Tech
Interrupting encoder training in diffusion models enables more efficient generative AI
A new framework for generative diffusion models was developed by researchers at Science Tokyo, significantly improving generative AI models. The method reinterpreted Schrödinger bridge models as variational autoencoders with infinitely many latent variables, reducing computational costs and preventing overfitting. By appropriately interrupting the training of the encoder, this approach enabled development of more efficient generative AI, with broad applicability beyond standard diffusion models.
Diffusion models are among the most widely used approaches in generative AI for creating images and audio. These models generate new data by gradually adding noise (noising) to real samples and then learning how to reverse that process (denoising) back into realistic data. A widely used version, the score-based model, achieves this by the diffusion process connecting the prior to the data with a sufficiently long-time interval. This method, however, has a limitation that when the data differs strongly from the prior, the time intervals of the noising and denoising processes become longer, which causes slowing down sample generation.
Now, a research team from Institute of Science Tokyo (Science Tokyo), Japan, has proposed a new framework for diffusion models that is faster and computationally less demanding. They achieved this by reinterpreting Schrödinger bridge (SB) models, a type of diffusion model, as variational autoencoders (VAEs).
The study was led by graduate student Mr. Kentaro Kaba and Professor Masayuki Ohzeki from the Department of Physics at Science Tokyo, in collaboration with Mr. Reo Shimizu (then a graduate student) and Associate Professor Yuki Sugiyama from the Graduate School of Information Sciences at Tohoku University, Japan. Their findings were published in the Physical Review Research on September 3, 2025.
SB models offer greater flexibility than standard score-based models because they can connect any two probability distributions over a finite time using a stochastic differential equation (SDE). This supports more complex noising processes and higher-quality sample generation. The trade-off, however, is that SB models are mathematically complex and expensive to train.
The proposed method addresses this by reformulating SB models as VAEs with multiple latent variables. “The key insight lies in extending the number of latent variables from one to infinity, leveraging the data-processing inequality. This perspective enables us to interpret SB-type models within the framework of VAEs,” says Kaba.
In this setup, the encoder represents the forward process that maps real data onto a noisy latent space, while the decoder reverses the process to reconstruct realistic samples, and both processes are modeled as SDEs learned by neural networks.
The model employs a training objective with two components. The first is the prior loss, which ensures that the encoder correctly maps the data distribution to the prior distribution. The second is drift matching, which trains the decoder to mimic the dynamics of the reverse encoder process. Moreover, once the prior loss stabilizes, encoder training can be stopped early. This allows us to complete learning faster, reducing the risk of overfitting and preserving high accuracy in SB models.
“The objective function is composed of the prior loss and drift matching parts, which characterizes the training of neural networks in the encoder and the decoder, respectively. Together, they reduce the computational cost of training SB-type models. It was demonstrated that interrupting the training of the encoder mitigated the challenge of overfitting,” explains Ohzeki.
This approach is flexible and can be applied to other probabilistic rule sets, even non-Markov processes, making it a broadly applicable training scheme.
More information:
Kentaro Kaba et al, Schrödinger bridge-type diffusion models as an extension of variational autoencoders, Physical Review Research (2025). DOI: 10.1103/dxp7-4hby
Citation:
Interrupting encoder training in diffusion models enables more efficient generative AI (2025, September 29)
retrieved 29 September 2025
from https://techxplore.com/news/2025-09-encoder-diffusion-enables-efficient-generative.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no
part may be reproduced without the written permission. The content is provided for information purposes only.
Tech
Hyperscaler datacentres set to dominate by 2031 | Computer Weekly
Immense concentration continues apace in the cloud industry, with hyperscalers expected to comprise 67% of global datacentre capacity by 2031, or 14 times the capacity they had in 2018. Back then, enterprise datacentres accounted for 56% of all datacentre capacity.
That’s according to figures from US-headquartered research organisation Synergy Research Group, which says artificial intelligence (AI) is driving huge and accelerated growth, with hyperscaler capacity expected to double in the next three years.
By the fourth quarter of 2025, Synergy found that hyperscaler-operated datacentres accounted for 1,360 of total sites and 48% of worldwide capacity. Datacentres built by hyperscalers form the bulk of that capacity – 60% of it – with the remaining capacity leased.
Non-hyperscale colocation capacity accounts for 20% of current totals, while enterprise datacentres account for 32%.
Synergy expects hyperscaler datacentre capacity to comprise 67% of all capacity in 2031. The share of colocation is expected to drop, although it is still increasing at double-digit rates.
