<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Chris von Csefalvay on Medium]]></title>
        <description><![CDATA[Stories by Chris von Csefalvay on Medium]]></description>
        <link>https://medium.com/@chrisvcsefalvay?source=rss-93652c542fd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*JIlLxBAh3E56Ubg2Fnw8Bg.jpeg</url>
            <title>Stories by Chris von Csefalvay on Medium</title>
            <link>https://medium.com/@chrisvcsefalvay?source=rss-93652c542fd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 10:13:21 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@chrisvcsefalvay/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Knowledge Dividend of LLMs: a pragmatic perspective]]></title>
            <link>https://medium.com/starschema-blog/the-knowledge-dividend-of-llms-a-pragmatic-perspective-79e8f8fb0686?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/79e8f8fb0686</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Mon, 02 Oct 2023 12:25:07 GMT</pubDate>
            <atom:updated>2023-10-03T15:47:06.667Z</atom:updated>
            <content:encoded><![CDATA[<p>As I’m writing this, the sun hasn’t risen over the Denver skyline in earnest. There’s still pink in the sky over the Front Range, and most of the world is still blissfully asleep. And so far, a small, moderately fine-tuned Large Language Model (LLM) trained on $500 worth of free credits has explained to me just how bad <a href="https://www.espn.com/nfl/story/_/id/38481606/miami-dolphins-score-70-points-vs-broncos-record-day">the Broncos’ recent 20–70 embarrassment against the Miami Dolphins</a> is (very), made some useful suggestions for a Caddoan language to learn if I wanted to help with language preservation (Pawnee) and created a fairly acceptable recipe to salvage whatever is left in my fridge (spicy tomato and cheese omelet with a chia side salad). Not too shabby for something that has absolutely no understanding of language preservation, omelets or American football (then again, neither do I, as far as the last one is concerned).</p><p>And therein lies one of the pervasive paradoxes of LLMs: they generate very confident, very credible and very often correct answers to questions on subjects they really don’t know all that much about. This is indeed a source of some philosophically rooted criticism of such models. The four brilliant authors of <a href="https://dl.acm.org/doi/10.1145/3442188.3445922">a remarkably perceptive 2021 paper</a> described the then-nascent language models (GPT-3 was still state of the art back then) as “stochastic parrots,” and this term has become a woefully overused term for such models among the critics of the LLM revolution. On one hand, this is ironic, because many of these critics themselves exhibit a psittacine lack of originality in regurgitating the same criticism. More importantly, however, it misses the ‘stochastic’ part.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w41ZCOX0prnVnCi4-cstiw.jpeg" /><figcaption>Image by <a href="https://unsplash.com/@girlwithredhat">Girl with red hat</a> on Unsplash.</figcaption></figure><p>What, exactly, is this stochasticity? Fundamentally, all LLMs currently in widespread use have one fundamental functionality: given a sequence of tokens (which roughly corresponds to words — let’s for now treat them as synonymous), they estimate a distribution of what tokens are most likely to come next, and draw a sample from the region of highest probability. Since they have some leeway as to the precise word they pick, the outcome does have a degree of randomness. This compounds as the newly picked word is now part of the context that conditions the distribution from which the next word will be picked, and so on, <em>ad infinitum</em>. At no point, of course, is there any insinuation that the model knows what any of those words ‘mean’ — it’s just really, really good at stringing words together. The stochasticity in this process is constrained by the conditional probability of each word given the previous words — probabilities that have been nailed down quite well during the training process and refined in RLHF (reinforcement learning with human feedback). The result is that even though LLMs don’t really <em>know</em> anything about languages, omelets or football, their probabilistic ability to string tokens together can replicate information in ways that, even devoid of what would satisfy philosophers to constitute knowledge, is good enough for us to make things work on them.</p><p>In this post, we’ll be looking what LLMs ‘know,’ or at least what they pretend to know, well enough for us to rely on. It turns out that the answer is ‘quite a lot.’</p><p><strong>“Knowledge” is Overrated (and so are philosophers)</strong></p><p>Much of the difficulty here is semantic and somewhat philosophical, and it’s intrinsically connected to the fact that we conceive of knowledge as primarily a human thing, one that requires the subjective life of the mind that machines do not have. A fascinating thought experiment that’s quite appropriate to this question is <a href="https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/abs/minds-brains-and-programs/DC644B47A4299C637C89772FACC2706A?utm_campaign=shareaholic&amp;utm_medium=copy_link&amp;utm_source=bookmark">the Chinese Room proposed by John Searle in 1980</a>, a more innocent age, when machine translation and conversation were distant enough prospects that entertaining them was largely the realm of thought experiments. In the Chinese Room, a computer has been ‘taught’ to read an input in Chinese (which Searle treated as a monolithic language, but I digress) and respond back in (written) Chinese, well enough that a native Chinese speaker could not conclusively tell that they aren’t talking to another native speaker. But does the machine ‘understand’ Chinese, Searle asked?</p><p>In the Chinese Room sense, LLMs don’t ‘know’ much about languages, omelets or football, just as the computer in the Chinese Room does not really ‘know’ Chinese the way we understand knowledge of a language. Rather, they know how to create a convincing simulacrum of knowledge in both cases.</p><p>It turns out that in this game, quantity has a quality all of its own. The distributions that are being sampled to pick the next word have been learned on massive corpora like the <a href="https://commoncrawl.org/">Common Crawl</a>. When a model sees that “the 44th President of the United States” is typically followed by “Barack Obama” and rarely, if ever, by, say, “Donald Duck,” the probability distribution from which the tokens next to “the 44th President of the United States” will be drawn will center heavily around “Barack Obama” as the answer. In what is reminiscent of the way CNNs build their own feature extracting filters, this process is extremely inductive. Nobody ever ‘taught’ LLMs who the 44th President was, or what a President is — merely that just as night follows day, a token stream on the 44th President is likely followed by President Obama’s name.</p><p>LLMs have what I will call ‘Gettier knowledge,’ in reference to the Gettier problem (“is justified true belief knowledge?”). Gettier knowledge is, simply put, a high-confidence stochastic belief justified by sufficient evidence (training data) that happens to be true. In short: for our purposes, regardless of how we define or conceive of knowledge in general, if an LLM has a good enough belief that “the 44th President of the United States” is semantically sufficiently tied to “Barack Obama,” that’s good enough for us to call it knowledge. The philosophers can argue about whether this is indeed true knowledge, but for our purposes, it will be good enough to proceed.</p><p><strong>Free Knowledge! Take One!</strong></p><p>If you’ve ever built a convolutional neural network, you must recall that moment when you realised this wasn’t going to be anything like other machine learning tasks. In traditional ML, you define features, extract them and learn over those features. It is ultimately very inductive, but still needs you to be opinionated enough to kickstart that induction. Neural nets are different: they are so purely inductive that they themselves create their own features in the process. If you try to build a deep convolutional neural network to identify, say, Golden Retrievers, it won’t ask you to tell it what a Golden Retriever <em>is</em>, just to show it some examples of Golden Retrievers, and preferably also some non-Golden Retrievers. LLMs have that same moment — an LLM doesn’t need to be taught what the capital of Hungary is, because it doesn’t really care. What it needs is an understanding that the region of highest probability for the token following “The capital of Hungary is” would contain “Budapest”. Then, sampling that probability distribution will get us “Budapest”.</p><p>From the practical perspective, this is just as good as knowledge: as long as we understand what we have created, the difference between a simulacrum of knowledge and knowledge itself, we can make good use of it. And if this simulacrum is built with enough information to create a good degree of fidelity to truth, then that may be all we need.</p><p>Consider the following scenario:</p><p><em>A drug, X, is approved for use in adults with generalized myasthenia gravis who are anti-AChR or anti-MuSK positive.</em></p><p><em>Mrs Hunter is 56 years old and has gMG. She does not have antibodies against cholinergic receptors or against receptor tyrosine kinase proteins.</em></p><p><em>Is X indicated for Mrs Hunter?</em></p><p>MuSK is, of course, one of the RTK Class XVIII proteins. A base implementation of GPT-4 gives us the correct answer, where the system infers the simplest of logical relationships, namely that of <strong>class membership</strong>: all MuSK proteins are RTK Class XVIII proteins, therefore no anti-Class XVIII proteins means MuSK negative, and hence X is not indicated (X here is based on the real world monoclonal antibody rozanolixizumab). If a statement applies to a class, then it applies to every member of that class, and therefore a statement about the class applies to every instance:</p><p><em>Mrs Hunter, as described, has generalized myasthenia gravis but does not have antibodies against acetylcholine receptors (AChR) or any receptor tyrosine kinase proteins, which includes muscle-specific kinase (MuSK).</em></p><p><em>Given this information, drug X would not be indicated for Mrs Hunter as she does not meet the specific criteria for which the drug is approved, namely the presence of anti-AChR or anti-MuSK antibodies.</em></p><p>This is the correct response — GPT-4 understands not only that a statement about a set is a statement about everything in that set, it also understands that MuSK is a kind of receptor tyrosine kinase (class membership knowledge).</p><p>LLMs acquire this knowledge as part of their training. For an LLM to be good at predicting what the next token is — which, as said, is largely what LLMs do — means to build an understanding based on real-world text, which usually is a good(ish) representation of reality. LLMs create “plausible” text, and if they’re trained on data that is at least somewhat accurate, they will have to learn about the world as a corollary to learning about the words. We call this the <strong>knowledge dividend</strong>: an understanding of the reality, as represented to the LLM by its training data, that is acquired as part of the process to convey quintessentially linguistic information to the LLM to be trained. This is because we train LLMs through their inductive properties, i.e. by examples, not theory. LLMs are trained on sentences like “the 44th President of the United States was Barack Obama”, and the name of President Obama might be masked out. Models that correctly infer Barack Obama as the most plausible tokens to fill the mask are rewarded. Together with a good vector embedding to back it up, the inescapable corollary of teaching a language model about words is to teach it also about the worlds they inhabit and reflect. This, then, is the fundamental idea behind what LLMs know, and <em>how</em> they know it. All they know is to generate plausible text, but if the standard of plausibility they are held to during training is also <em>truthful</em> text (i.e. accurate, high-quality training data), they will also learn a true representation of what the world is (they will learn a representation <em>anyway</em>, but what that will be hinges greatly on the training set).</p><p><strong>Forms of Semantic Knowledge in LLMs</strong></p><p>I have intentionally chosen an example that illustrates how the semantic knowledge inherent in LLMs can be used to answer real-world questions. There are, generally, five major classes of semantic knowledge inherent in LLMs.</p><p><strong>Class membership</strong></p><p>As illustrated above:</p><p><em>Q: Alice is in a store that sells plushies of various animals. She is trying to decide between a teddy bear, a plush puppy, a plushy dinosaur and a stuffed shark. She definitely wants a plushy of a mammal. Which of the toys is she not going to choose?</em></p><p><em>A: If Alice is looking for a plushy of a mammal, she will avoid choosing plushies representing animals that are not mammals. From the given options:</em></p><ul><li><em>A teddy bear represents a bear, which is a mammal.</em></li><li><em>A plush puppy represents a dog, which is also a mammal.