back to article IBM Watson Health cuts back Drug Discovery 'artificial intelligence' after lackluster sales

IBM Watson Health is tapering off its Drug Discovery program, which uses "AI" software to help companies develop new pharmaceuticals, blaming poor sales. The service isn’t completely shutting down, however. IBM spokesperson Ed Barbini told The Register: “We are not discontinuing our Watson for Drug Discovery offering, and we …

  1. I.Geller Bronze badge

    1. Does IBM train data on dictionary definitions or terabytes of data?

    - If on dictionary definitions a user does not need a centralized IBM Watson computer and can structure data, create his profile and search on his computer. In addition, with the dictionary definitions an IBM product begins to search for information based on the information true meaning, becomes a true AI and truly learns new (adds structured texts/ Machine Learning). I.e. the dictionary definitions become subtexts, identify the unique meanings of data patterns' words (each pattern has a few words) through the creation of long tuples. (In mathematics, a tuple is a final ordered list (sequence) of elements. In AI case it's a list of clustered synonymous phrases.)

    - If IBM trains data on texts then there is a purely mechanical comparison of pieces of text (contexts), without the understanding of the words' meanings.

    2. The second and extremely important problem - the quality of the data which IBM Watson uses: it must have sufficiently large texts and, if possible, annotated with other texts - AI needs at least a few large paragraphs, the more the better. Thus, we are talking again about the words’ meanings by their unique dictionary definitions/ subtexts - by comparison with the surrounding/ annotating text. (Plus other texts (as subtexts) provide examples how the words are used/ the same training on external data.)

    In order to find a unique dictionary definition for a word is necessary to know surrounding it words' contexts and subtexts - see my patents, the only thing I could publish, sorry.

    Or IBM should create an extremely highly qualified AI that would understand the entire pharmacological information, which - at the moment - is absolutely not realistic.

    3. It's more pragmatic to appeal AI technology primarily to areas where there is a lot of text and money. For example, to jurisprudence?

    But IBM is not trying to create a Patent Search! Why? Because IBM does not annotate words by their dictionary definitions, does not search for their meanings - and annotates patterns entirely, not seeing that they consist of many words. IBM cannot search for patents! IBM cannot search at all!

    So IBM is using the wrong technology, and the fact that IBM has abandoned pharmacology does not mean that it is not the gold bottom in the right hands and with the right AI technology.

    4. IBM is a giant on clay feet ... Not surprising that IBM will not be able to achieve commercial success and suffers losses. Because of IBM's refusal to use dictionary definitions IBM is losing the market. The evidence - its failing Debater project.

    1. I.Geller Bronze badge

      I know what I'm talking about, I was at NIST TREC QA and searched for a phrase in 6+ million texts. Without subtexts, dictionary definitions - forget about it!

    2. Bilious

      Software to analyse the potential for interaction between a drug candidate and selected receptors or enzymes have been in use for a long time - both for finding substances that may have a desired effect, and for weeding out substances with properties that will lead to failure in clinical trials or clinical use. Newer AI systems will have to deliver additional value beyond the expert knowledge of those already in the field. A general tool written by generalists would seem to have a slim chance of being of general use across all those specialist sub-fields within drug discovery - each one requiring biological and technical knowledge way beyond textbooks or published articles.

      The market has spoken: The tool wasn't relevant.

      1. I.Geller Bronze badge

        "...biological and technical knowledge way beyond textbooks or published articles" = extensive textual descriptions, which are either rare or don't exist now. Why they? Such the explanations have never been in demand, there was no a mechanism to structure them automatically/ AI search technology didn't exist. Instead they were manually structured (to several words and phrases, "specialist sub-fields within drug discovery") and SQL search technology was applied. No sufficient data for AI...

        The extremely highly qualified AI that would understand the entire pharmacological information and convert these short to extensive is hard to be created, though. IBM gave up.

  2. vaporland

    IBM’s Health division has been crumbling for a while.

    Last year, it laid off a large chunk of its employees after it failed to win hospital contracts. The failure was, apparently, caused by using subcontractors and cheap H1B visa holders.

    Fixed that for you.

    1. theblackhand Silver badge

      Re: IBM’s Health division has been crumbling for a while.

      "The failure was, apparently, caused by using subcontractors and cheap H1B visa holders."

      The attempt to make a failing service more cost effective after failing to deliver any measurable success (over and above competitors) in the last 20 years resulted in the use of subcontractors and H1B visa holders. And failed to deliver a different result.

      And there's the old adage...no one ever ever got fired for buying IBM. They all got fired for working at IBM...

