The role of scientific research in pushing the frontiers of artificial intelligence cannot be overstated. The researchers working at MITs Computer Science and Artificial Intelligence Laboratory, Stanford Artificial Intelligence Laboratory, Oxford University and many other top labs are shaping the future of humanity. In addition, most top AI labs, even the private players such as DeepMind and OpenAI, publish on preprint servers to democratise and share knowledge.

But, how useful are these papers for the community at large?

Recently, a Reddit user published a post titled, I dont really trust papers out of Top Labs anymore. In the post, the user asked: Why should the AI community trust these papers published by a handful of corporations and the occasional universities? Why should I trust that your ideas are even any good? I cant check them; I cant apply them to my own projects.

Citing the research paper titled An Evolutionary Approach to Dynamic Introduction of Tasks in Large-scale Multitask Learning Systems, the Reddit user said, Its 18 pages of talking through this pretty convoluted evolutionary and multitask learning algorithm; its pretty interesting, solves a bunch of problems. But two notes. One, the big number they cite as the success metric is 99.43 on CIFAR-10, against a SotA of 99.40.

The Reddit user also referred to a chart towards the end of the paper that details how many TPU core-hours were used for just the training regimens that resulted in the final results.

The total is 17,810 core-hours. Lets assume that for someone who doesnt work at Google, youd have to use on-demand pricing of USD3.22 per hour. This means that these trained models cost USD57,348.

Strictly speaking, throwing enough compute at a general enough genetic algorithm will eventually produce arbitrarily good performance, so while you can read this paper and collect interesting ideas about how to use genetic algorithms to accomplish multitask learning by having each new task leverage learned weights from previous tasks by defining modifications to a subset of components of a pre-existing model, he said.

Jathan Sadowski, a senior fellow at Emerging Tech Lab, responded: AI/ML research at places like Google and OpenAI is based on spending absurd amounts of money, compute, and electricity to brute force arbitrary improvements. The inequality, the trade-offs, the wasteall for incremental progress toward a bad future.

The Reddit post has been a source of much debate on social media. Many pointed out that there should be a new journal for papers where one can replicate their results in under eight hours on a single GPU.

Findings that cant be replicated are intrinsically less reliable. And the fact that the ML community is maturing towards decent scientific practices instead of anecdotes is a positive sign, said Leon Derczynski, assistant professor at IT University of Copenhagen.

The replication crisis has been gripping the scientific community for ages. The AI domain is also grappling with it, mostly because researchers often dont share their source code. A replication crisis refers to when scientific studies are difficult or impossible to reproduce.

According to a 2016 Nature survey, more than 70 percent of researchers have tried and failed to reproduce another scientists experiments. Further, more than 50 percent of them have failed to reproduce their own experiments.

Reproducibility is the basis of quality assurance in science as it enables past findings to be independently verified.

The scientific and research community strongly believes that withholding important aspects of studies, especially in domains where larger public good and societal well-being are concerned, does a great disservice.

According to the 2020 State of AI report, only 15 percent of AI studies share their code, and industry researchers are often the culprits. The report criticises OpenAI and DeepMind, two of the worlds best AI research labs, for not open sourcing their code.

In 2020, Google Health published a paper in Nature that described how AI was leveraged to look for signs of breast cancer in medical images. But Google drew flak as it provided little information about its code and how it was tested. Many questioned the viability of the paper, and a group of 31 researchers published another paper in Nature titled Transparency and reproducibility in artificial intelligence. Benjamin Haibe-Kains, one of the papers authors, called Googles paper an advertisement for cool technology with no practical use.

However, things are changing. NeurIPS now asks authors/researchers to produce a reproducibility checklist along with their submissions. This checklist consists of information such as the number of models trained, computing power used, and links to code and datasets. Another initiative called the Papers with Code project was started with a mission to create free and open-source ML papers, code and evaluation tables.

Visit link:
'I don't really trust papers out of top AI labs anymore' - Analytics India Magazine

As information becomes more and more web-based, so too does the attention of the younger generations.