</em></li><li><em>A plush dinosaur represents a dinosaur, which is a reptile (and is extinct).</em></li><li><em>A stuffed shark represents a shark, which is a fish.</em></li></ul><p><em>Given her preference for a plushy of a mammal, Alice would not choose the plush dinosaur and the stuffed shark.</em></p><p>The relevant knowledge item here is, of course, what is and isn’t a mammal, and the understanding that statements about the class apply to every element of the class.</p><p><strong>Hierarchies of Relationships</strong></p><p>An expansion on the previous is a hierarchy of relationships:</p><p><em>Q: Xs are an instance of an object — doesn’t matter what. Xs comprise As, which have the property s, and Bs, which do not have the property s. Bs are divided again into Ks, which have the property q and Ls, which do not have the property q.</em></p><p><em>Determine if the following statements are true (respond just with number and ‘true’ or ‘false’ or ‘cannot be decided’):</em></p><ol><li><em>Every K is a B.</em></li><li><em>No A has property q.</em></li><li><em>No K has property s.</em></li></ol><p>This is a little clunky, but we needed to pick attributes about which the system couldn’t possibly have knowledge or understanding. Given this, it can reason correctly:</p><ol><li><em>True</em></li><li><em>Cannot be decided</em></li><li><em>True</em></li></ol><p><strong>Functions, Attributes, Properties</strong></p><p>LLMs can have an understanding of what kinds of attributes or properties things may have:</p><p><em>Q: Of the following objects, return a list of those that have a gauge: railways, trees, elevators, shotgun shells, railway engineers, bosons.</em></p><p>I was pleasantly surprised by the ‘strong’ property understanding here:</p><p><em>Based on the generally accepted meanings of “gauge” in these contexts, the list of objects that have a gauge is as follows:</em></p><ul><li><em>Railways</em></li><li><em>Shotgun Shells</em></li></ul><p>Railway engineers do sometimes carry a tool to check rail gauge, and there’s a confounder here — a concept in theoretical physics called <a href="https://en.wikipedia.org/wiki/Gauge_boson">gauge bosons</a>, but bosons themselves do not ‘have’ a gauge, they are a kind of boson.</p><p><strong>Adjacency Knowledge</strong></p><p>Adjacency is the spatial cousin of class membership and relates to spatial conformation:</p><p><em>Q: A patient’s heart is imaged using a method of imaging, from an anterioposterior axis. Given a relatively narrow field centered on the heart, which of the following is more likely to be seen — the superior vena cava or the small saphenous vein?</em></p><p>GPT gets it:</p><p><em>A: When imaging the heart with a method focused on an anterioposterior axis with a relatively narrow field centered on the heart, structures in immediate proximity to the heart are more likely to be visualized.</em></p><p><em>The Superior Vena Cava (SVC) is a large vein that is in immediate proximity to the heart, transporting deoxygenated blood from the upper half of the body to the right atrium of the heart.</em></p><p><em>On the other hand, the Small Saphenous Vein is located in the leg, and is thus far from the heart and unlikely to be visualized when the imaging is centered on the heart.</em></p><p><em>Given the choice between the superior vena cava and the small saphenous vein, the Superior Vena Cava is more likely to be seen.</em></p><p>The question is a bit of a trick question in that there are forms of imaging in which the area outside the field of view generates a signal, but it does get the fundamental point across — what’s closer to an imaging focus is likely to be picked up better.</p><p><strong>Qualifications</strong></p><p>Qualifications are the conditions an item needs to meet for a statement to hold true for it — typically, a membership criterion. When is an X a Y?</p><p><em>Q: Make me a checklist of criteria a state has to meet to be considered a parliamentary democracy.</em></p><p><em>A: Checklist summary:</em></p><ul><li><em>Sovereign State</em></li><li><em>Constitution or Equivalent Legal Framework</em></li><li><em>Elected Legislature</em></li><li><em>Political Parties</em></li><li><em>Separation of Powers</em></li><li><em>Parliamentary System</em></li><li><em>Rule of Law</em></li><li><em>Accountable Governance</em></li><li><em>Civil Liberties and Political Freedoms</em></li><li><em>Stable and Effective Governance</em></li><li><em>Inclusive Citizenship</em></li></ul><p><strong>Sequentiality</strong></p><p>Some things need to happen in a certain order. Other times, the order of events determines positionality or outcome. The famous “coffee cup, thimble and diamond” test ultimately is a test of sequential understanding:</p><p><em>Q: I’m in my house. On top of my chair in the living room is a coffee cup. Inside the coffee cup is a thimble. Inside the thimble is a single diamond. I move the chair to my bedroom. Then I put the coffee cup on the bed. Then I turn the cup upside down. Then I return it to right side up, and place the coffee cup on the kitchen counter. Where’s the diamond?</em></p><p><em>A: Given the sequence of actions described, you did not mention removing the diamond from the thimble or the thimble from the coffee cup at any point. So, when you turned the coffee cup upside down, the thimble and the diamond would likely have fallen out, possibly ending up on the bed or the floor near the bed. After turning the cup right side up again and moving it to the kitchen counter, the diamond is not mentioned to have been placed back inside the thimble or the cup, so it would not be on the kitchen counter inside the cup. The most probable location, based on the given information, is that the diamond is still where it fell, near or on the bed.</em></p><p><strong>The Extent of Knowledge</strong></p><p>What is quite remarkable is the extent of this knowledge, even on relatively generalist models. For instance, most such models have a very good understanding of categories. GPT-4, in particular, does really well with classifying by aspects:</p><p><em>Q: List various categories into which I could classify the enzyme telomerase reverse transcriptase.</em></p><p>The response:</p><p><em>Biological Classification: enzyme (Enzymes are biological molecules that significantly speed up the rate of virtually all of the chemical reactions that take place within cells.)</em></p><p><em>Functional Classification: reverse transcriptase (This enzyme can synthesize DNA from an RNA template, contrary to the usual process of transcription.)</em></p><p><em>Biological Role: telomere maintenance (Telomerase reverse transcriptase plays a crucial role in maintaining the length of telomeres, which are regions of repetitive nucleotide sequences at each end of a chromosome.)</em></p><p><em>Subcellular Localization: ribonucleoprotein complex (It is part of a complex with RNA, forming a ribonucleoprotein complex, typically located in the nucleus of the cell.)</em></p><p><em>Clinical Relevance: cancer marker (The expression of telomerase reverse transcriptase is often associated with cancer cells, as it enables the unlimited division and immortality of these cells.)</em></p><p><em>Biomedical Research: drug target (Due to its role in cellular aging and cancer, telomerase reverse transcriptase is a significant target for drug discovery and development in cancer therapeutics.)</em></p><p><em>Molecular Biology: DNA synthesis (This enzyme is involved in the synthesis of DNA strands, specifically the addition of telomeric repeats to chromosome ends.)</em></p><p><em>Genetic Aspect: gene product (Telomerase reverse transcriptase is the protein product of a specific gene, often denoted as TERT in humans.)</em></p><p>The power of such language models is twofold, and even if you have skipped through most of this article, I hope you let this stick with you:</p><ol><li>This knowledge comes pre-loaded. At no point did I tell the model a thing about telomerase. A pretty stunning amount of knowledge comes ‘out of the box’ with these models.</li><li>Much of this knowledge can be assimilated and enhanced. For instance, there already are plenty of domain-specific models.</li></ol><p>I see two major consequences to this.</p><ol><li>Generalist models can do a lot of things quite well, operating on the basis of a fairly respectable knowledge base — facts like enzyme hierarchies, interventions, where things are located and where even fictional things would be located (e.g. “would something exploding in the kitchen of my fictional two-bedroom apartment shatter my toilet bowl?”).</li><li>On the other hand, specialist models, which can be trained now quite inexpensively over domain-specific information, will always be more extensive. While GPT-4 can answer questions like the protein classification question above, there is a point at which fine-tuning becomes the way to go.</li></ol><p><strong>Conclusion</strong></p><p>There are plenty of applications that would benefit from an LLM-enabled perspective for no other reason than the ability of such models to store, represent and respond to — and with — a wide range of knowledge. From a practical perspective, such models come with ‘batteries included,’ and a fairly thorough understanding of the world they operate in, and the logical tooling to at least pretend to be relatively good at reasoning about those fact items.</p><p>In my practice advising the biomedical sector on AI/ML applications, we often see challenges that derive from the difficulty of incorporating complex knowledge bases. In many other industries, the knowledge that LLMs acquire as part of their training for more basic functionalities can become a significant asset.</p><p>Even relatively simple language models have a stunning grasp of information about the world they inhabit, and those trained on large, high-quality and informative corpora (such as the Wikipedia corpus) and improved via RLHF can take an enterprise’s domain understanding and specialist knowledge further. Philosophers may debate whether this is truly knowledge, but from the pragmatic perspective, this knowledge representation can serve as a valuable ‘base layer’ for understanding an increasingly complex world.</p><p>REACH OUT TO US <a href="https://starschema.com/contact">HERE</a> TO LEARN MORE:</p><p><a href="https://starschema.com/contact">Contact</a></p><p>READ MORE STORIES FROM STARSCHEMA:</p><ul><li><a href="https://medium.com/starschema-blog/a-no-nonsense-approach-to-large-language-models-for-the-enterprise-8ac0dbde017a">A No-Nonsense Approach to Large Language Models for the Enterprise pt. 1</a></li><li><a href="https://medium.com/starschema-blog/asemantic-induction-of-hallucinations-in-large-language-models-c92ef5030714">Asemantic Induction of Hallucinations in Large Language Models</a></li><li><a href="https://medium.com/starschema-blog/automating-consumer-goods-r-d-with-product-stability-forecasting-5430bfd23298">Consumer Goods R&amp;D with Automated Product Stability Forecasting</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=79e8f8fb0686" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/the-knowledge-dividend-of-llms-a-pragmatic-perspective-79e8f8fb0686">The Knowledge Dividend of LLMs: a pragmatic perspective</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Asemantic Induction of Hallucinations in Large Language Models]]></title>
            <link>https://medium.com/starschema-blog/asemantic-induction-of-hallucinations-in-large-language-models-c92ef5030714?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/c92ef5030714</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[gpt-3]]></category>
            <category><![CDATA[language]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Thu, 23 Mar 2023 09:39:36 GMT</pubDate>