      1. I.Geller Bronze badge

        Re: IBM’s Health division has been crumbling for a while.

        Plus the wrong technology: most likely IBM trains AI on examples from terabytes of texts that say HOW words are used, but do not explain what they mean. And so IBM is not able to create long enough tuples of structured texts, the main advantage of AI search technology is lost... IBM sees no dictionary, ignores it completely!

        1. This post has been deleted by its author

    2. bombastic bob Silver badge
      WTF?

      Re: IBM’s Health division has been crumbling for a while.

      well even if it really WAS caused by internal competition and workplace drama., who'd want to work in THAT kind of environment?

      Someone that creates internal competition and allows workplace drama needs to be FIRED INSTEAD.

      1. aj69

        Re: IBM’s Health division has been crumbling for a while.

        “Internal competition and workplace drama.“

        Sounds like a rat race.

  3. I.Geller Bronze badge

    IBM does not understand that parsing and weights are only a small part of AI technology, that although the texts created with the help of AI-parsing seem to be sufficient - it is not so, they are not sufficient though necessary! These structured texts do not explain what the words mean, but merely mechanically explain how they are used.

    Look up the definition of almost any word in a dictionary? It says what the word means and usually gives an example of how it is used. Training AI on terabytes you get only an example of how it is used! Only a portion of that dictionary definition!

    Patent Search is a litmus test and you can claim that you have AI only if you have it, because it can be done ONLY using the true the words' meanings. Which are in the dictionary. However IBM has only Augmented Intelligence, it uses only the explanatory, second part of dictionary definitions... Thus IBM abandoned pharmacology, it has no Artificial Intelligence.

    1. ATeal

      RE: No actually AI is...

      I'm getting board of bumbling around trying to sell people on functional analysis like some sort of deranged Jehovah Witness (but armed with proof)

      The problem we are all suffering from is that you can get convergence with pretty much any method, the conditions you can show sufficient for convergence are one of the rare cases where the conditions are actually really really broad (typically if you can show if you have X then it converges, X is some tiny bit of information, very specific, here X is very broad and the "NN (structure activiation function of ( sum of inputs * weights ) + bias ) )" ) (FTW (LISP))

      Nice to see some philosophy in there with schemas of knowledge, ironic you are bickering about what AI is, as if philosophy hasn't asked all those questions and formed an array of answers for us to muse on for 5 minute before thinking we've come to some deep conclusions.

      Seriously quantify more.

      Addendum: you know how people think stats can assign a number to something with perfectly describes that thing? Well polynomials are dense in continuous functions - that is to say a polynomial can get as close as you like (under a metric like the sup norm - I hate being vague I set off my own noob-dar) to any continuous function. Well floating point and the "network" structure can get pretty damn close too. So people are imbuing it with "statistics says okay" sort of reverence.

      1. I.Geller Bronze badge

        Actually AI is...

        I'm making philosophy an exact science. For example, I believe that IBM sales fell for the reason that IBM doesn't find what patterns' words mean (doesn't detect their unique dictionary definitions, as subtexts), but finds how the patterns are used (trains them on terabytes of data).

        I mean this: there are two sentences

        - Alice.

        - Alice laughs, rejoices and sings.

        The first sentence is the limit to the differential function of the second: all three phrases

        -- Alice laughs

        -- Alice rejoices

        -- Alice sings

        have subtexts, i.e. unique dictionary definitions for each word; while IBM finds in what contexts these patterns are used and doesn't see that each has two words in it.

        The first "Alice." has no a unique dictionary definition and means everything and nothing at the same time. In other words, the second's function finds its limit into the first's radically new quality; which is the basic principle of Differential Analysis - a differential function is absolutely different from its limit. (This is my Differential Linguistics.)

        Thus, not employing dictionary definitions (as subtext of words) IBM violates the Laws of Nature, rejects the rules of Differential Analysis. There lies the true reason why IBM is losing money on pharmacology, and I believe that IBM should find the way to annotate pharmacological data patterns' words first (by dictionary definitions), and make money next.

        Am I clear? Do you think I made Philosophy an exact science?

        1. I.Geller Bronze badge

          Re: Actually AI is...

          Why? I was at NIST TREC QA and found that without the definitions as subtexts nothing works, truly... I just know.

        2. GrumpenKraut Silver badge
          Facepalm

          Re: Actually AI is...

          How often do you plan to re-post this silly comment? Are you a bloody perl script?

          1. herman Silver badge
            Devil

            Re: Actually AI is...

            No, the NIST uses Expect, not Perl.

        3. IT Poser

          Re: Am I clear? Do you think I made Philosophy an exact science?

