Nelson Sample Submission: Fill & Download for Free

You don’t want them to share your stuff on social media platforms. Publishers want to buy first rights. That is, the rights to something that’s never been published elsewhere before. Putting something online amounts to publishing it and greatly decreases the value of your submission.

This June, I celebrate two distinctions: I completed my sixth decade on this planet, and my 35th year of active research and academic publications in AI/ML. My first academic papers were published in IJCAI 1985, and we’ve just been notified that our 2020 ICML submission on extending reinforcement learning to nonstationary Markov decision processes has been accepted for presentation and publication. Having published over 150 papers in AI and ML over the past 35 years gives me some perspective to answer your important question.So, how does one go about doing research in AI and ML? For that matter, how does one go about doing research in any field? What is “research” anyway? How does it differ from other activities, like programming a Python compiler or changing oil in a car?At the heart of all foundational research is a question. It’s best to illustrate this with an example. For many years, I used the story of Cecilia Payne-Gaposchkin, a pioneering woman astronomer, as my role model of what it means to do a PhD thesis for my graduate students. Few have made as important a research discovery in their PhD thesis as this brilliant woman.Cecilia Payne-Gaposchkin - WikipediaAt the age of 25, her PhD thesis at Harvard answered the following question: what is the most commonly occurring chemical element in the universe? She conjectured it was hydrogen, against prevailing scientific wisdom and against her PhD advisor’s personal view, and her thesis firmly proved using measurements that she was right. Alas, my first academic papers couldn’t compare with such an earth shattering contribution.Perhaps it’s unfair to compare AI and ML with physics. After all, AI and ML are more engineering endeavors where researchers tend to build software artifacts that demonstrate some intelligent behavior, rather that discover some basic property of the universe, as Cecilia did. But there are broad similarities between the scientific goals of any research field.The most important and overlooked component of research is asking the right question. Alas, no one teaches you this in a course or a textbook. Textbooks comprise of knowledge. As Einstein often remarked:Imagination leads one to ask the right questions. Invariably, in AI and ML, these constitute formalizations of computational algorithms. I’ll take an example of the most famous PhD student to graduate from my former UMass Amherst lab in 1984: Richard Sutton did a pioneering PhD thesis under Andrew Barto, exploring how agents can learn to solve the “temporal credit assignment” problem. His question was simple and yet profound: if you only get delayed feedback, such as a win or loss in a game that is only known at the end, how could you possibly learn during the game to play against your opponent without knowing the final verdict.He was not the first to ask this basic question. Arthur Samuel, the pioneering IBM researcher who coined the phrase “machine learning”, asked the same question in 1959, and demonstrated the first ML program that learned to play checkers from self play on a vacuum tube IBM 701 computer that had no keyboard or display or programming language!Samuel and Sutton both studied the same problem, temporal credit assignment, and the same algorithm, temporal difference learning, but Rich Sutton brought the study of TD learning to new mathematical heights with a much deeper analysis. Remarkably, and in my experience, completely unprecedented, Rich continues to work on his PhD thesis question 36 years later. More than any other machine learning researcher, he exemplifies in my mind the ideal of a computational scientist, one who is deeply interested in foundational questions about intelligent behavior.So, how does one actually do research in ML or AI. Above all, you need passion and dedication. There’s no room for dilettantes. Rich exemplifies this dedication. He’s thought about TD learning literally every day for the past 36+ years. I guarantee you he understands it better than anyone else, even his former PhD advisor and my former UMass colleague, Andrew Barto, with whom I was privileged to co-direct the Autonomous Learning Lab for over 15 years. All the wonderful work by Deep Mind on Alpha Go and Atari video games using deep reinforcement learning could not have happened without Rich’s insights.Another brilliant PhD dissertation that brought the study of reinforcement learning to new heights was done by Chris Watkins in Kings College, England. Chris’ PhD thesis was simply titled “Learning from Delayed Reward”. Chris studied the same problem as Samuel and Sutton, but crucially linked TD learning to operations research, dynamic programming, and Markov decision process. His thesis contributed the fundamental Q learning algorithm, the first model free optimal control method. This illustrates another important component of research: connecting ideas from one field to another. Q learning remains the most widely used algorithm in reinforcement learning 30 years after Watkins’ PhD thesis. Many thousands of papers extending Q learning have been written in the past three decades.I’ll end with a final remark on how to do research. Many scientists deeply believe in the importance of simplicity. Einstein said it best: everything should be as simple as possible, but no simpler. This is often called Occam’s razor.One fundamental problem with Q learning is that it does not reliably converge when combined even with simple linear function approximation, let alone the huge nonlinear neural nets that Deep Mind’s engineers are fond of using. My research into this question led to a new formulation of Q learning, one that explored how to use powerful tools from optimization theory such as proximal gradients to shed new insight into this problem. My former PhD student received the Facebook Best Student Paper award in 2015 at the prestigious Uncertainty in AI (UAI) conference for his paper on this problem, which led to a longer journal article in the Journal of AI Research.Proximal Gradient Temporal Difference Learning: Stable Reinforcement Learning with Polynomial Sample ComplexityThis new formulation of gradient TD is mathematically elegant but nontrivial to understand. You need a deep background in optimization to understand the algorithm. Needless to say, this complexity goes against the grain of computational scientists like Sutton. He came up with a simpler formulation of TD called emphatic TD that is more stable under linear function approximation.This illustrates how science works. Scientists are not cold blooded but passionate. They argue a lot. They hold strong opinions. Until his dying day, Einstein refused to believe in quantum mechanics. His most favorite comment was:He said this so often in his debates with Nils Bohr, a pioneering quantum theorist, who finally told Einstein in exasperation: stop telling God what to do!Scientific research in any field is ultimately a spiritual experience. The moment of inspiration, for anyone who has experienced it, is one of exaltation. There is no better reward for a lifetime of doing research, of the endless toil.I remember vividly about 15 years ago, I was trying to understand how reinforcement learning agents exploring their environment through trial and error could discover its underlying symmetries. In short: how does structure emerge from randomness? That’s the question I was struggling to understand. My solution was an idea called “proto-value functions”, which, unlike regular value functions that Q learning estimates, are not derived from task specific rewards but task independent domain geometry. The underlying mathematics used the beautiful idea of the Laplacian operator, which has been called the most beautiful object in math and physics.“The Laplace operator in its various manifestations is the most beautiful and central object in all of mathematics. Probability theory, mathematical physics, Fourier analysis, partial differential equations, the theory of Lie groups, and differential geometry all revolve around this sun, and its light even penetrates such obscure regions as number theory and algebraic geometry.” — Nelson, Tensor Analysis.My ICML 2005 paper on proto-value functions has been extended by other researchers in many interesting and novel ways. A recent paper by researchers at Deep Mind and University of Alberta showed that there’s a nice connection to successor representations, which there’s evidence from neuroscience are present in the hippocampus.The hippocampus as a predictive mapAs you see, mathematics, physics, neuroscience, AI and ML come together beautifully sometimes in research, and I was fortunate enough to see this happen with a piece of my own research.To bring the story to its end, research ultimately is about the quest for beauty and truth, synonyms for many scientists. No one put this better than the legendary poet, Keats, in a famous poem called an Ode to a Grecian Urn. The last stanza has served to motivate many a scientist.Ode on a Grecian Urn by John Keats | Poetry Foundation"Beauty is truth, truth beauty,—that is allYe know on earth, and all ye need to know."

