HSC J1631+4426 broke the record for the lowest oxygen abundance. Credit: NAOJ/Kojima et al.

Breaking the lowest oxygen abundance record.

New results achieved by combining big data captured by the Subaru Telescope and the power of machine learning have discovered a galaxy with an extremely low oxygen abundance of 1.6% solar abundance, breaking the previous record of the lowest oxygen abundance. The measured oxygen abundance suggests that most of the stars in this galaxy formed very recently.

To understand galaxy evolution, astronomers need to study galaxies in various stages of formation and evolution. Most of the galaxies in the modern Universe are mature galaxies, but standard cosmology predicts that there may still be a few galaxies in the early formation stage in the modern Universe. Because these early-stage galaxies are rare, an international research team searched for them in wide-field imaging data taken with the Subaru Telescope. To find the very faint, rare galaxies, deep, wide-field data taken with the Subaru Telescope was indispensable, emphasizes Dr. Takashi Kojima, the leader of the team.

However, it was difficult to find galaxies in the early stage of galaxy formation from the data because the wide-field data includes as many as 40 million objects. So the research team developed a new machine learning method to find such galaxies from the vast amount of data. They had a computer repeatedly learn the galaxy colors expected from theoretical models, and then let the computer select only galaxies in the early stage of galaxy formation.

The research team then performed follow-up observations to determine the elemental abundance ratios of 4 of the 27 candidates selected by the computer. They have found that one galaxy (HSC J1631+4426), located 430 million light-years away in the constellation Hercules, has an oxygen abundance only 1.6 percent of that of the Sun. This is the lowest values ever reported for a galaxy. The measured oxygen abundance suggests that most of the stars in this galaxy formed very recently. In other words, this galaxy is undergoing an early stage of the galaxy evolution.

What is surprising is that the stellar mass of the HSC J1631+4426 galaxy is very small, 0.8 million solar masses. This stellar mass is only about 1/100,000 of our Milky Way galaxy, and comparable to the mass of a star cluster in our Milky Way, said Prof. Ouchi of the National Astronomical Observatory of Japan and the University of Tokyo. This small mass also supports the primordial nature of the HSC J1631+4426 galaxy.

The research team thinks that there are two interesting indications from this discovery. First, this is the evidence about a galaxy at such an early stage of galaxy evolution existing today. In the framework of the standard cosmology, new galaxies are thought to be born in the present universe. The discovery of the HSC J1631+4426 galaxy backs up the picture of the standard cosmology. Second, we may witness a new-born galaxy at the latest epoch of the cosmic history. The standard cosmology suggests that the matter density of the universe rapidly drops in our universe whose expansion accelerates. In the future universe with the rapid expansion, matter does not assemble by gravity, and new galaxies wont be born. The HSC J1631+4426 galaxy may be the last generation galaxy in the long cosmic history.

Read this article:
Surprisingly Recent Galaxy Discovered Using Machine Learning May Be the Last Generation Galaxy in the Long Cosmic History - SciTechDaily

Who you are together is more important than who you are alone.

What makes us happy in a romantic relationship? The question might seem too complex to answer, too varied couple to couple. But a new study in the Proceedings of the National Academy of Sciences attempts to answer just that - using machine learning.

Previous studies on romantic satisfaction were limited in size. By using machine learning, however, researchers were able to analyze a massive amount of data, which included over 11,000 different couples from 43 data sets. Individual studies are many times limited - it is difficult and expensive to recruit couples for the studies. Its also exhausting for the participants. Using machine learning to analyze a large amount of data from pre-existing studies bypasses these problems.

The researchers looked at variables that could predict happiness within a relationship. Some of these, such as neuroticism, political orientation, conscientiousness or family history were qualities of the individuals involved. Others, such as appreciation, affection and perceived partner commitment were qualities of the relationship.

Of these, qualities of the relationship, rather than the individuals involved, contributed more to overall satisfaction. The five most important were how much they believed their partner was committed to the relationship, how much they appreciated their partner, sexual satisfaction, how much they believed their partner was happy in the relationship, and not fighting often.

Appreciation and commitment are key for a fulfilling relationship.

Qualities of the individuals contribute too - but not as much. In fact, 45% of the variability in a relationship is due to the qualities of the relationship. 21% were due to the individuals themselves. In addition, once qualities of the relationship were taken into account, the differences due to the individuals were not as important.

