded results that showcased both accuracy in estimating causal effects and enhanced computational efficiency when compared to existing methods.

This advancement could significantly impact the fields of epidemiology, economics, and social sciences. By providing a more pragmatic approach to causal analysis, researchers may now explore intricate relationships in data sets previously deemed too complex or resource-intensive to analyze effectively.

Standard Chartered has long positioned itself as a leader in banking and finance, emphasizing the importance of skilled human capital. This image was recently challenged by CEO Bill Winters’ comments regarding the role of artificial intelligence.

Following Winters’ remarks describing certain roles as “lower-value human capital,” regulators in Hong Kong and Singapore expressed their concerns. They requested further clarification on how these views might affect employment practices and corporate responsibility.

The queries highlight tensions between advancing technology and traditional workforce norms. Standard Chartered is now navigating these regulatory landscapes while attempting to reassure stakeholders about its commitment to its employees.

The repercussions could shape the bank’s reputation in the industry. As regulators probe deeper, the dialogue around AI’s impact and the future of jobs remains a critical issue for both Standard Chartered and the banking sector at large.

DeepSeek, a rising player in the artificial intelligence market, is shifting gears. Previously focused on immediate product applications, the company is now prioritizing advanced AI research. This change aligns with its long-term vision of achieving artificial general intelligence (AGI).

In recent meetings, DeepSeek’s senior management discussed its new direction with potential investors. The ongoing funding round, valued at 70 billion yuan ($10 billion), will enable the startup to funnel resources into groundbreaking research instead of short-term commercial projects. This pivot aims to place DeepSeek at the forefront of AI innovation.

As a result, DeepSeek is set to alter its organizational strategy significantly. The fresh funding will direct teams towards exploratory projects that push technological boundaries. This ambitious approach signals a departure from traditional incremental development common in the industry.

The implications are substantial. Should DeepSeek succeed in its pursuit of AGI, it could reshape the landscape of artificial intelligence. Investors are watching closely, aware that this gamble could redefine competition and innovation in tech.

Nintendo Co. has set expectations for its upcoming Switch 2 console. Initially, the company projected sales that aligned closely with conservative forecasts. However, recent developments have prompted a shift in production strategy.

Sources reveal that Nintendo has instructed its partners and suppliers to prepare for the assembly of approximately 20 million units. This figure reflects a 20% increase over the company’s earlier public sales outlook, indicating a strong belief in the system’s demand.

In response to the unexpected revision, suppliers are ramping up production capabilities to meet the anticipated demand. Analysts are closely monitoring the market to see how this increased output will influence overall sales in a competitive gaming landscape.

The decision underscores Nintendo’s confidence in the Switch 2’s appeal. If successful, this could lead to a significant boost in the company’s market position and financial performance, shaping the future of gaming consoles.