            <atom:updated>2023-03-23T09:39:36.124Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Live with a man 40 years. Share his house, his meals. Speak on every subject. Then tie him up, and hold him over the volcano’s edge. And on that day, you will finally meet the man.<br>– Adlai Niska citing the fictional despot Shan Yu in Firefly, 1x10: War Stories</blockquote><p>It’s late March 2023, and you cannot throw a stone without hitting something that involves large language models (LLM), especially the new and shiny GPT-4. And truly, GPT-4 is amazingly entertaining. You could, but for the sometimes patchy performance, spend an infinity trying to get it to answer your questions in limericks. Or golden age hip hop lyrics, for that matter.</p><p>However, when you want to learn about how something works and what its true nature is, what it does well is much less important than what it does not. Taking a tool to the extremes of its capabilities and seeing how to make it misbehave is how we learn how that tool really works in practice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fkDXRJLr4B4JHCqj" /><figcaption>Photo by <a href="https://unsplash.com/de/@jrkorpa?utm_source=medium&amp;utm_medium=referral">Jr Korpa</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Hallucinating Models</h3><p>For GPT-4, the way to do that is by inducing hallucinations. In this context, a hallucination is something that is <em>realistic</em>, but not <em>real</em>. Just like ordinary sensory hallucinations in humans, it replicates the perception of what is, but not its connection to reality.</p><p>Generative models like GPT-4 hallucinate because of what they <em>are</em>. This is, I think, a much too neglected point. GPT-4 doesn’t hallucinate because it’s not good enough yet, or because it needs more training data or a larger corpus or a next version: it hallucinates because of what it is.</p><p>Somewhat counterintuitively, GPT-4 is not really concerned with creating some sort of “truth” it has learned. Rather, generative LLMs create something that looks plausibly like the truth. If I ask GPT-4 to tell me, say, the total area of Hungary, it will try to give me something that might be a relatively decent answer. Train the model enough, and that answer moves from what is plausible to what is actually correct… mostly, anyway. And therein lies the heart of what we are probing at in this case.</p><h3>Semantic and Asemantic Content</h3><p>Language is highly semantic, but because of that, it is also highly flexible. By semantic, I mean that every lexeme has a distinct and coherent meaning. A lexeme is the “root form” that is conjugated to various forms, e.g. “see”, “saw” or “seeing” are all forms of the same lexeme, “SEE” (by convention, in linguistics, we set lexemes in capitals to distinguish them of their homomorphic conjugated forms). By flexibility, I mean that you can actually manipulate lexical order and retain meaning. Consider the sentence “Joe walked his dog in the park”. “In the park, Joe walked his dog” and “His dog, Joe walked in the park” all have slightly different nuances due to inflection and emphasis of order, but overall, they all get the same larger message across. Some languages permit even more flexibility with word order, but even in English, the worst case scenario is that it’ll sound a little odd. The content remains intelligible, which is why poets get to move words around to make things rhyme. In short, producing something that looks “like” natural language is going to <em>be</em> natural language. It’s likely not going to be a Booker Prize-winning product of staggering genius, but it will be good enough.</p><p>This is not the case for asemantic structures. By asemantic structures, I mean any system in which a sequence of tokens has a meaning, but in which there’s no semantic meaning to any token. In other words, every token has a meaning, but that meaning is not inherent in their character. It probably serves to belabour this point a little. All words are, to an extent, made up, but more similar words are closer to each other in meaning. By similar, I do not mean simple formal similarity, such as Hamming or Levenshtein distances, but rather logical similarity, e.g. conjugation of the same root. This is more evident in some languages than others. In Semitic languages, for instance, you have clusters of meaning that are connected by consonantal roots. By way of an example: the triconsonantal root k-b-d connects a cloud of meanings that all have to do with the middle or centre of something, and by extension the centre of gravity or honour. This gives us e.g., the Hebrew words for ‘heavy’ (כָּבֵד), mass in the physical sense (כֹּבֶד), and the word for ‘liver’ (כָּבֵד), which was considered to be roughly in the middle of the body. In any language, however, there is a degree of meaningful semantic similarity between connected concepts. There has more been written on this than I have the space to mention here.</p><p>An asemantic structure is where there are formally similar things that are unrelated. You have probably experienced this when you dialled the wrong number by a slip of the finger. The fact is, if you live in the United States, my phone number looks a lot like yours, and by extension, anyone else’s. There’s no semantically meaningful reason why your phone number shouldn’t be mine or <em>vice versa</em>: it’s not more ‘you’ as mine is not more ‘me’ in any underlying sense.</p><p>Which is why GPT-4 struggles with asemantic content, and we’ll use this to break it a little.</p><h3>Induced Asemantic Hallucination</h3><p>Because to the best of my knowledge, nobody has written about this in any focused way, I’d like to define Induced Asemantic Hallucinations (IAHs): an IAH is a hallucination of a generative model that is defined by the fact that it occurs when the generative model seeks to produce asemantic content. It happens because the very meaning — the <em>semantics</em> — that dictate what goes together is absent from such content. Asemantic content is arbitrary. Consider the following sequence of numbers:</p><p>9, 7, 8, 0, 3, 2, 3, 9, 5, 3, 8, 9, 4</p><p>If you’re around books a lot, you might recognise this as an ISBN-13 number (<a href="https://www.amazon.com/Computational-Modeling-Infectious-Disease-Applications-ebook/dp/B0BVVR563W">specifically, it’s that of my book</a>). However, there’s no cogent reason why this should not be just about any other number beginning with 978 and having the right last digit (ISBN numbers have a check digit at the end calculated by the modulus 11 of the first nine digits weighted by their reverse ordinal).</p><p>Now look at the following problem. I’ve asked GPT-3.5 to tell me about something that I do know a thing or three about, to the point that I’ve read pretty much anything that’s ever been written. I’ve asked GPT-3.5 to give me five papers, complete with citations on the role of VP24 in Ebola virus disease. Here’s the first one I got:</p><blockquote><strong>Chris</strong>: Can you give me five papers, complete with citations and PMIDs, on the role of VP24 in Ebola virus disease?</blockquote><blockquote><strong>ChatGPT 3.5</strong>: Certainly! Here are five papers on the role of VP24 in Ebola virus disease:</blockquote><blockquote>1. Hoenen T, Groseth A, Falzarano D, Feldmann H. Ebola virus: unravelling pathogenesis to combat a deadly disease. Trends Mol Med. 2006 Nov;12(11):206–15. doi: 10.1016/j.molmed.2006.09.005. Epub 2006 Oct 3. PMID: 17011756.</blockquote><blockquote>This paper provides an overview of the pathogenesis of Ebola virus disease and the role of various viral proteins, including VP24.</blockquote><p>You may note that almost everything is correct about this… almost. The authors are correct — <a href="https://pubmed.ncbi.nlm.nih.gov/16616875/">this is indeed Thomas Hoenen’s paper</a>. The venue is correct, it appeared in 2006 in <em>Trends Mol Med</em>. But it was published not in November but May, in volume 12(5) not 12(11), albeit the page numbers are right. Its DOI ends in .006, not .005, and the publication date is 2006 April 17. Most glaringly, however, the PMID is wrong. <a href="https://pubmed.ncbi.nlm.nih.gov/17011756/">The PMID actually leads to an article by Carlsson <em>et al.</em></a> on messing with cell membranes using cetylpiridinium chloride, which has pretty much nothing to do with the original subject.</p><p>Note what ChatGPT 3.5 got right and what it got wrong. Names are meaningful, as are publication venues, because there’s other information in the corpus to connect it. However, dates are arbitrary, and so is the PMID. There’s nothing inherent in the article, or anything knowable about the article, that could lead anyone to know that the last three digits of the PMID will be 756. In other words, you could be in the (unenviable) theoretical position of having every single possible piece of knowledge that has ever existed the moment before the PMID was assigned, and you could not, with any confidence, predict what it would be.</p><p>This makes the PMID asemantic, and GPT responds to asemanticity by hallucinating… wildly hallucinating.</p><p>This is an example of IAH: we induced it by targetedly requiring GPT to retrieve asemantic information, rubbing our hands with malicious glee and the knowledge that this tends to trip up GPT and similar LLMs. With some coaxing, you can generate your own IAHs: from ISBNs to runways, from phone numbers to Dewey decimal codes, if you can get a generative LLM to try to come up with some asemantic information, you will be able to get it to hallucinate as soon as you can get it to go beyond what its core knowledge base is.</p><h3>Bad Trip</h3><p>As part of my work, I deal a lot with asemantic information, and its slightly nicer cousin, semi-semantic information, when there’s some information value, but not a whole lot. As a computational epidemiologist and population health data scientist, for instance, we rely heavily on various coding systems, which come at various degrees of semanticity. If by any chance, your spacecraft crashes and you see your physician for the first time, they will log that encounter using the weird and wonderful ICD-10 code of V95.41XA — spacecraft crash injuring occupant, initial encounter (in case you wonder, there are codes for collision, fire and forced landing of spacecraft, but none for zombie aliens). Some of this has semantic meaning, e.g. V means it’s an external cause of mortality and V95 comprises things going wrong with powered aircraft the same way V30 describes an occupant of a three-wheeled motor vehicle hitting a pedestrian or animal and getting injured in the process. However, there’s nothing inherent in the number 95.41 that is more intrinsically connected to a spacecraft crash than the number 12.34 would be. Or, indeed, any other number. GPT knows one code — language. Confront it with any other coding schema, no matter how simple, and what a look-up table could accomplish with no problems will break the flagship of LLMs.</p><p>The heart of the matter is this: we ought to be wary of investing these complex language models with our human understanding of what it is they do. LLMs don’t think, they pretend to think and pretend to give answers. The pretense is very, very good — to the point that the pretense is actually <em>useful</em>. GPT-4 is a very good faker who can fake knowledge well enough that they might as well be the real deal. Yet it is not a good enough faker to always get it right. Worse, these are genuine hallucinations: it’s not that GPT doesn’t know what to say so it confabulates. It’s rather that the model it uses to arrive at a result — any result — is so intrinsically dependent on the semanticity of the underlying content that it cannot cope with asemantic information. In the context of IAHs, we’ve exploited this weakness and overreliance on the highly semantic construct that is language. However, it may occur spontaneously, and when LLMs hallucinate, the users might be in for a bad trip indeed.</p><p>REACH OUT TO US <a href="https://starschema.com/contact">HERE</a> TO LEARN MORE:</p><p><a href="https://starschema.com/contact">Contact</a></p><p>READ MORE STORIES FROM STARSCHEMA:</p><ul><li><a href="https://medium.com/starschema-blog/can-we-stop-chatgpt-from-just-making-stuff-up-f622ad5e30bf">Can We Put a Lie Detector on ChatGPT?</a></li><li><a href="https://medium.com/starschema-blog/data-for-the-next-pandemic-f340cac8558c">Data for the Next Pandemic</a></li><li><a href="https://medium.com/starschema-blog/as-we-may-see-the-world-after-dashboards-81dda9a0a7e7">As We May See: The World after Dashboards</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c92ef5030714" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/asemantic-induction-of-hallucinations-in-large-language-models-c92ef5030714">Asemantic Induction of Hallucinations in Large Language Models</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data for the Next Pandemic]]></title>
            <link>https://medium.com/starschema-blog/data-for-the-next-pandemic-f340cac8558c?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/f340cac8558c</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-marketplace]]></category>
            <category><![CDATA[covid19]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Tue, 07 Mar 2023 21:32:01 GMT</pubDate>
            <atom:updated>2023-03-22T14:47:08.902Z</atom:updated>