We rely on the internet to give us everything, from news to research, and it is having a huge impact on what we read and how we read it.

We have moved away from the deep reading of different texts and journals and moved towards the skim, light reading of web posts, and websites in order to get the information that we want.

This affects the way that our brains process and relay information.Instead of taking in each element of the text we read, we simply find the pieces of information that serve our agendas, rather than developing our brains to help us recall the information later.

When we read deeply, we read much slower; we take in the details of different sensory descriptions and we become much more immersed in the things we read.

Whereas with thelight reading, we read much faster, we look for the pieces of information that we want and we dont really look at everything else around it. This doesnt give us all of the information available from the text and it doesnt exercise our brains as much as deeper reading does.

Deep reading activates the centres in your brain which are responsible for speech, hearing, and vision, and helps them to work together to create an image in our heads.

Reading in this way also develops our ability to perceive and use language, and gives us the greater ability to create more complex sentence structures and fuller descriptions.

When we read deeply, we also take in the information much better than in the case of light reading. The information is stored in the brain when we deep read and is ready to be recalled later on.

Deep reading has also been shown to make us nicer. As we read we can articulate and develop the skill to understand our emotions much better than before, and this helps us to understand emotions in others as well.

Reading things such as poems, novels, and academic reports can massively develop your writing abilities. Make sure to take your time and soak in all of the available information to create a full picture, rather than skim-reading the information.

Putthese types of writings in your life rather than getting information only from online blogs and televisions shows, as they make your brain switch off almost immediately. Even though they are entertaining, they wont develop your writing ability whatsoever.

Unfortunately, in the modern world, reading has become an unpopular pastime, but the benefits of real reading are massive and shouldnt be ignored.

Next time youre writing a paper, a short story, or even just have some free time, take some time to read and read deeply, taking in all the information and really enjoying it.

References:

Contributing writer at Learning Mind

Francesca Forsythe is a professional writer who holds a dual award Master's degree in European Law and Philosophy of Law from Leiden University. She has written for several websites on a range of subjects across lifestyle, relationships, and health & fitness, as well as academic pieces in her fields of study.

Visit link:
Light Reading VS Deep Reading: What You Read Matters ... - Learning Mind

CHICAGO This month, DNAStar formally launched NovaFold AI, an update to its NovaFoldprotein structure prediction software that incorporates the AlphaFold2 artificial intelligence system from Google sister company DeepMind Technologies. The company had introduced a beta version of the product, dubbed NovaFold AI powered by AlphaFold2, in February.

Steve Darnell, team leader of DNAStar's structural biology and protein team, called NovaFold AI a cloud version of the AlphaFold 2 pipeline. The offering lets usersaccess the AlphaFold protein structure prediction method from within DNAStar'sNovaCloud Services interface of the firm's Protean 3D visualizationand analysis suite.

He said that AlphaFold 2 is designed to help researchers without a bioinformatics background predict structures.

"This release allows customers to more tightly integrate the analysis with running the prediction," added DNAStar General Manager Shawn Grass.

Madison, Wisconsin-based DNAStar sees NovaFold AI as a growth opportunity.

"The structure market has been a little more challenging for us to get into, because when people think of us, they think of just DNA. They don't think of us in the protein area," Grass said. "Our hope is with the excitement and buzz right now about structure prediction that this will give us an opportunity to expand our reach."

DNAStar, which has been in business since 1984, is actually quite active beyond DNA.

Last year, the privately held firm introduced version 17.3 of its Lasergene software for DNA, RNA, and protein sequencing assembly and analysis. That release added viral genome analysis workflow support for Pacific Biosciences and Oxford Nanopore long-read sequences, including data from PCR-amplified fragments generated according to Artic network protocols. The update was also optimized forgenomic analysis and variant identification to support COVID-19 research.