Enterprises’ on-premise datacentre capacity is expected to drop to 19% of the total by 2031, at a rate of about 2% per year, although even here that decline is not so rapid, largely due to the deployment of AI hardware.
Synergy’s data is based on several quarterly tracking research services in hyperscale, colocation and enterprise datacentres, and based on datacentre footprint and operations of the world’s major cloud colocation firms, plus tracking the datacentre hardware market.
John Dinsdale, a chief analyst at Synergy Research Group, said AI is driving the world’s datacentre market towards increased concentration in favour of the hyperscalers.
“Cloud and consumer-oriented digital services have been driving changes in datacentre deployment patterns for many years now, but over the last three years, AI technology has accelerated those changes,” he said.
“We are seeing a different mix of datacentre usage across the regions, but overall, the world is racing towards a situation where hyperscale operators are responsible for the bulk of global datacentre capacity. There are almost 800 hyperscale datacentres in our known future pipeline, enabling hyperscale capacity to double in just three years,” Dinsdale added.
By the third quarter of 2025, worldwide spend on cloud services had reached $107bn, up from $68bn two years before that, in 2023.
Among the big three, Amazon’s market share has been in a state of gradual decline since 2022. In the third quarter of 2025, it had a 29% market share, down from just under 34% in the third quarter of 2022.
Meanwhile, the third-quarter 2025 market share for Microsoft was 20%, and 13% for Google Cloud. Both of these are seeing increases in market share, with Microsoft up from 13% in the fourth quarter of 2020.
Meanwhile, so-called neocloud providers – those that specialise in AI datacentre capacity – have a market share of 2.5%.
Dinsdale said: “Beyond the three market giants, a wide mix of smaller players is competing for traction, but the reality is that third-placed Google remains nearly four times the size of fourth-placed Alibaba, underscoring the widening gulf between the market leaders and the rest of the field.”
Tech
XCOM RAN intros end-to-end private 5G for physical AI | Computer Weekly
Looking to boost the adoption of physical artificial intelligence (AI) across several key applications areas for industrial automation, which the company believes will become the new norm, XCOM RAN by Globalstar has announced the launch of an end-to-end private 5G solution.
The company believes that its mission is to provide the next generation of private 5G infrastructure, which is designed to support “tomorrow’s mission-critical industrial automation requirements”. XCOM RAN claims that it is delivering “unprecedented” performance by taking a new approach to private 5G, increasing capacity by more than four times over current private 5G offerings for “flawless” connectivity in the densest automation environments.
XCOM RAN runs on private 5G shared spectrum allocated around the world, and it can use Globalstar’s licensed Band n53 as a dedicated band for “worry-free” private 5G deployments. Its Supercell architecture is designed to reduce the need for site surveys and RF network design, leading to a private 5G solution that “deploys quickly, is easy to manage, and provides full capacity and coverage” in industrial environments.
The company predicted that the amount and types of physical AI optimisations that can be applied will increase exponentially. It noted that its customers are asking for an underlying wireless network architecture that is comprehensive, can adapt and grow with their automation strategies, and can address the needs of customers and for partners.
The launch introduces an orchestration layer for managing private 5G environments, which the company said speaks to the operational complexity enterprises are running into as deployments scale in the AI era.
The company’s offerings include XCOM RAN’s Supercell architecture, based on O-RAN standards, with XCOM Radio Series with indoor and outdoor options; XCOM Core, which is now offered in addition to private 5G cores from partners; and the XCOM Orchestrator, a multi-tenant management and orchestration system designed to streamline operations and minimise the learning curve for enterprise teams new to private 5G.
XCOM RAN is designed to offer spectrum flexibility with support for Band n48 shared spectrum in the US and Band n78 allocated for private 5G and industrial use in Europe and parts of Asia, while it uses Globalstar Band n53 for licensed, dedicated use. The solution includes the XCOM Industrial Router, an Industry 4.0 CPE device that supports all three spectrum bands, enabling customers to integrate XCOM RAN private 5G into their AI-driven industrial automation environments.
XCOM RAN also works with a set of industry partners to offer a private 5G solution and services that are described as “thoroughly tested, integrated and ready for deployment”. The expanding network of partners is said to be intended to ensure customers benefit from “proven technology, seamless integration” and an end-to-end solution built to scale with their business.
A number of these partners have declared support for the new tech, such as ruggedised industrial solutions provider Zebra Technologies.