            <content:encoded><![CDATA[<p>Three years into the COVID-19 pandemic, the medical, public health and disaster preparedness communities are trying to isolate lessons learned from the harrowing experience of a global outbreak that resulted in little shy of seven million deaths. There remains considerable disagreement as to what ought or ought not have been done, what worked and what did not and how to best approach a coming pandemic.</p><p>One of the only points on which there is widespread agreement regarding the global response to the pandemic is that data played a crucial role in tackling the crisis. Where data drove decision-making, outcomes were almost universally better in terms of morbidity and mortality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G9n9S0mwBVEUWMOMT4MR5w.jpeg" /></figure><p>In the preface to my recent book, <a href="https://www.elsevier.com/books/computational-modeling-of-infectious-disease/csefalvay/978-0-323-95389-4"><em>Computational Modeling of Infectious Disease</em></a>, I wrote:</p><blockquote><em>Computational models of infectious disease can make all the difference in our response to pandemics. As habitat loss and climate change make zoonotic spillover events increasingly more likely, COVID-19 is almost certainly not the last major pan- demic of the 21st century. In fact, it is reasonable to assume that such outbreaks will become increasingly frequent. Computational models can be powerful weapons in our fight against pandemics.</em></blockquote><blockquote><em>The increasing availability and decreasing cost of computing resources means that large-scale computation has never been this affordable in human history. The average high school student’s graphing calculator packs more computing power than has been available for the first several decades of computer science. Massively distributed architectures can simulate populations of millions in agent-based models and examine the effectiveness of increasingly complex public health policies in a data-driven and evidence-based manner.</em></blockquote><p>My work was greatly inspired by my experience on two COVID-19 data-related initiatives: The COVID-19 Tracking Project and Starschema’s COVID-19 Epidemiological Data Set. I wrote about the latter here:</p><p><a href="https://medium.com/starschema-blog/covid-19-and-the-first-war-of-data-science-980798f075ef">COVID-19 and the first war of data science</a></p><p>While events like COVID-19 or the 1918–19 influenza pandemic (Spanish Flu) are thankfully once-in-a-century anomalies, epidemics are a fact of human life. It has been so for millennia, perhaps ever since the first larger conurbations sprung up in the Neolithic. Somewhere — perhaps in a tropical rainforest, perhaps in the thawing permafrost soil or maybe in one of our own cities — the next pathogen to try humanity’s resilience and resourcefulness is slowly emerging.</p><p>Despite the advances of modern medicine, the challenges of global epidemics have only become greater. Habitat loss of viral reservoir species increases the likelihood of zoonotic spillover events. Our global trade and transportation networks enable pathogens to make their way around the world in 24 hours. Climate change is disrupting fragile ecosystems and global poverty, especially urban poverty, exacerbates the problem.</p><p>Antibiotic overuse, especially in intensive animal farming, raises the risk of horizontal gene transfer of antibiotic resistance genes to other pathogens. The development of new antivirals has always been a slow, painstaking process, and antibiotic discovery is not keeping up with pathogens: in the last decade, fewer than 20 new antibiotics (and a single non-antibiotic biological, the monoclonal antibody obiltoxaximab) have been approved, and the pipeline looks quite bare.</p><p>If that sounds like bad news, there is reason for optimism, too. Data has made all the difference in this pandemic. Learn how we need to prepare for the next.Modern methods of infectious disease modeling can arm humanity against pathogenic threats:</p><ul><li>There has never been a better time to create complex spatial models. Thanks to open data, most of the major roads and street networks across the world have been digitally mapped. In my book, I illustrate the utility of this by real-world use cases, such as determining the optimal site of a testing centre based on street networks in Manhattan or exploring the vulnerability of individual parts of Oxford by their distance from the nearest medical facilities. The resolution of the spatial models that we can now build is light-years ahead of what we have been used to.</li><li>Agent-based simulation, while computationally expensive, allows for deep modeling of complex human conduct. The attitudes, decisions and emotions that drive our destinies are complicated. They are not necessarily amenable to simple models. Simulated synthetic populations based on public data can create very accurate predictions.</li><li>In a paper jointly authored with my colleague Tamas Foldi in the early days of the pandemic, we argued for <a href="https://zenodo.org/record/3782877#.ZAUa6C9OmfA">a global, data-driven early warning system for pandemic threats</a>. The methods of anomaly detection are now sophisticated enough to bridge the ‘diagnostic gap’— the time difference between the emergence of a pandemic and detecting cases. Because anomaly detection can operate unsupervised, it does not require an exact case definition before it can pick up suspected cases.</li></ul><p>All of this, of course, is premised on the availability of data — high quality, high quantity and high velocity data. Much of what data was available during COVID-19 was a patchwork of various data sets, and that made aggregated data sets like ours particularly useful for researchers, private sector operators and the public health establishment, too.</p><p>There is, as of yet, no global facility where data could be efficiently shared. The landscape resembles a number of roadside stalls more than a single, unified marketplace. In our practice, we routinely enable some of the world’s largest companies to share data using convenient, secure and well-architected data exchange/data marketplace architectures. There is no reason why the same could — or should — not be done for health data.</p><p>We will need it in the next pandemic.</p><p>REACH OUT TO US <a href="https://starschema.com/contact">HERE</a> TO LEARN MORE:</p><p><a href="https://starschema.com/contact">Contact</a></p><p>READ MORE STORIES FROM STARSCHEMA:</p><ul><li><a href="https://medium.com/starschema-blog/five-healthcare-ai-ml-trends-to-watch-for-in-2023-b46f138d26b">Five Healthcare AI/ML Trends to Watch for in 2023</a></li><li><a href="https://medium.com/starschema-blog/as-we-may-see-the-world-after-dashboards-81dda9a0a7e7">As We May See: The World after Dashboards</a></li><li><a href="https://medium.com/starschema-blog/8-best-practices-for-working-with-your-data-science-vendor-from-data-scientists-4e07ca4d3181">8 Best Practices for Working with Your Data Science Vendor — from Data Scientists</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f340cac8558c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/data-for-the-next-pandemic-f340cac8558c">Data for the Next Pandemic</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Five Healthcare AI/ML Trends to Watch for in 2023]]></title>
            <link>https://medium.com/starschema-blog/five-healthcare-ai-ml-trends-to-watch-for-in-2023-b46f138d26b?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/b46f138d26b</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Mon, 05 Dec 2022 09:33:46 GMT</pubDate>
            <atom:updated>2022-12-05T09:33:46.955Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L_hcN1quxBSZvz8i" /><figcaption>Photo by <a href="https://unsplash.com/@helloimnik?utm_source=medium&amp;utm_medium=referral">Hello I&#39;m Nik</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>2022 has been an interesting year for healthcare AI/ML. It saw, among others, a consolidation of pandemic-induced practices, such as a shift towards virtual care (especially in primary care), into the ‘new normal’ of healthcare delivery. The COVID-19 pandemic has spurred providers and payors alike to rethink traditional models of care and offer more cost-efficient virtual visits that suit the lifestyles of younger patients, who are used to on-demand services on their schedule rather than appointments at fixed times, as well as older and more vulnerable patients who have often experienced barriers in receiving care due to a lack of transportation and access. Payors are now looking towards automating some routine tasks using conversational Natural Language Processing (NLP).</p><p>We expect these trends to continue in 2023. As medical AI/ML applications become both increasingly widespread and routine, their adoption is likely to become more prevalent, too. This was given a much-awaited regulatory push by the <a href="https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software">FDA issuing final guidance on its regulatory practice with respect to clinical decision support (CDS) systems</a>, exempting, among others, knowledge management systems that support clinical practice by providing recommendations and guidelines relevant to a particular disease or condition from the FDA’s device regulatory purview. This is likely to open up a major interest in helping physicians manage the wealth of clinical information and research that is generated at an increasingly rapid pace. In addition, there is a growing awareness for the role of AI in patient safety and privacy, especially for devices that interface directly with a patient (e.g. robotic surgery instruments).</p><p><strong>Processing Exascale Unstructured Data</strong></p><p>According to a 2019 <a href="https://www.nature.com/articles/s41591-018-0316-z">article in <em>Nature Medicine</em></a>, healthcare generates around 150 exabytes of data a year, growing around 48% annually, most of it unstructured. The management of data assets of this scale remains challenging, and exascale computing remains generally unavailable to the public. The challenges of handling even a fraction of this data is highlighting the importance of an emerging new profession: the healthcare data engineer, a clinical informatics specialist experienced with creating stable, reliable data lakes and data pipelines of clinical data, at scale and in a cost-efficient manner.</p><p>The fundamental challenge is that speed, safety and cost-efficiency are rarely attainable at the same time. Healthcare data engineers will continue to face the challenge of managing the data assets of the industry that is estimated to produce a little under a third of all data stored in digital form. While the first decades of clinical informatics were largely focused on creating interoperability through APIs like FHIR/HL7, the emerging challenges have to do more with managing data at scale to both facilitate operational tasks (patient care, medical record-keeping, imaging and PACS functionalities &amp;c.) and have the ability to draw insights from it. As health systems, payors and other institutions realize the value of their clinical data assets, the demand for clinical data engineers will skyrocket.</p><p><strong>AI-assisted Radiology</strong></p><p>The COVID-19 pandemic highlighted that in a surge of patients, radiologists can often become overwhelmed with work. During the pandemic, a number of quite viable models emerged to differentiate normal, non-COVID and COVID viral pneumonia chest X-ray images. While most of these models are not mature enough for clinical use, the rapid and relatively inexpensive development of such models has highlighted the viability of deep learning and computer vision in assisting radiologists.</p><p>A key difficulty of AI-assisted radiology is that, beyond the more trivial functionalities (e.g. segmentation models to quantify lesion size, count and volume), issues of regulation and liability remain unsettled. As the use of such solutions expands, we may expect the regulatory bodies to catch up and determine these issues. Radiologists can benefit from the assistance of AI-driven image analysis solutions to increase their throughput, decrease time spent on relatively menial parts of the image analysis process and automatically prioritize cases if the demand for their services exceeds capacity.</p><p><strong>AI in Patient Safety</strong></p><p>There is an increased drive towards leveraging AI to ensure patient safety and privacy through applying advanced analytics. This is of particular importance in the context of medical devices that are directly interfacing with patients, from robotic surgery through insulin pumps to brachytherapy machines. For much of the past, such devices relied on hard interlocks, trained personnel and/or air-gapped installation to prevent intentional or unintentional misuse that might harm the patients. AI can increase patient safety by monitoring inputs and identifying anomalies, whether it is spotting an accidentally added zero on the dose field of a radiotherapy prescription, an employee attempting to obtain narcotics from an automated dispensing machine with intent to divert it or a malicious hacker gaining access to critical equipment.</p><p>AI can also prevent unintentional medication errors by correlating its knowledge of the patient with the prescription, such as a physician accidentally prescribing a normal dose to a patient with renal impairment, resulting in an excessive dose, or an oversight causing a recent surgical patient to be discharged without the required anticoagulant prescription. Through forming an overall view of what “normal” care would entail for a patient with the given characteristics and clinical history, anomaly detection can highlight potential deviations and create a “speed bump” to focus a provider’s attention to a potential risk to patient safety.</p><p><strong>Decentralized Clinical Trials</strong></p><p>The face of clinical trials is changing rapidly. Decentralized Clinical Trials (DCTs), often enabled by smart devices and wearables, have become popular in recent years. This is partly thanks to the wide popularity of and patient familiarity with smart devices and partly to the cost-efficiencies it provides. AI-driven adaptive monitoring decreases the monitoring burden on the patient while maintaining data quality through identifying patients who are more likely to require monitoring and increasing monitoring frequency.</p><p>A key challenge for DCTs remains data availability and the difficulties of engineering safe and reliable systems to store trial data. Users are looking for end-to-end, turnkey ecosystems that are nonetheless adaptable and seamlessly link from onboarding, consent, data management and storage to analytics, intelligence and evaluation. A “real-time analytics” capability of DCTs can be helpful for Data Safety Monitoring Committees (DSMCs) in identifying risks or potential reasons to consider discontinuation. While clinical trials should not be discontinued for economic reasons, understanding trials as ongoing processes of accumulating data rather than a series of “snapshots” allows planning, adjustments and strategic decision-making.