Well, you don’t appear to “sell” anything, so the site generates revenue from Ads.That means you predominately want Readers, not Converters.So you need to find sites that have a high readership with an interest in BlockChain.That could be Blogs, forums or Social Followings from specific individuals/brands.You can start by looking up some recent stories, and noting which sites have them.Look at social statistics to get a rough (and likely inaccurate!) idea of readership/engagement.Prioritise those sites.Then use tools like Alexa and Similar Sites to get a ballpark on their traffic.Then use backlink checking tools to see a sample of sites that link to those sites.That will give you an idea of where to get links from.(You can also then partially repeat the process and identify slightly lower-tier sites - which may be a better place to start from :D).The question is then - “how do I convince site owners to link to me”.Aside from the “quality” of your content … there are several other factors you need to nail down;Speed and freshness.As you are presenting as a news site - you need to be showing you get the story sooner than the others. This will require you to identify the “story chain”, work your way up it and get as close to the sources as possible.From there - when more established, you can try to insert yourself as a release point, and you can get stories directly.Veracity/accuracy.Far to often the “chain” is little more than a digital version of Chinese-Whispers.Incorrect information, perspectives etc. all creep in and warp the truth.You need to be seen to try and avoid that - and if it occurs, be seen to correct it!Authority/Trust.Writing is not the same as reporting! You need to be able to show that you are worth paying attention to, and that you can be trusted. Part of this is obtained by Speed and Veracity (above).But displaying knowledge, referencing prior instances, citing legitimate sources, interlinking to relevant stories etc. all contribute to the perception of being knowledgeable.Another set of facets you can look at is service/technology.Availability.Being a source of news is all well and good - but how do people know what’s on your site?You should look at supplying feeds (partial - not full syndication!).Widgets that will allow others to include your stories, ideally by category/topic/date.Contactability.What about submissions and suggestions?There may be some folk out there with a story - but it’s “small” and of little interest to the major players.If you provide a method of story submission - you can potentially get these little stories. Not only does that provide you with generally non-competed content, but it has promotional benefits (they will often share their story for you).