Lasergene, which serves the molecular biology and genomics markets, is one DNAStar's two product lines. It includes a series of applications for tasks like genome visualization, sequence assembly, sequence alignment, and in silico cloning. DNAStar's molecular visualization and analysis platform, Protean 3D, is technically part of Lasergene as well.

"Protean 3D has a tight integration between its sequence representations and its structural representations through a selection model," Darnell explained.

Lasergene products are "general tools" aimed at anyone in molecular biology or genomics looking to manage pipelines and workflows, according to Grass. He said that a fair number of customers mistakenly refer to the company as Lasergene because the product name is well known in some circles.

The other DNAStar product line, for structural biology applications including structure prediction, docking, and antibody modeling, is called Nova.

NovaFold is the commercialization of I-TASSER, protein structure prediction software developed by University of Michigan bioinformatician Yang Zhang. In 2015, DNAStar licensed the exclusive commercial rights to I-TASSER and markets the technology as part of Protean 3D.

The firm also licenses technology for its NovaDock protein-protein docking software from Cancer Research UK. Darnell said that the company's products represent a mixture of DNAStar-developed tools, freely available external tools like AlphaSeek, and licensed technology.

DNAStar is not directly partnering with DeepMind, which makes some of its AlphaFold technology available through open-source channels.

Grass said that DNAStar has not made any additions or other changes to the AI that AlphaFold provides. "It's something we are always looking at, but right now, we felt it was best to leave it just how it was," he said.

What the company does offer with the new release is the ability to add AlphaFold structure predictions as templates to the core NovaFold software to guide the modeling process toward these predictions. "What we also bring to the table is the additional downstream analysis outside of just coordinate generation," according to Darnell.

Most of the tools in both product lines are integrated on DNAStar's NovaCloud infrastructure, which the firm has been building up for the last five or six years.

DeepMind's AlphaFold had the highest score in the 14th and most recent Critical Assessment of Structure Prediction (CASP) competition in 2020, in which entrants are given the amino acid sequences for about 100 proteins to then predict their structures.

"It certainly excels wonderfully at fold recognition and modeling," Darnell said. He added that AlphaFold will be entered in CASP 15 this summer, a competition that will have a greater focus on predicting multimers.

That is a direction DNAStar wants to go in with NovaFold 2.

"The direct structure prediction of multimers is definitely where I believe the field is certainly looking forward toward, as well as different and better approaches for doing multi-domain proteins," Darnell said. "A tool like AlphaFold can identify and actually model different domains in isolation within a multi-domain sequence."

Darnell also expects to see more activity around protein-protein docking technology in subsequent years.

Read more from the original source:
DNAStar Beefs up Protein Structure Prediction With DeepMind-Powered NovaFold AI - GenomeWeb

It is with great skill that Bergner places Carolines story in context of the history of modern psychiatry. Its hard to do justice to the sweep of the larger story he tells, but probably the most shocking part is the utter randomness that has characterized so much of the modern search for psycho-pharmaceuticals, combined with the utterly devastating side effects they can have. Bergner tracks the history of treatments like lithium, S.S.R.I.s and antipsychotics. In many cases, researchers only stumbled across the drugs potential to ameliorate symptoms. Of lithium, he writes that 19th-century doctors used it to treat kidney stones. Later it was among the ingredients in 7-Up. Even though lithium was approved by the F.D.A. for psychiatric use in 1970, no one had more than a vague concept of how the drug worked neurologically, Bergner notes, and they still dont.

Bergner interviews a group of researchers who, despite the accidental origins of numerous pharmaceuticals, strive today to develop them into substances that will truly improve peoples lives. This is an interesting set of interviewees, all dedicated, hardworking, highly knowledgeable scientists, who frankly acknowledge how poor the efficacy of many drugs is, how much of a toll they can take on people who use them and how little we know about how the brain actually works.

Bergners subjects, as well as the scientists and clinicians he interviews, also attest to the fuzziness of many diagnostic and behavioral boundaries. Standard diagnoses often collapse what some scientists believe are different conditions into one, whereas other diagnoses wall off conditions that are perhaps not so different at all. Its possible that psychosis, for example, is not really one disorder but dozens of them.