          Where you see three phases,

          -- Alice laughs

          -- Alice rejoices

          -- Alice sings

          I see two phases,

          -- Alice laughs

          -- Alice rejoices and sings

          Now riddle me this, based on the given information, do you think I think you made Philosophy an exact science?

    2. devTrail Bronze badge

      Hang on. According to the article the system failed to provide useful result to domain experts. Your post and all the other posts I read above point to supposed failures in ML/Computer Science design, but I can't believe that IBM is lacking computer science expertise. On the other hand I find more credible the idea that they failed to enroll in the project enough domain experts. After all if only after they tried to deliver the system to the customers they found out it was not useful at all it means they couldn't understand the value of the results.

      1. I.Geller Bronze badge

        I don't know either, just a guess. IBM has no patent search. Why? Patent search does not work without dictionary definitions. And IBM stopped pharmacology... I concluded that the reason is IBM does not comment on patterns' words by dictionary definitions.

        It may be a plenty of texts in there, a lot of money. Or there are no texts? I don't know, perhaps that's the problem.

        But IBM doesn't do Artificial Intelligence. Augmented only. Why? What is the difference? Dictionary definitions? IBM trains on data?

      2. I.Geller Bronze badge

        If IBM trains its patterns on textual data, it explains why Watson is a centralized computer/ why IBM doesn't comment using definitions / doesn't have Artificial Intelligence / why Watson lost the debate.

        If IBM does it has no technology, it will soon lose to Google and Amazon, which annotate using dictionary.

        I claim that AI is my theory, and the fate of my theory may depend upon the answers "Why?" and "How?"

        1. devTrail Bronze badge

          But IBM doesn't do Artificial Intelligence. Augmented only. Why? What is the difference? Dictionary definitions? IBM trains on data?

          If IBM trains its patterns on textual data, it explains why Watson is a centralized computer/ why IBM doesn't comment using definitions / doesn't have Artificial Intelligence / why Watson lost the debate.

          IBM has been working on Watson for 20 years, In the meantime a lot of research has been published. So they had plenty of time to update it.

          If IBM does it has no technology, it will soon lose to Google and Amazon, which annotate using dictionary.

          I claim that AI is my theory, and the fate of my theory may depend upon the answers "Why?" and "How?"

          Basically you repeated that it must be a problem related to ML/Computer science, but I'm still unconvinced, the main clue remains the one I mentioned above, they didn't realised that their system was underperforming until they delivered it to the customers. Since domain expertise requires a lot of experience and IBM as all the other corporations prefers to hire fresh graduates a lack of domain experts is a reasonable explanation

          1. I.Geller Bronze badge

            We both are guessing. Above I saw a version that this is the result of IBM internal strife.

  4. Schultz Silver badge
    Boffin

    Peak AI?

    Maybe we can move beyond this fashionable nonsense and start focusing on productive issues again. In my mind, AI is a typical case of practical magic. You don't have to worry about the details (the 'magic' part takes care of that), and it'll solve all your problems / make you rich and famous / get you a big budget or grant.

    Humans love this magic stuff. All the results without the hard work. So I guess we'll have a few more years before people move on to the next big thing.

    1. This post has been deleted by its author

    2. I.Geller Bronze badge
      Facepalm

      Re: Peak AI?

      Who cares about people, their desires and money, what they need or no - if AI ​​changes the status quo and many business giants could become bankrupt. We should buy what is on the shelves, not what we need!

      I patented AI 10 years ago, so what? Only some applications of AI are allowed to exist, only those that do not undermining the status quo. All giant corporations lobbied for me to remain unknown ... That's it! Have you read somewhere about who came up with AI? Not? I did.

      That is why I am raping the Internet, using all methods and means to bring technology to those who can turn it into a business.

      1. bombastic bob Silver badge
        WTF?

        Re: Peak AI?

        you patented AI ???

        1. I.Geller Bronze badge

          Re: Peak AI?

          Yes, I did. I patented AI parsing, annotation by dictionary and many other things.

          See the US PTO for Ilya Geller

          1. lowwall

            Re: Peak AI?

            Now it all makes perfect sense.

            1. I.Geller Bronze badge

              Re: Peak AI?

              It's all new... New terminology, new knowledge.

          2. Tom 7 Silver badge

            Re: Peak AI?

            But you didnt add 'on a phone'!

            1. I.Geller Bronze badge

              Re: Peak AI?

              Yes, I'm on the phone promoting artificial intelligence technology. I guess I'm the only one who insists. All the rest under the guise of AI sell stale goods on a new, catchy name. For example, IBM.