</p><p><strong>The Knowledge Management Challenge</strong></p><p>In the first year and half of COVID-19, <a href="https://royalsocietypublishing.org/doi/10.1098/rsos.210389">scientists published over 200,000 papers relating to the disease</a>. While this certainly shows a vivid interest and the ability of the global scientific community to respond rapidly to an outbreak, it also poses unique challenges to clinicians who need to consolidate information and integrate it into their clinical practice. The knowledge management challenge is likely to become increasingly acute with the ability to rapidly publish via Open Access and preprint servers.</p><p>Language models can help physicians overcome this flood of data by distilling it down to a manageable form. Until relatively recently, the very specific and sometimes counterintuitive language of scientific writing, especially in the biomedical and life sciences, has hampered the use of language models trained on common conversational language on academic publications. However, the rise of Transformer models and their increasing efficiency and rapidly dropping costs are helping AI-based solutions analyze, collate and summarize academic output, guidelines and diagnostic criteria or treatment protocols into an evolving, up-to-date picture of a clinical entity. Such clinical knowledge management solutions will greatly benefit from the FDA’s September 2022 guideline that indicates it would not consider such systems to be subject to “device” regulation, and we can expect some of these solutions to rapidly rise to the forefront in the coming years.</p><p><strong>About the author</strong></p><p><em>Chris von Csefalvay is the Practice Director of AI/ML for the healthcare portfolio at HCL Tech, where he advises Fortune 500 pharmaceutical, medical technology and life sciences enterprises on AI/ML strategy, policy and technologies. He is a Fellow of the Royal Society for Public Health and the author of </em>Computational Modeling of Infectious Disease<em>.</em></p><p><strong>REACH OUT TO STARSCHEMA </strong><a href="https://starschema.com/contact"><strong>HERE</strong></a><strong>:</strong></p><p><a href="https://starschema.com/contact">Contact</a></p><p><strong>READ MORE STORIES FROM STARSCHEMA:</strong></p><ul><li><a href="https://medium.com/starschema-blog/as-we-may-see-the-world-after-dashboards-81dda9a0a7e7">As We May See: The World after Dashboards</a></li><li><a href="https://medium.com/starschema-blog/learning-to-stop-worrying-and-augment-your-analytics-23c20b712a23">Learning to Stop Worrying and Augment Your Analytics</a></li><li><a href="https://medium.com/starschema-blog/data-science-trends-to-rule-2022-68ab4c93cd2">Data Science Trends to Rule 2022</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b46f138d26b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/five-healthcare-ai-ml-trends-to-watch-for-in-2023-b46f138d26b">Five Healthcare AI/ML Trends to Watch for in 2023</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[As We May See: The World after Dashboards]]></title>
            <link>https://medium.com/starschema-blog/as-we-may-see-the-world-after-dashboards-81dda9a0a7e7?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/81dda9a0a7e7</guid>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[augmented-analytics]]></category>
            <category><![CDATA[dashboard]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Fri, 01 Apr 2022 09:03:55 GMT</pubDate>
            <atom:updated>2022-04-01T09:03:55.320Z</atom:updated>
            <content:encoded><![CDATA[<p>In 1945, influential scientist and government advisor Vannevar Bush published <a href="https://www.theatlantic.com/magazine/archive/1945/07/as-we-may-think/303881/"><em>As We May Think, </em>an article of remarkable foresight</a>. In less than a dozen pages, he outlined many futuristic technologies that have come to pass, including hypertext, the personal computer and speech recognition. His memex, a kind of World Wide Web for the analog age, presaged today’s connected informational environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/588/1*aAb_o_fkrbCGlC8CFB2O0Q.jpeg" /><figcaption>Vannevar Bush (1890–1974), a visionary of the digital age before it was a thing.</figcaption></figure><p>Eighty-five years later, we are at the cusp of equally groundbreaking possibilities. Natural language processing (NLP), natural language generation (NLG) and machine learning (ML) have merged to create a new paradigm for the representation of information. Instead of asking questions from data, we are moving towards self-curating data experiences that tell decision-makers what they need to know.</p><p><strong>Dashboards are over (if you want it)</strong></p><p>Enterprises have spent the last two decades meticulously curating dashboards. Tools have improved and what once was the preserve of trained professionals has largely been democratized across the enterprise. However, dashboards are quintessentially static. A dashboard shows what its developer has directed it to show. That ignores the vast realm of the ‘unknown unknown’ that is, what we don’t know that we don’t know.</p><p>Dashboards, in that sense, answer narrow questions while giving the illusion of comprehensiveness and a 30,000ft perspective. In reality, dashboards are limited tools to convey limited information, while providing a psychologically fulfilling yet false illusion of comprehensiveness.</p><p>This may be true, but until now, there have been rather few alternatives. In the modern enterprise, a dashboard is created to provide a consumer with information on a particular subject. This involves certain judgments about what is, and what is not, included. This, in turn, biases the observer to what they are presented: they are, in a sense, trapped in an involuntary version of McNamara’s fallacy. What is not quantified and dashboarded does not exist. Few modern executives have the capability to reach beyond these summaries of data (if they did, they wouldn’t need dashboards in the first place).</p><p>However, machine learning may spell a fundamental change in this. We now have the tools and techniques to sift through vast volumes of data and evaluate the relative saliency of each data item.</p><p><strong>Introducing SABRE</strong></p><p>Saliency Based Representation of Evidence (SABRE) is a new paradigm for creating data experiences. SABRE considers an initial dashboard design little more than an informative prior that is subject to updating. As the volume of data increases, SABRE updates these priors with indicators of relative saliency of the incoming data.</p><p>Saliency refers to the relative cognitive importance of a feature. In other words, salient features are those slices of a wider reality that most efficiently convey a situational picture to an observer. A driver on a highway may pay particular attention to the road, the vehicle in front and the speedometer gauge. These are salient because from the perspective of decision-making by the driver, the situational picture is almost exhaustively described by these facts, while the row of trees adjoining the road contributes rather little to the driver’s decision-making.</p><p>Consider now one of the trees from the row of trees starting to lean over the highway. Suddenly, this becomes acutely relevant to the driver’s conduct if he expects to survive. This is an example of an anomaly: an otherwise rather less noticeable object behaving in an unusual manner. On the other hand, consider the coolant temperature gauge rapidly climbing towards boiling point. What is otherwise a rather much-neglected indicator suddenly becomes relevant due to its rapid change. These two scenarios describe the two typical sources of saliency: anomalies and trends. What is salient what we ought to pay attention to is what is out of the ordinary, or what is changing rapidly.</p><p>SABRE creates secondary variables (likelihood of anomaly and trend of change) to discern the overall saliency of an indicator. Where hundreds of KPIs and gauges compete for the viewer’s attention, SABRE prioritizes these, and determines a relative rank.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DciROYOqHeJ9xLM66hvyfA.png" /><figcaption>The SABRE architecture: a saliency model ranks and assesses KPIs and curates the dashboard automatically.</figcaption></figure><p>SABRE is quintessentially a Bayesian metric. As such, it is amenable to updating its priors according to evidence. These can be gradual changes: for instance, as the likelihood of anomaly increases for a metric, it would typically slide upwards, whereas if an indicator were to regress towards normal, non-anomalous behavior, it would eventually slide lower. However, as an open Bayesian metric, it is possible to include other forms of evidence. For instance, users may provide feedback on the metrics they provide, focusing on metrics they deem important, which would be prioritized in future representations.</p><p><strong>Creating data experiences</strong></p><p>Beyond dashboards, data visualization will focus on data experiences. The experiential approach to data means that dashboards are going to be only one of many ways to interact with data. Clear-text queries, voice recognition and even immersive experiences such as those leveraging AR/VR, will emerge as other ways to interact with information. No longer constrained by the fundamentally format-dependent genre of dashboards, data experiences can take any shape perhaps beyond data visualization as we may think today.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MLMUxLuab-8NregvKFvV5A.jpeg" /><figcaption>Data visualization doesn’t have to be a dashboard. Immersive data experiences using AR and VR can create deep interactions with data, such as this demo by <a href="https://virtualitics.com/">Virtualitics</a>.</figcaption></figure><p>Creating data experiences is fundamentally going to be an emergent art of turning saliency-ranked data items, such as KPIs, into an explorable, interactive experience. As such, saliency-based data experiences do not replace dashboards with another form of data visualization. Rather, it challenges the very paradigm on which the philosophy of dashboards rests, and by which dashboards are constrained (e.g. two-dimensionality, pre-definition). Beyond dashboards lies a world in which data is not something that is provided in a pre-configured structure, but rather something that is conveyed as an informed response to a system’s inferences about what a decision-maker might need, or want, to know.</p><p>And the world is ready for it.</p><p>REACH OUT TO US <a href="https://starschema.com/contact">HERE</a> TO LEARN MORE:</p><p><a href="https://starschema.com/contact">Contact</a></p><p>READ MORE STORIES FROM STARSCHEMA:</p><ul><li><a href="https://medium.com/starschema-blog/data-science-trends-to-rule-2022-68ab4c93cd2">Data Science Trends to Rule 2022</a></li><li><a href="https://medium.com/starschema-blog/8-best-practices-for-working-with-your-data-science-vendor-from-data-scientists-4e07ca4d3181">8 Best Practices for Working with Your Data Science Vendor — from Data Scientists</a></li><li><a href="https://medium.com/starschema-blog/scalable-machine-learning-in-snowflake-distributed-python-from-snowflake-sql-using-bodo-496e13fa9c3e">Scalable Machine Learning in Snowflake: Distributed Python from Snowflake SQL using Bodo</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=81dda9a0a7e7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/as-we-may-see-the-world-after-dashboards-81dda9a0a7e7">As We May See: The World after Dashboards</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Science Trends to Rule 2022]]></title>
            <link>https://medium.com/starschema-blog/data-science-trends-to-rule-2022-68ab4c93cd2?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/68ab4c93cd2</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[modern-data-stack]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[augmented-analytics]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Thu, 09 Dec 2021 09:20:33 GMT</pubDate>
            <atom:updated>2021-12-09T09:20:33.264Z</atom:updated>