Experiencing negative affect, depression, or insecure attachment are surely relationship risk factors. But if people nevertheless manage to establish a relationship characterized by appreciation, sexual satisfaction, and a lack of conflictand they perceive their partner to be committed and responsivethose individual risk factors may matter little, say the authors.

In other words, for a happy relationship, its more important who you are together than who you are apart.

Originally posted here:
Machine Learning Reveals What Makes People Happy In A Relationship - Forbes

(MENAFN - EIN Presswire)

Activeloop user interface and toolset work with NVIDIA processing to help InteinAir achieve great ML results

Activeloop.ai logo

Y Combinator alum achieves better aerial data pipelines for IntelinAir in an industry-leading Agriculture Tech solution

MOUNTAIN VIEW, CA, USA, August 4, 2020 / EINPresswire.com / -- In a case study now available online, Activeloop ( [To enable links contact MENAFN] ), a Y Combinator-backed startup, is announcing a major success in helping an early customer, IntelinAir , improve the efficiency of their AI analysis of aerial footage. Activeloop's software builds plug-and-play data pipelines for unstructured data. The software helps data scientists streamline their data aggregation and preparation, and automates and optimizes their training of machine learning models. Together with NVIDIA , Activeloop has achieved a massive reduction in the time-to-value and cost of machine learning / deep learning efforts. The case study documents a breakthrough in the field of aerial imagery with their joint customer IntelinAir, a leading crop intelligence firm.

Activeloop's solution is becoming available just in time for the exploding artificial intelligence and advanced machine learning market, projected to grow up to $281.24 billion by 2026 with CAGR of 37.95%. This coincides with the massive growth of data available to be analyzed by AI. All data generated by the end of 2020 will be about 40 trillion gigabytes (40 zettabytes), with IBM estimating that 90% of it has been created over the past 2 years. As data gets bigger faster than ever, translating it into actionable insights is becoming increasingly difficult and expensive. As a result, the effort needed to set up a new model and get it running efficiently can be beyond the reach of many teams who could otherwise benefit from machine learning. Existing solutions often have large cloud storage and processing costs. These solutions can't be made more efficient without radical changes.

'Unstructured data - including text, images, or videos, comprises about 80-90% of the data people generate today', says Davit Buniatyan, Activeloop Founder and CEO. 'As it comes in different forms, sizes, and even shapes, analyzing and managing it is an extremely difficult and costly task. In fact, data scientists spend about 50 to 80% of their time setting up their unstructured dataset rather than analyzing it via machine or deep learning. We're changing that by creating a fast, simple platform for building and scaling data pipelines for machine learning.'

'We operate in an agile fashion: we want to focus on building high-quality models instead of fighting with data pipelines, infrastructure, and deployment challenges' says Jennifer Hobbs, Director of Machine Learning at IntelinAir. 'Thanks to Activeloop, we've been able to deploy new models in a matter of days instead of weeks. With the help of Activeloop's platform and NVIDIA's powerful GPUs, we were able to increase the inference speed threefold and improve the accuracy of the trained models at half the cost."

You can read more about the success story here: [To enable links contact MENAFN] .

###

About Activeloop

Activeloop ( [To enable links contact MENAFN] ), is a startup backed by Y Combinator and prominent Silicon Valley investors. The company has already been featured by major outlets including TechCrunch and is now coming out of stealth mode to make its product available to the machine learning community. Formerly named Snark AI, Activeloop aims to optimize the way machine and deep learning models are trained and streamline the huge amounts of data required for this work. Activeloop is a member of NVIDIA's Inception program for AI/ML development.

About IntelinAir

IntelinAir ( [To enable links contact MENAFN] ) is a full-season and full-spectrum crop intelligence company focused on agriculture that delivers actionable intelligence to help farmers make data-driven decisions to improve operational efficiency, yields, and ultimately their profitability.

Mikayel HarutyunyanActiveloop.ai+1 415-876-5667email us here Visit us on social media:Facebook Twitter LinkedIn

Activeloop introduction and demo

Follow this link:
Preparing new machine learning models used to take weeks Activeloop teams up with NVIDIA to reduce that time to hours - MENAFN.COM

Machine learning typically is used to solve a host of diverse problems within an organization, extracting predictive knowledge from both structured and unstructured data and using them to deliver value. The technology has already made its way into different aspects of a business ranging from finding data patterns to detect anomalies and making recommendations. Machine learning helps organizations gain a competitive edge by processing a voluminous amount of data and applying complex computations.

