Toggle light / dark theme

Reinforcement learning (RL) has become central to advancing Large Language Models (LLMs), empowering them with improved reasoning capabilities necessary for complex tasks. However, the research community faces considerable challenges in reproducing state-of-the-art RL techniques due to incomplete disclosure of key training details by major industry players. This opacity has limited the progress of broader scientific efforts and collaborative research.

Researchers from ByteDance, Tsinghua University, and the University of Hong Kong recently introduced DAPO (Dynamic Sampling Policy Optimization), an open-source large-scale reinforcement learning system designed for enhancing the reasoning abilities of Large Language Models. The DAPO system seeks to bridge the gap in reproducibility by openly sharing all algorithmic details, training procedures, and datasets. Built upon the verl framework, DAPO includes training codes and a thoroughly prepared dataset called DAPO-Math-17K, specifically designed for mathematical reasoning tasks.

DAPO’s technical foundation includes four core innovations aimed at resolving key challenges in reinforcement learning. The first, “Clip-Higher,” addresses the issue of entropy collapse, a situation where models prematurely settle into limited exploration patterns. By carefully managing the clipping ratio in policy updates, this technique encourages greater diversity in model outputs. “Dynamic Sampling” counters inefficiencies in training by dynamically filtering samples based on their usefulness, thus ensuring a more consistent gradient signal. The “Token-level Policy Gradient Loss” offers a refined loss calculation method, emphasizing token-level rather than sample-level adjustments to better accommodate varying lengths of reasoning sequences. Lastly, “Overlong Reward Shaping” introduces a controlled penalty for excessively long responses, gently guiding models toward concise and efficient reasoning.

Researchers at the University of Gothenburg have developed a novel Ising machine that utilizes surface acoustic waves as an effective carrier of dense information flow. This approach enables fast, energy-efficient solutions to complex optimization problems, offering a promising alternative to conventional computing methods based on von-Neumann architecture. The findings are published in the journal Communications Physics.

Traditional computers can stumble when tackling —tasks of scheduling logistic operations, financial portfolio optimization and high frequency trading, optimizing communication channels in complex wireless networks, or predicting how proteins fold among countless structural possibilities.

In these cases, each added node—an additional logistic hub, network user, or molecular bond causes the number of possible configurations to explode exponentially. In contrast to linear or polynomial growth, an exponential increase in the number of possible solutions makes even the most powerful computers and algorithms lack the computational power and memory to evaluate every scenario in search of vanishingly small subsets representing satisfactorily optimal solutions.

The use of artificial intelligence (AI) scares many people as neural networks, modeled after the human brain, are so complex that even experts do not understand them. However, the risk to society of applying opaque algorithms varies depending on the application.

While AI can cause great damage in democratic elections through the manipulation of social media, in astrophysics it at worst leads to an incorrect view of the cosmos, says Dr. Jonas Glombitza from the Erlangen Center for Astroparticle Physics (ECAP) at Friedrich-Alexander Universität Erlangen-Nürnberg (FAU).

The astrophysicist uses AI to accelerate the analysis of data from an observatory that researches cosmic radiation.

A new study probing quantum phenomena in neurons as they transmit messages in the brain could provide fresh insight into how our brains function.

In this project, described in the Computational and Structural Biotechnology Journal, theoretical physicist Partha Ghose from the Tagore Centre for Natural Sciences and Philosophy in India, together with theoretical neuroscientist Dimitris Pinotsis from City St George’s, University of London and the MillerLab of MIT, proved that established equations describing the classical physics of brain responses are mathematically equivalent to equations describing quantum mechanics. Ghose and Pinotsis then derived a Schrödinger-like equation specifically for neurons.

Our brains process information via a vast network containing many millions of neurons, which can each send and receive chemical and electrical signals. Information is transmitted by nerve impulses that pass from one neuron to the next, thanks to a flow of ions across the neuron’s cell membrane. This results in an experimentally detectable change in electrical potential difference across the membrane known as the “action potential” or “spike”

Based on how an AI model transcribes audio into text, the researchers behind the study could map brain activity that takes place during conversation more accurately than traditional models that encode specific features of language structure — such as phonemes (the simple sounds that make up words) and parts of speech (such as nouns, verbs and adjectives).

The model used in the study, called Whisper, instead takes audio files and their text transcripts, which are used as training data to map the audio to the text. It then uses the statistics of that mapping to “learn” to predict text from new audio files that it hasn’t previously heard.

The future of AI is here—and it’s running on human brain cells! In a groundbreaking development, scientists have created the first AI system powered by biological neurons, blurring the line between technology and biology. But what does this mean for the future of artificial intelligence, and how does it work?

This revolutionary AI, known as “Brainoware,” uses lab-grown human brain cells to perform complex tasks like speech recognition and decision-making. By combining the adaptability of biological neurons with the precision of AI algorithms, researchers have unlocked a new frontier in computing. But with this innovation comes ethical questions and concerns about the implications of merging human biology with machines.

In this video, we’ll explore how Brainoware works, its potential applications, and the challenges it faces. Could this be the key to creating truly intelligent machines? Or does it raise red flags about the ethical boundaries of AI research?

What is Brainoware, and how does it work? What are the benefits and risks of AI powered by human brain cells? How will this technology shape the future of AI? This video answers all these questions and more. Don’t miss the full story—watch until the end!

#ai.
#artificialintelligence.
#ainews.

******************

Compare news coverage. Spot media bias. Avoid algorithms. Try Ground News today and get 40% off your subscription by going to https://ground.news/upandatom.