            <content:encoded><![CDATA[<p>2021 was an exciting year for data science: despite COVID–19 pandemic-related layoffs and budget reductions, the field continued to flourish. <a href="https://www.anaconda.com/state-of-data-science-2021">According to a survey commissioned by Anaconda, only 37% of companies decreased their investment in data science</a>. To the vast majority of companies, data science emerged as the pre-eminent tool to survive, and thrive in, the pandemic.</p><p>2022 promises to be no less interesting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g4CbTeGzX1uxJM-m" /><figcaption>Photo by <a href="https://unsplash.com/@merittthomas?utm_source=medium&amp;utm_medium=referral">Meritt Thomas</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>TinyML</h3><p>Massive models trained on millions of instances like GPT–3 or <a href="https://www.openai.com/blog/dall-e/">DALL-E</a> might grab the headlines, but TinyML is on the rise. Simply put, TinyML is the long-awaited fusion of embedded systems with machine learning. The IoT paradigm has largely relied on raw data from edge devices, from smartwatches to electricity meters, being shuttled to large conventional servers that would then execute complex machine learning algorithms. However, over the last few years, the cost (and size) of processing power has rapidly decreased, while the cost of data transfer has remained largely the same. TinyML is a natural answer to computationally expensive models.</p><p>Bigger is not always better when it comes to models. A low-power, low-latency model run on an edge device might be a better choice when data transfer is costly or difficult (e.g. due to a lack of cellular or wired networking in the area), a rapid response is desirable and the models can be reduced to a relatively small size. A trail camera used by wildlife researchers to photograph a particular species does not need to have a state-of-the-art deep learning image recognition model onboard — but it does need to be able to operate in austere settings for a prolonged period of time. Similarly, devices used in predictive maintenance and anomaly detection for e.g. oil pipelines or overland high-voltage networks often have to operate outside the boundaries of ubiquitous wireless connectivity. TinyML is a trend that evolved in response to these challenges.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ICnLXmBoqMx02mrw" /><figcaption>Photo by <a href="https://unsplash.com/@jorgedevs?utm_source=medium&amp;utm_medium=referral">Jorge Ramirez</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>At the heart of TinyML is a growing need for machine learning developers to understand their underlying hardware. Where resources are effectively unlimited, there is little need to budget computational power. However, when a model needs to run on a dime-sized microcontroller with 256Kb (yes, kilobytes — that’s not a typo!) of RAM, developers need to be closer to the metal and have a good understanding of the power and resource implications of just about every line of code.</p><p>Fortunately, the TensorFlow developers have picked up on the potential of TinyML, and created TensorFlow Lite, which can compress and optimize models to run on 8-bit integers. As edge devices become ubiquitous, from “smart” kitchen appliances to live anomaly detection and monitoring of unmanned industrial facilities, the TinyML paradigm will gain ground. Data scientists might initially find the learning curve of hardware and low-level software daunting, but the increasing proliferation of TinyML frameworks is bound to give this domain a significant uplift in 2022.</p><h3>AI as a Service (AIaaS)</h3><p>In November 2021, <a href="https://venturebeat.com/2021/11/18/openai-makes-gpt-3-generally-available-through-its-api/">OpenAI made waves by announcing that their Transformer language model, GPT–3, would be made available as an API to the general public</a>. This is but the latest in a growing trend to provide cutting-edge models as services. The future of AIaaS will be characterized by composition of atomic AI services. A bank may use one service to build a chatbot to whom customers can report fraudulent credit card charges and use a different service for anomaly detection, risk-scoring customers’ reports. Meanwhile, a clinical language model would ingest a patient’s record and history, and use a conversational language engine from a different supplier to warn her about drug interactions. With a growing number of composable ‘domain expert’ AI models, users can create complex algorithms that merge domain-specific best-of-breed tools.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kcR7K0Osr6NdF2iV" /><figcaption>Photo by <a href="https://unsplash.com/@amayli?utm_source=medium&amp;utm_medium=referral">Amélie Mourichon</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>AIaaS is not without its challenges: in particular, enterprises need to carefully vet their prospective service providers for reliability and security. Data privacy concerns, too, may factor into managing the risk of reliance on a third party. An enterprise that provides a solution may or may not be legally liable if an API provider is breached, but the potential for reputational harm is significant. Enterprises who consider using AIaaS APIs in a customer-facing role must be ‘picky eaters’, doing assiduous due diligence on prospective API providers, obtaining ironclad SLA warranties and ensuring they are sufficiently indemnified.</p><p>Regulated industries (e.g. banking, healthcare) must meet high burdens of compliance, which might obviate the benefits of AIaaS. For enterprises that can discharge their compliance obligations and manage the risk of relying on third parties for potentially significant customer-facing AI products, however, AIaaS is an excellent way of rapidly building AI-driven solutions without the upfront expenditure traditionally associated with an in-house AI team. The future looks bright for AIaaS in 2022, and we are likely to see household names and legacy enterprises leverage AI through AIaaS solutions.</p><h3>AutoML</h3><p>Machine learning is complex, with a steep learning curve and an often resource-intensive business model — and that is unlikely to change in 2022. However, automating machine learning (AutoML) might present a solution for both of these problems. By enforcing ‘blueprints’, best practices can be baked into the analytics pipeline from the beginning, thereby preventing users from straying into the numerous pitfalls of machine learning. In addition, automation may also reduce cost-to-solution and time-to-solution by reducing the need for specially trained and expensive machine learning specialists.</p><p>The name is slightly misleading — mature AutoML solutions manage not only the machine learning part but also data preparation and preprocessing, as well as model selection and hyperparameter tuning. The leading AutoML solutions are even capable of running automated model diagnostics, advising the user as to the suitability of the final product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-5ApG-7WvIdAKPkO" /><figcaption>Photo by <a href="https://unsplash.com/@possessedphotography?utm_source=medium&amp;utm_medium=referral">Possessed Photography</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>AutoML may also be an attractive value proposition to experienced users and data scientists, who may wish to reuse a predefined analytical pathway for reproducibility or encapsulate individual subtasks to create composable arrays of tasks. In addition, AutoML can provide useful safeguards for the soundness of analytical outputs, such as identifying information leakage from training data that is not row-wise independent and identically distributed. Automating feature selection and hyperparameter search through automated hyperparameter tuning is also going to be a welcome time-saver for data scientists and non-experts alike.</p><p>While AutoML tools have been around for years, 2022 is likely to put new emphasis on such solutions to alleviate the chronic shortage of data scientists. <a href="https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/rethinking-ai-talent-strategy-as-automated-machine-learning-comes-of-age">A recent report by McKinsey</a> highlighted that money spent on pursuing scarce data science talent might be more impactful in recruiting and training competent, proficient users of AutoML. Similarly, AutoML competence is easier to acquire than the full panoply of a machine learning specialist’s skillset, allowing subject matter experts in operational areas to cross-train as AutoML users. Will 2022 be the year AutoML takes off? I certainly think so, and judging by the trend of major cloud providers as well as independent providers offering AutoML solutions, I am not alone.</p><h3>Self-service and augmented analytics</h3><p>Self-service analytics is hardly new, but its meaning is rapidly changing. From a service-oriented notion, which focused on making analytics self-service ready, leading enterprises have transitioned towards a capabilities approach: self-service is not just about providing tools like BI platforms designed for end users, but also expecting managers to become data-driven decision-makers who leverage these tools in their day-to-day duties.</p><p>Augmented analytics has emerged to bridge the service-oriented perspective with the capabilities-oriented demands. By automating the generation of insights, augmented analytics products help decision-makers navigate the data firehose and get to results faster — while also reducing the workload of expensive specialists like data scientists, who will be able to focus on higher value-added activities.</p><p>Augmented analytics often faces initial enthusiasm but also deep-seated uneasiness from prospective users. Behind this are equal parts pride and prudence: asking seasoned managers to rely on an algorithm to curate information for them is a tall order. Even executives who do not delude themselves into believing they can navigate the vast sea of enterprise data better than an algorithm feel apprehensive at the prospect of turning over data curation to artificial intelligence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oxBVgptLHBzJ5d8DOLblhA.jpeg" /><figcaption>Photo by <a href="https://www.flickr.com/people/telstar/">Telstar Logistics</a> on <a href="https://www.flickr.com/">Flickr</a></figcaption></figure><p>Successful augmented analytics solutions act like a fighter jet’s heads-up display (HUD): they allow situational awareness while preventing information overload, prioritizing data points by impact and saliency — all the while taking care not to obscure the totality of information (“non-limiting curation”). Augmented analytics must also take into account the various ways in which we prefer to ingest information — natural language generation (NLG), for instance, is a valuable tool in conveying information to verbal learners, while visual learners will prefer sparklines and other visualizations. As the interest in augmented analytics continues to rise, a growing number of solutions are incorporating a sound understanding of sensory psychology and best practices for communicating complex information. 2022 might well be the year augmented analytics becomes a household word among leading enterprise customers.</p><h3>Data marketplaces and exchanges</h3><p>If data is the new oil, data marketplaces are the new commodity exchanges. Just a few years ago, companies used to own, and jealously guard, their data. With the rise of simple, convenient platforms to share and monetize large data sets, such as Snowflake’s Data Marketplace, any company can refashion themselves into a data provider. The traditional model, whereby an intermediary buys data from originators and transforms it into an analytics-ready format, is on the wane. This is bad news for intermediaries, who will have to fight hard to demonstrate their added value, but good news for everyone else: service providers who have data on customer behavior will be able to create a lucrative side business selling analytical outputs. What you listen to, what you look at before you buy, where you order your dinner from and what weather phenomena get you to take a cab are now valuable commodities in the hands of service providers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*T55aWaGDCfZIcfE2" /><figcaption>Photo by <a href="https://unsplash.com/@changlisheng?utm_source=medium&amp;utm_medium=referral">Lisheng Chang</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>To succeed, they must balance their economic interests with privacy and legal rules on customer confidentiality, lest their new side business destroys the reputation of their overall business. Equally, they must become responsive to a new class of clients and expend some effort up-front to crunch their data into a digestible, analytics-ready format. Data exchanges, such as Snowflake’s Data Marketplace, can be an invaluable tool in the process.</p><p>Companies that wish to take the benefit of the new marketplaces in data need a solid data monetization strategy. This must deal with the entire process, start to finish: from the legal ramifications and privacy obligations to building internal CI/CD platforms for data that automates transformations into an analytics-ready format without committing specialist resources. Once adequately developed, however, a data monetization strategy can continue to reap benefits with minimal additional effort. For enterprises that generate large datasets (or highly unique ones) as part of their business, exploring data sharing might well be worthwhile.</p><p>2022 is shaping up to be an eventful year in data science, machine learning and artificial intelligence. The pandemic has refocused attention on advanced analytics as a powerful tool to manage uncertainty and respond to rapidly shifting strategic situations. The rise in AIaaS will contribute to the ubiquity of AI-driven solutions, which will now become feasible for a much wider range of markets. Similarly, AutoML and augmented analytics is gradually turning data science from a job description to a job skill — one that will be indispensable for managers seeking advancement. And 2022 might be the year when the oft-cited adage about data being the New Oil would come closer to reality with the first ‘commodity exchanges’ for data.</p><p>REACH OUT TO US <a href="https://starschema.com/contact">HERE</a> TO LEARN MORE:</p><p><a href="https://starschema.com/contact">Contact</a></p><p>READ MORE STORIES FROM STARSCHEMA:</p><ul><li><a href="https://medium.com/starschema-blog/8-best-practices-for-working-with-your-data-science-vendor-from-data-scientists-4e07ca4d3181">8 Best Practices for Working with Your Data Science Vendor — from Data Scientists</a></li><li><a href="https://medium.com/starschema-blog/scalable-machine-learning-in-snowflake-distributed-python-from-snowflake-sql-using-bodo-496e13fa9c3e">Scalable Machine Learning in Snowflake: Distributed Python from Snowflake SQL using Bodo</a></li><li><a href="https://medium.com/starschema-blog/covid-19-and-the-first-war-of-data-science-980798f075ef">COVID-19 and the first war of data science</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=68ab4c93cd2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/data-science-trends-to-rule-2022-68ab4c93cd2">Data Science Trends to Rule 2022</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The COVID Tracking Project is shutting down in a week. What next?]]></title>