Where the history of drug development has been astonishingly haphazard, and our grasp of brain function is disturbingly low-level, the history of psycho-pharmaceutical marketing has been clever and effective. I still recall when an undergraduate friend confidently told me that her recent bout with depression had resulted from a chemical imbalance in her brain. I was dazzled by the explanation. It made her sadness cleaner, more easily resolved, less unglamorous.

It turns out that we had both signed on to the chemical imbalance theory, which proposed, in the 1960s, that depression could result from a deficiency of neurotransmitters. This ultimately evolved into the idea that too many or too few neurochemicals could cause different kinds of mental illness, such as psychosis. Biology became ascendant in our understanding of psychiatric conditions, which led to a vision of medicalized mental health that one of Bergners scientists calls a house of cards. The idea that S.S.R.I.s, for example, could further our understanding of disorders, the scientist observed, was like saying, I have pain so I must have an aspirin deficiency.

Read the original:
Review: The Mind and the Moon, by Daniel Bergner - The New York Times

Music - 2 hours ago

Ahead of the release of his upcoming album, Mr. Morale & The Big Steppers, Kendrick Lamar has released another installment of his The Heart series. The Heart Part 5 debuted on Sunday; accompanying its release was a pretty surreal music video where Lamar, against a red backdrop, transforms into different Black celebrities as he performs the song. Through the use of deepfake technology (the use of AI to replace the likeness of one person with another in video and other digital media), Kendrick Lamar becomes O.J. Simpson, Kanye West, JussieSmollett, Will Smith, Kobe Bryant, and Nipsey Hussle throughout the six-minute long video.

Now, deepfakes are inherently weird, and this is surely the first time weve seen them used in such a way by a mainstream rap artist. As inventive as it is, there is a bizarreness to the music video as Lamars face isnt just replaced with the faces of living (and dead) Black celebrities, but he raps some of his verses while donning those faces, too.

But which deepfakes were the most bizarre? How about the least? Well, rather than try and interpret every second of the music video (Kendrick fans on Twitter are already going above and beyond around that), weve done the real hard work of ranking the deepfakes used in The Heart Part 5 from least to most bizarre.

As with all six deepfakes, theyre all pretty accurate in their likeness, especially Jussie Smolletts. Although hes one of the most unexpected of the bunch, hes not the most bizarre.

Considering Kanye is actually a rapper, its not too unnerving seeing his face rapping Kendricks lyrics. However, Kendricks hair paired with Kanyes face might be the most bizarre if our ranking was based solely on that.

Is it just me or does the deepfake of Will Smith just look like a light skin Andr 3000?

The Kobe deepfake is one of two deceased figures used in the video, and that adds to the bizarreness of it all. But the likeness just feels too uncanny.

Its the one that starts everything off and its so unexpected. The moment Kendrick covered his face only for a deepfake of O.J. to appear, I had to scroll back a few seconds to make sure I wasnt losing my mind. That, paired with the fact that its O.J., is why this deepfake is one of the most bizarre ones from the video.

What makes the Nipsey deepfake the most bizarre is that not only is this the other deceased figure used in the video, but that Kendrick also raps from the perspective of the late rapper when he dons his face. As Nipsey, he directs a few lines at the Crenshaw rappers brother Sam Asghedom:

And Sam, Ill be watchin over youMake sure my kids watch all my interviewsMake sure you live all the dreams we produceKeep that genius in your brain on the move

And he also exonerates Nipseys killer (I forgive you, just know your souls in question) which, depending on how you view it, could be seen as taking ones creative license a little too far. But theres no denying how eerie it is to see Kendrick transform into Nipsey and essentially stay as him until the song comes to a close.

Read this article:
Ranking Kendrick Lamars The Heart Part 5 Deepfakes From Least To Most Bizarre - Okayplayer