Special thanks to Chuankun Zhang, Tian Ooi, Jacob S. Higgins, and Jack F. Doyle from Prof. Jun Ye’s lab at JILA/NIST/University of Colorado, as well as Prof. Victor Flambaum from UNSW’s Department of Theoretical Physics, for their valuable assistance and consultation on this video.

Hi! I’m Jade. If you’d like to consider supporting Up and Atom, head over to my Patreon page smile
https://www.patreon.com/upandatom.

Visit the Up and Atom store.
https://store.nebula.app/collections/up-and-atom.

https://www.youtube.com/c/upandatom.

For a one time donation, head over to my PayPal smile https://www.paypal.me/upandatomshows.

Global optimization-based approaches such as basin hopping28,29,30,31, evolutionary algorithms32 and random structure search33 offer principled approaches to comprehensively navigating the ambiguity of active phase. However, these methods usually rely on skillful parameter adjustments and predefined conditions, and face challenges in exploring the entire configuration space and dealing with amorphous structures. The graph theory-based algorithms34,35,36,37, which can enumerate configurations for a specific adsorbate coverage on the surface with graph isomorphism algorithms, even on an asymmetric one. Nevertheless, these methods can only study the adsorbate coverage effect on the surface because the graph representation is insensitive to three-dimensional information, making it unable to consider subsurface and bulk structure sampling. Other geometric-based methods38,39 also have been developed for determining surface adsorption sites but still face difficulties when dealing with non-uniform materials or embedding sites in subsurface.

Topology, independent of metrics or coordinates, presents a novel approach that could potentially offer a comprehensive traversal of structural complexity. Persistent homology, an emerging technique in the field of topological data analysis, bridges the topology and real geometry by capturing geometric structures over various spatial scales through filtration and persistence40. Through embedding geometric information into topological invariants, which are the properties of topological spaces that remain unchanged under specific continuous deformations, it allows the monitoring of the “birth,” “death,” and “persistence” of isolated components, loops, and cavities across all geometric scales using topological measurements. Topological persistence is usually represented by persistent barcodes, where different horizontal line segments or bars denote homology generators41. Persistent homology has been successfully employed to the feature representation for machine learning42,43, molecular science44,45, materials science46,47,48,49,50,51,52,53,54,55, and computational biology56,57. The successful application motivates us to explore its potential as a sampling algorithm due to its capability of characterizing material structures multidimensionally.

In this work, we introduce a topology-based automatic active phase exploration framework, enabling the thorough configuration sampling and efficient computation via MLFF. The core of this framework is a sampling algorithm (PH-SA) in which the persistent homology analysis is leveraged to detect the possible adsorption/embedding sites in space via a bottom-up approach. The PH-SA enables the exploration of interactions between surface, subsurface and even bulk phases with active species, without being limited by morphology and thus can be applied to periodical and amorphous structures. MLFF are then trained through transfer learning to enable rapid structural optimization of sampled configurations. Based on the energetic information, Pourbaix diagram is constructed to describe the response of active phase to external environmental conditions. We validated the effectiveness of the framework with two examples: the formation of Pd hydrides with slab models and the oxidation of Pt clusters in electrochemical conditions. The structure evolution process of these two systems was elucidated by screening 50,000 and 100,000 possible configurations, respectively. The predicted phase diagrams with varying external potentials and their intricate roles in shaping the mechanisms of CO2 electroreduction and oxygen reduction reaction were discussed, demonstrating close alignment with experimental observations. Our algorithm can be easily applied to other heterogeneous catalytic structures of interest and pave the way for the realization of automatic active phase analysis under realistic conditions.

The electrically readable complex dynamics of robust and scalable magnetic tunnel junctions (MTJs) offer promising opportunities for advancing neuromorphic computing. In this work, we present an MTJ design with a free layer and two polarizers capable of computing the sigmoidal activation function and its gradient at the device level. This design enables both feedforward and backpropagation computations within a single device, extending neuromorphic computing frameworks previously explored in the literature by introducing the ability to perform backpropagation directly in hardware. Our algorithm implementation reveals two key findings: (i) the small discrepancies between the MTJ-generated curves and the exact software-generated curves have a negligible impact on the performance of the backpropagation algorithm, (ii) the device implementation is highly robust to inter-device variation and noise, and (iii) the proposed method effectively supports transfer learning and knowledge distillation. To demonstrate this, we evaluated the performance of an edge computing network using weights from a software-trained model implemented with our MTJ design. The results show a minimal loss of accuracy of only 0.4% for the Fashion MNIST dataset and 1.7% for the CIFAR-100 dataset compared to the original software implementation. These results highlight the potential of our MTJ design for compact, hardware-based neural networks in edge computing applications, particularly for transfer learning.

Quantum systems hold the promise of tackling some complex problems faster and more efficiently than classical computers. Despite their potential, so far only a limited number of studies have conclusively demonstrated that quantum computers can outperform classical computers on specific tasks. Most of these studies focused on tasks that involve advanced computations, simulations or optimization, which can be difficult for non-experts to grasp.

Researchers at the University of Oxford and the University of Sevilla recently demonstrated a over a classical scenario on a cooperation task called the odd-cycle game. Their paper, published in Physical Review Letters, shows that a team with can win this game more often than a team without.

“There is a lot of talk about quantum advantage and how will revolutionize entire industries, but if you look closely, in many cases, there is no mathematical proof that classical methods definitely cannot find solutions as efficiently as quantum algorithms,” Peter Drmota, first author of the paper, told Phys.org.