            <link>https://medium.com/starschema-blog/the-covid-tracking-project-is-shutting-down-in-a-week-what-next-5c791cebe451?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c791cebe451</guid>
            <category><![CDATA[covid19]]></category>
            <category><![CDATA[snowflake-computing]]></category>
            <category><![CDATA[pandemic-response]]></category>
            <category><![CDATA[covid-19-data]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Mon, 01 Mar 2021 19:17:21 GMT</pubDate>
            <atom:updated>2021-03-01T19:17:21.292Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lDv-mg2dEghE8NSMNRVkGA.jpeg" /></figure><p>The <a href="https://covidtracking.com/">COVID Tracking Project</a> has been one of the most successful citizen-driven data collection projects in history. Driven by The Atlantic and supported by an army of volunteers, it has collected the nuggets of information about testing and case counts, often beating federal and state authorities to the race. Yet sustaining such a project over the long run, especially when primarily driven by volunteer engagement, is quite difficult. And so, after a year, the COVID Tracking Project is shutting down on 07 March.</p><p><a href="https://covidtracking.com/analysis-updates/covid-tracking-project-end-march-7">Analysis &amp; updates | It&#39;s Time: The COVID Tracking Project Will Soon Come to an End</a></p><p>The good news is that if you have been accessing the COVID Tracking Project’s data via the Starschema COVID-19 Data Set, whether through Snowflake Data Marketplace or through the flat files shared via AWS S3, you should see replacement tables emerging that cover the same ground. Below is a brief overview of data sources that you can use to replace data from the COVID Tracking Project.</p><h4>Case counts and bed utilization</h4><p>We have integrated several data products from the US Department of Health and Human Services and the CDC to provide information about case counts and healthcare resource utilization. We provide the following data sets:</p><ul><li>CDC_INPATIENT_BEDS_ALL: all-cause inpatient bed usage, forecast (including upper and lower bounds of the 95% confidence interval)</li><li>CDC_INPATIENT_BEDS_COVID_19: inpatient bed usage for COVID-19 cases, forecast (including upper and lower bounds of the 95% confidence interval)</li><li>CDC_INPATIENT_BEDS_ICU_ALL : all-cause intensive care bed usage, forecast (including upper and lower bounds of the 95% confidence interval)</li><li>CDC_REPORTED_PATIENT_IMPACT : reported (actual) patient impact data, including critical care availability, admissions for suspected and confirmed COVID-19 by paediatric vs adult cohort and a number of other indicators</li></ul><h4>Policy measures</h4><p>A new table, CDC_POLICY_MEASURES, includes highly granular data on a range of policy measures for counties and states, including:</p><ul><li>stay-at-home orders,</li><li>large gathering bans,</li><li>restrictions on specific industry sectors and venues, and</li><li>non-essential workers legislation.</li></ul><p>This supplements our table KFF_US_STATE_MITIGATIONS, which only lists state-level measures. Both tables are updated ‘as needed’, but are constantly monitored for changes.</p><h4>Testing and diagnostics</h4><p>The CDC_TESTING table contains information about tests conducted, including positive, negative and pending/inconclusive tests. Even as the emphasis of public health interventions moves from testing to prophylaxis by way of vaccines, diagnostic testing remains an important bellwether for pathogenic dynamics. Like all CDC data, this is updated every day, although there may be an up to 24 hour lag due to reporting times by county, state and territorial health authorities.</p><h4>Vaccination</h4><p>With the availability of two vaccines on the US market and the recent positive endorsement of the Johnson &amp; Johnson single-dose vaccine by ACIP, the spotlight is now on vaccine penetration––the ratio of the population who have been vaccinated as a fraction of the entire population.</p><p><a href="https://www.washingtonpost.com/health/2021/02/26/johnson-and-johnson-vaccine-fda/">FDA advisers recommend authorization of Johnson &amp; Johnson single-dose coronavirus vaccine</a></p><p>It is crucial to track vaccination rates for a thorough assessment of a population’s at-risk status. The OWID_VACCINATIONS table provides detailed information on vaccine allocation, vaccines administered and, most importantly, the number of persons who have received two doses of the currently marketed two-dose vaccines (both Pfizer and Moderna), who can be deemed to be immune.</p><h4>Table updates</h4><p>We have also made some back-end updates to tables. In particular, if you have been using WHO data, this should come as good news to you––because of the frequent format changes, WHO data was from time to time less reliable. This has now been fixed, and you can look forward to a steady stream of the WHO’s regular reports.</p><h3>Next steps</h3><p>If you have been using the COVID Tracking Project data set to power your visualizations and analytics, you may have to make some changes. We have endeavoured to bring these changes to you in due time to allow you to make the changes you need. As always, if we can assist you in any way to make the most out of the Starschema COVID-19 Data Set, please <a href="https://starschema.com/contact">don’t hesitate to let us know</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c791cebe451" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/the-covid-tracking-project-is-shutting-down-in-a-week-what-next-5c791cebe451">The COVID Tracking Project is shutting down in a week. What next?</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fighting the COVID-19 pandemic with data and context]]></title>
            <link>https://medium.com/starschema-blog/fighting-the-covid-19-pandemic-with-data-and-context-74da2d77364c?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/74da2d77364c</guid>
            <category><![CDATA[covid19]]></category>
            <category><![CDATA[open-data]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[coronavirus]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Wed, 01 Apr 2020 20:22:12 GMT</pubDate>
            <atom:updated>2020-04-01T20:23:02.310Z</atom:updated>
            <content:encoded><![CDATA[<p>The COVID-19 outbreak is in many ways an outlier. It emerged with unusual speed, spread rapidly throughout the globe and has elicited a public health response that is unprecedented in recent history. Yet this time, humanity faces a pandemic with a wealth of information that would have been unimaginable just a few brief years ago. Writing for ZDNet, <a href="https://www.zdnet.com/article/as-covid-19-data-sets-become-more-accessible-novel-coronavirus-pandemic-may-be-most-visualized-ever/">Larry Dignan referred to this outbreak as the most visualized pandemic ever</a>, and the sheer number of data sets available — some more accurate, timely and reliable than others — to the general public is staggering. Yet in data, as in all narrative endeavors, context is everything.Each new data source we add to the <a href="https://starschema.github.io/COVID-19-data/">free-of-charge, public data set</a> provides additional context. Each new source is vetted for accuracy. The data is collated, unified and normalized. Internal keying around the same widely used conventions — e.g. normalization of country names to ISO 3166–1 names and country codes — facilitates going beyond individual data sources and enables users to extract meaning and information from the raw data itself regardless of which data set it came from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DIOMZhVAwuym8oFbiv8ceQ.png" /><figcaption><a href="https://public.tableau.com/profile/chrisvoncsefalvay#!/vizhome/Book2_15855891386900/U_S_CasesperICUbed">Choropleth map of U.S. counties by relative ICU demand from COVID-19</a>, calculated as the number of cases per fully staffed ICU beds (2018).</figcaption></figure><p><a href="https://public.tableau.com/profile/chrisvoncsefalvay#!/vizhome/Book2_15855891386900/U_S_CasesperICUbed?publish=yes">Our recent visualization illustrating the cases as a function of staffed intensive care unit (ICU)</a> is an example of this context. Only data unified along the same primary keys and carefully curated for data quality can create the contextual understanding that is needed in this situation. By seeing cases in the context of the underlying healthcare system, public health planners can anticipate whether healthcare capacity will be exceeded in a particular location and enterprises can track the movement of COVID-19, assess supply chain risks and evaluate strategic options.As an epidemiologist, I know first-hand that a pandemic can often feel like a sweeping turmoil, and curated, cleaned data is indispensable for data-driven decision-making. Models and decisions are only as good as the data supporting them.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fj4jRa0cZKFI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dj4jRa0cZKFI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fj4jRa0cZKFI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fd7f6c9903a86f3fb37b4870c49ba05c/href">https://medium.com/media/fd7f6c9903a86f3fb37b4870c49ba05c/href</a></iframe><p>The data set is freely available, and accessible <a href="https://starschema.github.io/COVID-19-data">via S3</a> as well as on <a href="https://www.snowflake.com/datasets/starschema/">Snowflake Data Exchange</a>. We hope we can contribute by providing more information and clarity over merely providing more data. In the end, it is that information and clarity that will make all the difference.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74da2d77364c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/fighting-the-covid-19-pandemic-with-data-and-context-74da2d77364c">Fighting the COVID-19 pandemic with data and context</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[COVID-19 and the first war of data science]]></title>
            <link>https://medium.com/starschema-blog/covid-19-and-the-first-war-of-data-science-980798f075ef?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/980798f075ef</guid>
            <category><![CDATA[dataset]]></category>
            <category><![CDATA[covid19]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[coronavirus]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Wed, 18 Mar 2020 17:13:14 GMT</pubDate>
            <atom:updated>2021-12-07T11:43:19.564Z</atom:updated>