With machine learning, companies can develop better applications according to their business requirements. This technology is mainly designed to make everything programmatic. Applications of ML have the potential to drive business outcomes that can extensively affect a companys bottom line. The rapid evolution of new techniques in recent years has further expanded the machine learning application to nearly boundless possibilities.

Industries relying on massive volumes of data are significantly leveraging ML techniques to process their data and to build models, strategize, and plan.

While implementing the effective application of machine learning enables businesses to grow, gain competitive advantage and prepare for the future, there are some key practical issues in ML and their business implications organizations must consider.

As machine learning significantly relies on data, the occurrence of noisy data can considerably impact any information prediction. Generally, data from a dataset carries extraneous and meaningless information which can significantly affect data analysis, clustering and association analysis. Having a lack of quality data can also restrain the capabilities of building ML models. In order to cope with quality data and noise, businesses need to apply better and effective machine learning strategies through data cleansing and overall processing of data.

There is no doubt that the development of machine learning has made it possible to learn directly from data rather than human knowledge with a strong emphasis on accuracy. However, the lack of the ability to explain or present data in understandable terms to a human, often called interpretability, is one of the biggest issues in machine learning. The introduction of possible biases in data has also led to ethical and legal issues with ML models. Theinterpretabilitylevels in the field of machine learning and algorithms may significantly vary. Some methods are human-compatible as they are highly interpretable, while some are too complex to apprehend, thus require ad-hoc methods to gain an interpretation.

In the context of supervised machine learning, an imbalanced dataset often involves two or more classes. There is an imbalance among labels in the training data in several real-world datasets. This imbalance in a dataset has the potential to affect the choice of learning, the process of selecting algorithms, model evaluation and verification. The models can even suffer large biases, and the learning will not be effective if the right techniques are not employed properly. ML algorithms can generate insufficient classifiers when faced with imbalanced datasets. When trying to resolve certain business challenges with imbalanced data sets, the classifiers produced by standard ML algorithms might not deliver precise outcomes.

Thus, to address imbalanced datasets requires strategies like enhancing classification algorithms or balancing classes in the training data before providing the data as input to machine learning algorithms.

Share This ArticleDo the sharing thingy

About AuthorMore info about author

Originally posted here:
Decoding Practical Problems and Business Implications of Machine Learning - Analytics Insight

WASHINGTON - New software developed by BAE Systems could help the Department of Defense build confidence in decisions and intelligence produced by machine learning algorithms, the company claims.

BAE Systems said it recently delivered its new MindfuL software program to the Defense Advanced Research Projects Agency in a July 14 announcement. Developed in collaboration with the Massachusetts Institute of Technologys Computer Science and Artificial Intelligence Laboratory, the software is designed to increase transparency in machine learning systemsartificial intelligence algorithms that learn and change over time as they are fed ever more databy auditing them to provide insights about how it reached its decisions.

The technology that underpins machine learning and artificial intelligence applications is rapidly advancing, and now its time to ensure these systems can be integrated, utilized, and ultimately trusted in the field, said Chris Eisenbies, product line director of the cmpanys Autonomy, Control, and Estimation group. The MindfuL system stores relevant data in order to compare the current environment to past experiences and deliver findings that are easy to understand.

While machine learning algorithms show promise for DoD systems, determining how much users can trust their output remains a challenge. Intelligence officials have repeatedly noted that analysts cannot rely on black box artificial intelligence systems that simply produce a decision or piece of intelligencethey need to understand how the system came to that decision and what unseen biases (in the training data or otherwise) might be influencing that decision.

MindfuL is designed to help address that gap by providing more context around those outputs. For instance, the company says its program will issue statements such as The machine learning system has navigated obstacles in sunny, dry environments 1,000 times and completed the task with greater than 99 percent accuracy under similar conditions; or The machine learning system has only navigated obstacles in rain 100 times with 80 percent accuracy in similar conditions; manual override recommended. Those types of statements can help users evaluate how much confidence they should place in any individual decision produced by the system.

This is the first release of the MindfuL software as part of a $5 million, three-year contract under DARPAs Competency-Aware Machine Learning (CAML) program. BAE Systems plans to demonstrate their software in both simulation and in prototype hardware later this year.

Continue reading here:
Could this software help users trust machine learning decisions? - C4ISRNet