LLMs use “Bag of Heuristics” to do approximate retrieval.
Category: information science – Page 32
AI Breakthrough Solves Supercomputer Math on Desktop PCs in Seconds
A breakthrough in artificial intelligence.
Artificial Intelligence (AI) is a branch of computer science focused on creating systems that can perform tasks typically requiring human intelligence. These tasks include understanding natural language, recognizing patterns, solving problems, and learning from experience. AI technologies use algorithms and massive amounts of data to train models that can make decisions, automate processes, and improve over time through machine learning. The applications of AI are diverse, impacting fields such as healthcare, finance, automotive, and entertainment, fundamentally changing the way we interact with technology.
Drilling into Neutron Stars with Computers
Simulations of neutron stars provide new bounds on their properties, such as their internal pressure and their maximum mass.
Studying neutron stars is tricky. The nearest one is about 400 light-years away, so sending a probe would likely take half a million years with current space-faring technology. Telescopes don’t reveal much detail from our vantage point, since neutron stars are only the size of a small city and thus appear as mere points in the sky. And no laboratory on Earth can reproduce the inside of neutron stars, because their density is too great, being several times that of atomic nuclei. That high density also poses a problem for theory, as the equations for neutron-star matter cannot be solved with standard computational techniques. But these difficulties have not stopped efforts to understand these mysterious objects. Using a combination of theory-based methods and computer simulations, Ryan Abbott from MIT and colleagues have obtained new, rigorous constraints for the properties of the interior of neutron stars [1].
Quantum simulators: When nature reveals its natural laws
Quantum physics is a very diverse field: it describes particle collisions shortly after the Big Bang as well as electrons in solid materials or atoms far out in space. But not all quantum objects are equally easy to study. For some—such as the early universe—direct experiments are not possible at all.
However, in many cases, quantum simulators can be used instead: one quantum system (for example, a cloud of ultracold atoms) is studied in order to learn something about another system that looks physically very different, but still follows the same laws, i.e. adheres to the same mathematical equations.
It is often difficult to find out which equations determine a particular quantum system. Normally, one first has to make theoretical assumptions and then conduct experiments to check whether these assumptions prove correct.
A brain-inspired algorithm that mitigates catastrophic forgetting of artificial and spiking neural networks with low computational cost
Neuromodulators in the brain act globally at many forms of synaptic plasticity, represented as metaplasticity, which is rarely considered by existing spiking (SNNs) and nonspiking artificial neural networks (ANNs). Here, we report an efficient brain-inspired computing algorithm for SNNs and ANNs, referred to here as neuromodulation-assisted credit assignment (NACA), which uses expectation signals to induce defined levels of neuromodulators to selective synapses, whereby the long-term synaptic potentiation and depression are modified in a nonlinear manner depending on the neuromodulator level. The NACA algorithm achieved high recognition accuracy with substantially reduced computational cost in learning spatial and temporal classification tasks. Notably, NACA was also verified as efficient for learning five different class continuous learning tasks with varying degrees of complexity, exhibiting a markedly mitigated catastrophic forgetting at low computational cost. Mapping synaptic weight changes showed that these benefits could be explained by the sparse and targeted synaptic modifications attributed to expectation-based global neuromodulation.
Scientists develop technology to control cyborg insect swarms
Scientists have developed an advanced swarm navigation algorithm for cyborg insects that prevents them from becoming stuck while navigating challenging terrain.
Published in Nature Communications, the new algorithm represents a significant advance in swarm robotics. It could pave the way for applications in disaster relief, search-and-rescue missions, and infrastructure inspection.
Cyborg insects are real insects equipped with tiny electronic devices on their backs—consisting of various sensors like optical and infrared cameras, a battery, and an antenna for communication—that allow their movements to be remotely controlled for specific tasks.
A Multi-Objective Framework for Balancing Fairness and Accuracy in Debiasing Machine Learning Models. A Multi-Objective Framework for Balancing Fairne
Machine learning algorithms significantly impact decision-making in high-stakes domains, necessitating a balance between fairness and accuracy. This study introduces an in-processing, multi-objective framework that leverages the Reject Option Classification (ROC) algorithm to simultaneously optimize fairness and accuracy while safeguarding protected attributes such as age and gender.
A quantum walk simulation of extra dimensions with warped geometry
We investigate the properties of a quantum walk which can simulate the behavior of a spin 1/2 particle in a model with an ordinary spatial dimension, and one extra dimension with warped geometry between two branes. Such a setup constitutes a \(1+1\) dimensional version of the Randall–Sundrum model, which plays an important role in high energy physics. In the continuum spacetime limit, the quantum walk reproduces the Dirac equation corresponding to the model, which allows to anticipate some of the properties that can be reproduced by the quantum walk. In particular, we observe that the probability distribution becomes, at large time steps, concentrated near the “low energy” brane, and can be approximated as the lowest eigenstate of the continuum Hamiltonian that is compatible with the symmetries of the model. In this way, we obtain a localization effect whose strength is controlled by a warp coefficient. In other words, here localization arises from the geometry of the model, at variance with the usual effect that is originated from random irregularities, as in Anderson localization. In summary, we establish an interesting correspondence between a high energy physics model and localization in quantum walks.
Anglés-Castillo, A., Pérez, A. A quantum walk simulation of extra dimensions with warped geometry. Sci Rep 12, 1926 (2022). https://doi.org/10.1038/s41598-022-05673-2
A3ML: Anticipatory and Adaptive Anti-Money Laundering
DARPA seeks to revolutionize the practice of anti-money laundering through its A3ML program. A3ML aims to develop algorithms to sift through financial transactions graphs for suspicious patterns, learn new patterns to anticipate future activities, and develop techniques to represent patterns of illicit financial behavior in a concise, machine-readable format that is also easily understood by human analysts. The program’s success hinges on algorithms’ ability to learn a precise representation of how bad actors move money around the world without sharing sensitive data.
DARPA wants to eliminate global money laundering by replacing the current manual, reactive, and expensive analytic practices with agile, algorithmic methods.
Money laundering directly harms American citizens and global interests. Half of North Korea’s nuclear program is funded by laundered funds, according to statements by the White House1, while a federal indictment alleges that money launderers tied to Chinese underground banking are a primary source of financial services for Mexico’s Sinaloa cartel 2.
Despite recent anti-money laundering efforts, the United States (U.S.) still faces challenges in countering money laundering effectively for several reasons. According to Congressional research, money laundering schemes often evade detection and disruption, as anti-money laundering (AML) efforts today rely on manual analysis of large amounts of data and are limited by finite resources and human cognitive processing speed3.