“We are at the forefront of adding new technology and spectrum options to our devices to support our customers as they rapidly move toward AI-driven intelligent operations,” said James Poulton, senior vice-president and general manager of enterprise mobile computing at Zebra Technologies.
“We have recently added support for Globalstar Band n53 to our ET 401 Enterprise tablets, giving our customers the opportunity to securely run their most sensitive applications over private, dedicated spectrum on these devices.”
Michiel Lotter, CEO of smart signal booster manufacturer Nextivity, added: “One of the latest trends in enterprise wireless deployments is combining modern DAS systems with private 5G to deliver pervasive indoor and outdoor capacity and coverage.
“These solutions are on the cutting edge of development, and we’re grateful to have a partner like the XCOM RAN team who is working with us to address our customers’ requirements.”
Tech
Men Are Buying Hacking Tools to Use Against Their Wives and Friends
Thousands of men are members of Telegram groups and channels that advertise and sell hacking and surveillance services that can be used to harass friends, wives and girlfriends, and former partners, new research has uncovered. The findings, from a European nonprofit group, also say that the communities are involved in extensive trading, selling, and promotion of a huge variety of abusive content, including nonconsensual intimate images of women, so-called nudifying services, plus folders of images that sellers claim include child sexual abuse material and depictions of incest and rape.
Over six weeks earlier this year, researchers at the algorithmic auditing group AI Forensics analyzed nearly 2.8 million messages sent across 16 Italian and Spanish Telegram communities that are regularly posting abusive content targeting women and girls. More than 24,000 members of the Telegram groups and channels took part in posting 82,723 images, videos, and audio files over the course of the study, the analysis says. Many posts target celebrities and influencers, but men in the groups also frequently victimize women they know.
“We tend to forget that most victims are ordinary women who sometimes don’t even know that their pictures are shared or manipulated in these types of channels,” says Silvia Semenzin, a researcher at AI Forensics who previously exposed Italian Telegram channels engaging in similar behavior as far back as 2019. “The majority of this violence is directed towards people who the perpetrators know,” she says, suggesting that Telegram, which has over 1 billion monthly active users, according to company founder Pavel Durov, should be subject to stricter regulation and classed as a “very large online platform” under Europe’s online safety rules.
The findings come as Durov is fighting back against Russia’s efforts to block the messaging app in that country, which has long positioned itself as a messaging app that allows free speech but has simultaneously been used by some to share terrorist, sexual abuse, and cybercrime materials. Durov is under criminal investigation in France relating to alleged criminal activity taking place on Telegram, although he has consistently denied the allegations.
A Telegram spokesperson tells WIRED that the company removes “millions” of pieces of content per day using “custom AI tools” and has policies in Europe that do not allow the promotion of violence, illegal sexual content including nonconsensual imagery, and other content such as doxing and selling illegal goods and services.
Among the extensive types of abusive content and services observed by the AI Forensics researchers were frequent references to the access, publishing, and doxing of women’s private information, sharing their Instagram or TikTok content, as well as references to spying or hacking. “Victims are often named, tagged, and locatable via shared profile links,” the group’s report says.
One translated post on Telegram titled “Professional hacking on commission” claimed to be able to give customers “access to phone gallery and extraction of photos and videos,” as well as “anonymous social media hacking.” Another message says: “I hack and recover any type of social media service. I can spy on your partner’s account. Send me a private message.”
Across the dataset there were more than 18,000 references to spying or spy content. One post reads: “Hi, do you have the desire to spy on a girl’s gallery? We sell a bot that does it for info DM.” Meanwhile, users were observed asking if people could find phone numbers connected to Instagram accounts and other requests, “who exchanges spy photos and videos?”
-
Uncategorized6 days ago
[CinePlex360] Please moderate: “Trump signals p
-
Entertainment5 days agoJoe Jonas shares candid glimpse into parenthood with Sophie Turner
-
Tech5 days agoOur Favorite iPad Is $50 Off
-
Fashion1 week agoChina’s Anta Sports posts record $11.62 bn revenue in 2025
-
Politics5 days agoIran can sustain Strait of Hormuz closure for years, will cut US military logistics: Official
-
Business6 days agoJaguar Land Rover sees sales recover after cyber attack
-
Sports5 days agoUConn Final Four run could trigger a $50M furniture giveaway for Massachusetts-based Jordan’s Furniture
-
Politics1 week agoTrump considers asking Arab allies to help to pay for Iran war