              You can't win NIST TREC QA unless you have a structured piece of text that you match against the same way structured pieces, finding the best answer to your question. You can't look for solutions for the medicinal formulas unless you found the best chunks! AI is a database technology!

              These fragments are created both as contexts and their subtexts. For example, dictionary definitions create the subtexts. You cannot train these pieces on terabytes because you never know if you missed something, but with the definitions you cannot miss a grain of information.

              This is Differential Linguistics: paragraphs are integrals, and sets of dictionary definitions are their constants. This is the basics of differential analysis! You have to add definitions, simply must and not otherwise.

              And I suspect that IBM ignores the "definition part" and, therefore, left the project "pharmacology".

  5. Tom 7 Silver badge

    I suppose that when most of your profit comes from telling people

    its really expensive to make drugs then if its cheap to make them then your whole modus operandi goes away.

  6. Blockchain commentard Silver badge
    Coat

    What if

    they'd called it Watson for Medicinal Discovery? I'm sure even Tony Blair would have gone for it !!!

  7. I.Geller Bronze badge

    I mean that your search phrase, extended with synonyms, is filtered through your profile, and a textual array of considerable size is created. That is, you are looking for information not with one or two phrases, but with hundreds and thousands; where a very significant part of the search phrases comes from subtexts, both dictionary definitions and examples.

    If you train texts in your profile only by examples and not by dictionary definitions, you will learn how patterns are used without knowing what their words mean. That is, the computer is purely mechanically, without understanding, looking for context-close texts on examples.

    For instance Google and Yandex translations. This is not a translation! This is the tragedy! The computer does not understand what and how it does!

    It seems to me that IBM Watson also does not understand what it is searching for, for example, in pharmacology. Accordingly, IBM is losing this market, as well as all others.

    Apparently I'm the only one who understands AI technology.

  8. I.Geller Bronze badge

    I wonder how much of my speculation is true?

    There is a sentence:

    "After the publication was made, the correspondence was lost."

    There are two synonyms in the sentence: "publication" and "correspondence"; and therefore "was made" and "was lost" refer to both. But IBM Watson (probably!) works with only these two whole patterns: "publication was made" and "correspondence was lost", not noticing that there are two synonymous there.

    In other words, the sentence produces four patterns - two descriptive parts refer to each of the two synonyms. However, IBM Watson (probably!) only gets two, and therefore loses two possible matches, 50% of the information. IBM (presumably!) loses 50% of information! And only because IBM (presumably!) trains the whole patterns! And doesn't see the synonyms/ thesaurus.

    Therefore, it is surprising, if we accept the assumption that IBM trains whole patterns, that IBM Watson finds something at all. But it becomes clear why IBM throws away a very profitable project...

    I wonder how much of my speculation is true?

  9. HmmmYes

    IBM fails to undersrand its new customers and what they want.

    I say new as the number of old IBM customers is shrinking due to blues hook and rip practises.

    Can you leverage AI for medicine. Unproven. IBM cannot though.

    1. I.Geller Bronze badge

      Amazon already uses AI technology, annotates and trains its data (Alexa) by dictionary.

      IBM is too big and bureaucratic to change its technology and start the annotation by dictionary. Therefore IBM will lose to younger competitors in medicine, Amazon is one of them.

      The right AI technology is the key! I discovered AI and patented it, I know.

  10. I.Geller Bronze badge

    Another hypothesis

    When you make your searching profile, you get sets of patterns into their contexts. This method was applied by IBM in Jeopardy!: IBM initiated a context for each question, using dictionaries and encyclopedias as its subtext. Then IBM looked for pieces of the most relevant texts, which have the most of search patterns.

    IBM called this technology Annotation Query Language.

    A hypothesis: in pharmacology, this did not work: almost all pharmacological information is HISTORICALLY contained in manually structured data, i.e. has no explicit contexts/ very few patterns but only implicit subtexts. IBM, therefore, cannot search for answers, having a lot patterns and almost none in data! IBM knows not how to restore subtexts? But IBM knew at Jeopardy!...

    AI can restore because it annotates words by their definitions-subtexts.

    This is another idea why IBM dropped pharmacology.

    1. I.Geller Bronze badge

      Re: Another hypothesis

      WSJ today:

      "Poor quality, inaccurate and unreliable customer and business data is preventing companies from leveraging AI, according to recent studies. About 76% of firms are aiming to extract value from data they already have, but only 15% said they currently have the right kind of data needed to achieve that goal, according to PricewaterhouseCoopers. The firm in January polled about 300 executives at U.S. companies in a range of industries with revenue of $500 million or more."

      Looks that the quality of pharmacology data is the true reason!

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2020