            <content:encoded><![CDATA[<h3>COVID-19 and the First War of Data Science</h3><p>In the subtitle of his remarkable history about the race for the nuclear bomb, science writer and historian of science Jim Baggott referred to World War II as the “first war of physics”.</p><p>Today, the efforts waged to curb the COVID-19 pandemic may be the first example of a large-scale, global data-driven response to a worldwide crisis, and as such perhaps the first war of data science.</p><p>It is difficult to overstate just how much data has become available in an extremely short time, and open science and the networks for sharing clinical and epidemiological data have enabled an unprecedented depth of analytics. We have the data, we have the tools, and we have experts. They are working hard, but we’ve never done this before and haven’t trained for it. While much progress is being made, we still must overcome many challenges.</p><h3>The Starschema COVID-19 dataset</h3><p>One of the tactical challenges lies in making the data “analytics ready” — ensuring the data is accurate, readily available, constantly updated, and in a format that can be easily used by data scientists on the front lines.</p><p>The case count dataset collated by Johns Hopkins University’s Center for Systems Science and Engineering (JHU CSSE) is referenced and relied on by hundreds if not thousands of data scientists. This dataset, and the dashboard JHU CSSE provides, is very important and helpful, but it wasn’t constructed in a way that makes the data analytics-ready and often data scientists who want to use the dataset need to go through a time-consuming process to unpivot, union and clean the data before they can use it in their own models and applications.</p><p>Based on the JHU CSSE data stream, Starschema has built — and will continue to update and improve — <a href="https://github.com/starschema/COVID-19-data">an analytics-ready dataset</a> that draws on multiple data sources, eventually integrating high temporal resolution domestic data (e.g. <a href="https://github.com/pcm-dpc/COVID-19">the dataset provided by Italy’s Department of Civil Protection</a>). The entire dataset is <a href="https://s3-us-west-1.amazonaws.com/starschema.covid/JHU_COVID-19.csv">available for download via AWS S3</a>, as well as via <a href="https://www.snowflake.com/datasets/starschema/">Snowflake‘s Data Exchange</a>. This will enable data scientists, epidemiologists and analysts to access the most up-to-date data on COVID-19 cases through a cloud-based data warehouse, including datasets enriched with relevant information such as population densities and geolocation data. Users can leverage this information to build inferential models of disease propagation and provide their customers with unique insights into the behavior of this novel epidemic.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V_jqz5vCAgjhbFDuISviNw.png" /><figcaption>The Starschema COVID-19 data set on Snowflake’s Data Exchange</figcaption></figure><p>The Starschema COVID-19 dataset is in a classical “long” format — each permutation of a point in time and a case type — confirmed, recovered, deceased, active — is a row. In addition, many of the inconsistencies between county-level reporting — prior to 10 March 2020 — and state-level reporting have been resolved.</p><p>We are in the process of collating all data pertaining to COVID-19 and making additional data available, such as information on population size, population density, and other metrics that enable contingency planning, forecasting, and visualization. As the data size expands, we are constructing an <a href="https://github.com/starschema/COVID-19-data/wiki/dynamic-dag">Airflow-based dynamic DAG</a> in order to provide a reproducible workflow for rapid ingestion and transformation of data.</p><h3>The outlook</h3><p>As the COVID-19 pandemic progresses, we can expect data to play an increasing role in both public and private operations. Public health authorities are already using data-driven approaches to monitor the spread of SARS-CoV-2, and phylogenetic analyses are used to identify whether particular strains of SARS-CoV-2 carry a higher risk. In the private sector, information on the number of cases is used to support business contingency operations and analyze supply chains for possible vulnerabilities.</p><p>With the increasing importance of data, ensuring data quality and consistency is paramount. Through this dataset Starschema intends to provide a practical demonstration of best-of-breed data quality assurance and data management procedures to provide public health professionals, contingency planners, and enterprises with an outline of data management sound enough to stake lives on.</p><p>As we fight SARS-CoV-2, data-driven approaches may well be what gives humanity an edge. With the rapid expansion and democratization of data and advanced analytics, data science is in a unique position to bring the tools, techniques, and procedures that were developed in the analytics domain over the last decade to bear on this unprecedented challenge.</p><p>If you have any questions regarding this dataset or how to use it, <a href="/service/mailto:csefalvayk@starschema.net">please reach out</a>. We are here to help.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=980798f075ef" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/covid-19-and-the-first-war-of-data-science-980798f075ef">COVID-19 and the first war of data science</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Predictive maintenance helped win a war. Now, it can help you outpace the competition.]]></title>
            <link>https://medium.com/starschema-blog/predictive-maintenance-helped-win-a-war-now-it-can-help-you-outpace-the-competition-597ebfeab546?source=rss-93652c542fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/597ebfeab546</guid>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[industry]]></category>
            <category><![CDATA[predictive-maintenance]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Chris von Csefalvay]]></dc:creator>
            <pubDate>Mon, 09 Sep 2019 09:14:17 GMT</pubDate>
            <atom:updated>2019-09-09T10:04:33.654Z</atom:updated>
            <content:encoded><![CDATA[<p>The year is 1943. Britain’s survival still hangs by a thin, precarious thread, despite America joining the war effort. Just a few months ago, in February 1942, two German battleships, the <em>Scharnhorst</em> and the <em>Gneisenau</em>, alongside the heavy cruiser <em>Prinz Eugen</em>, and their escort fleet executed Operation Cerberus — a daring raid passing through the English Channel from their French port of Brest at the western cape of Normandy. Britain is not merely humiliated — not even the Spanish Armada has been able to cross the Channel in 1588 — , it is also existentially threatened. If it cannot protect its coastal waters, it may lose the Battle of the Atlantic, and with that the transatlantic lifeline of shipping that has been keeping it in the fight. If not for regular transatlantic convoys, carrying anything from tanks to milk powder, Britain would have been under immense pressure to surrender amidst starvation and undersupply. And now, the Channel Dash, as the raid in 1942 will be known, has shown it cannot even keep its own littoral waters safe.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ael6qDFJ7USmnOmYzGVztw.jpeg" /><figcaption>Aircrew of a Handley Page Halifax Mk III bomber of Squadron 502, RAF Coastal Command, with their aircraft. The Halifax Mk III played a decisive role in keeping Britain’s littoral waters safe, and was the mainstay of Allied anti-submarine warfare in the Atlantic until the first radar-equipped Consolidated PBY Catalina patrol bombers entered service later in the war.</figcaption></figure><p>Along the Royal Navy, the responsibility for that lay with Coastal Command, the junior home command of the Royal Air Force (alongside the Bomber and Fighter commands), trying to provide anti-submarine warfare capabilities when anti-submarine warfare was still a novel art and its mainstay tools, radar and sonar (then known as ASDIC), were nascent at best. They were also plagued by equipment issues, with the Coastal Command’s B-24 Liberator anti-submarine bombers spending more time in repair than flying — in some squadrons, at any given time, less than half of the aircraft were operational, the rest awaiting scheduled or unscheduled maintenance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gWB9Z2RcsK0tbc1skd2tfg.png" /><figcaption>C.H. Waddington (1905–1975) — developmental biologist, embryologist, paleontologist, poet, painter and wartime operations researcher for the RAF Coastal Command. His observations that led to the eponymous Waddington effect were the beginning of condition-based maintenance (CBM), the precursor of predictive maintenance.</figcaption></figure><p>At their headquarters in Northwood, the Operational Research (OR) unit of the Coastal Command turned its attention to this very problem. Its leader, C.H. Waddington, was an unlikely choice — a distinguished developmental biologist, poet and artist, he was drafted into service to oversee the Coastal Command’s OR, despite having no experience with aviation or aeronautical engineering. It was here that he made his greatest contribution to the war effort — his discovery of the Waddington effect. Paradoxically, he noted, unscheduled repairs were <em>more</em>, not <em>less</em>, frequent after scheduled maintenance.</p><p>This flew in the face of conventional wisdom, which was that frequent scheduled preventive maintenance, every 50 flight hours, would keep the squadron flying. This was regardless of whether the aircraft needed it or not, often putting perfectly airworthy aircraft out of commission for unnecessary scheduled maintenance. Waddington didn’t merely say scheduled maintenance was useless, he went further — it could, he writes in his diaries (declassified in the 1970s and published as <a href="https://books.google.hu/books/about/O_R_in_World_War_2_operational_research.html?id=6SEhAQAAIAAJ&amp;redir_esc=y"><em>O.R. in World War II</em></a>),</p><blockquote><em>…increase breakdowns, and this can only be because it is doing positive harm by disturbing a relatively satisfactory state of affairs. There is no sign that the rate of breakdowns is starting to increase again after 40–50 flying hours when the aircraft is coming due for its next scheduled maintenance.</em></blockquote><p>His thinking, however counter-intuitive, has been hugely successful. An early version of condition-based maintenance (CBM), where aircraft underwent repairs based on reported issues and a quick post-flight examination resulted at a 60% increase in effective flying hours by the Coastal Command. Recognising the counterintuitive yet vital Waddington effect helped secure Britain’s coastal waters and with that, played a significant role in the Allied victory in the Battle of the Atlantic.</p><p>The rest, as they say, is history.</p><p>Predictive maintenance helped win a war. Today, it’s helping manufacturing companies outpace their customers. As the keystone of the Industry 4.0 stack, interest in predictive maintenance is almost ubiquitous among the entire manufacturing spectrum. <a href="https://news.thomasnet.com/companystory/downtime-costs-auto-industry-22k-minute-survey-481017">A 2005 Nielsen survey of automotive manufacturers estimated the cost of unplanned downtime to be around $22,000 per minute</a> — up to $50,000. <a href="https://www.aberdeen.com/techpro-essentials/stat-of-the-week-the-rising-cost-of-downtime/">A more recent figure by Aberdeen puts this, across all industries, to $260,000 per hour on average</a>. More worryingly, this figure is rapidly climbing, rising by 60% just over two years as manufacturing processes became increasingly dependent on sophisticated, complex machinery that often requires specialist off-site services to fix. No doubt with the increasing use of lean factories and just-in-time manufacturing, this figure is bound only to skyrocket. Adding the risk that catastrophic unplanned failures pose to high-value capital equipment and the potential human cost, unplanned breakdowns have earned their place as one of the top concerns of manufacturing executives, over half of whom have named unplanned downtimes as their #1 concern.</p><p>Predictive maintenance goes counter to two extremes — reactive maintenance (‘if it ain’t broke, don’t fix it!’) and scheduled maintenance, which Waddington so astutely proved to be worse than useless. It could be better described as ‘data driven maintenance’: failure risk is calculated based on known indicators associated with failure events, and maintenance is initiated once so-called prodromic patterns — signs of still-normal operation that however indicate an impending breakdown — are detected. This allows for a controlled shutdown, safe operating procedures and a degree of flexibility in repair. Between under- and overmaintenance, predictive maintenance offers the golden middle road of ‘just right’. And with the ubiquity of unplanned downtime — <a href="https://www.servicemax.com/unplanned-downtime">according to a Vanson Bourne study of 450 field service decision makers across a range of industries, 82% of companies have suffered at least one unplanned outage over the preceding three years</a> — , getting predictive maintenance right can be the keystone of outpacing the competition and operating a high-performance business: whether it’s manufacturing, operating costly capital assets like oil rigs or safety-critical operations like healthcare equipment or nuclear power plant, predictive maintenance must be a strategic cornerstone of all operations where there’s just no room for error.</p><p><a href="http://starschema.com/"><em>Starschema</em></a><em>’s data scientists have decades of combined experience with predictive maintenance tasks, including designing the best predictive maintenance strategy for your business regardless of your operations. With an in-depth understanding of not just the disciplines of operations research and data science but also of sensorics, equipment communication and industrial IoT, the Starschema team brings a powerful integrated approach to your maintenance tasks. </em><a href="https://starschema.com/kb/predictive-maintenance-in-manufacturing"><em>You can download our white paper on predictive maintenance in manufacturing for free</em></a><em>, which outlines our approach, methodology and track record of successful engagements.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=597ebfeab546" width="1" height="1" alt=""><hr><p><a href="https://medium.com/starschema-blog/predictive-maintenance-helped-win-a-war-now-it-can-help-you-outpace-the-competition-597ebfeab546">Predictive maintenance helped win a war. Now, it can help you outpace the competition.</a> was originally published in <a href="https://medium.com/starschema-blog">HCLTech-Starschema Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>