What a terrible, terrible article. Why is this blog-able information, there is nothing substantial here. Everything is future possibilities and no concrete details.
There have been many people working on similar projects to this in academia and otherwise. "Identifying the proteins that modify p53" is incredibly easy with the database information in Medline and doesn't require a "Watson" to do so. That being said, finding connections within medical literature could be interesting, but only if it is combined with other public databases that describe protein interactions and gene information. This blog post could be much more detailed in this sense.
I'm reaaaally getting tired of IBM's PR department plastering Watson press releases all over the place in the guise of "news". Especially when "Watson" is more of a brand name, and there's next to no evidence that whatever "it" is is actually better or more cost-effective than off-the-shelf techniques (let alone the frontier of research) at any particular machine learning application.
"Ford sees broader role for Taurus in aiding transportation networks of the future!" Of course they do.
Watson may be overhyped, but it isn't a replacement for off the shelf machine learning algorithms. Watson is more like a search engine that takes advantage of a lot of natural language processing. It is for answering questions in natural language and searching through huge amounts of literature automatically.
I don't mean to be glib; and I do understand that "answering natural language questions" is distinct from "search engine". But I have the feeling that it's a capability distribution that is much more impressive in the context of a full-court IBM sales pitch to an executive who saw it on Jeopardy once and wants to get in on this machine learning thing that everyone's talking about, than as a tool used by experts to do a task or set of tasks better. Their shameless press release marketing isn't exactly dispelling that notion.
As someone who works in the biological sciences and who has played around with various machine learning algorithms (ann, k-nn) this sounds like an advertising gimmick. I am all for using computers to process and analyze data, but frankly machine learning has not reached a point yet where algorithms can "identify" new hypotheses by sifting through the scientific literature.
The article is scarce in details but this is a real product with real applications. The software uses NLP and ML techniques to sift through large amount of scientific documents, i.e., patents or publications, and extracts relevant entities, e.g., proteins, drugs, genes and their relationships, e.g., how a protein acts on a gene, and lets the user visualize all this and query it using natural language.
I can give more details if anybody is curious. (Disclaimer: I work for Watson)
1. It feels a lot like literature-based-discovery. Is it ? if so , how is watson better than current literature based discovery methods that did create some scientific hypothesis AFAIK?
If not please share more details(if you can of course) about how it reaches hypothesis).
2. Does it work well in non biological literature ? Are there any examples ?
1. I am not super familiar with the space (I am not responsible for this specific product) so I may not answer your question exactly. The discovery product is an interactive tool more than a static tool generating hypothesis. It lets you search, navigate, visualize, query in natural language the literature and the extracted entities/relationships.
It uses a lot of pretty advanced tech to do that:
* State-of-the-art statistical annotators
* Some deep domain knowledge understanding of the entities themselves (e.g., how chemical entities are composed)
* Some nifty visualizations
* Many component from the Jeopardy stack (you can read the Jeopardy papers for more insights) - deep question parsing, candidate generation techniques, scorers, etc.
* Down the line it'll integrate more reasoning techniques along the lines of http://www.research.ibm.com/cognitive-computing/watson/watso...
2. Our first version was focused on life sciences and intelligence, we are expanding it to finance, law, and education.
Is there an API for Watson? Is it possible as a developer to obtain access?
More abstract: Watson, based on this article (http://www.wired.co.uk/news/archive/2013-02/11/ibm-watson-me...), can already diagnose cancer better than a human doctor. Will we see more of this functionality rolled out publicly? Or will it be reserved for use by healthcare organizations?
Platform 2.0 is coming in October. It will have many new APIs, the docs and code examples will be public, and we will start opening out for beta testing to a much larger set of developers. The goal ultimately is to offer all these capabilities through APIs to anybody.
I went through the paperwork pile to apply to round one access of the Watson APIs and discovered the shocking restrictions/requirement that the system apparently "needs" unstructured text, and if your dataset is more structured than unstructured they won't accept you into to the program.
Needless to say I was a bit annoyed, because I was already using fact extraction in the system I wanted to test Watson's query 'skill' on. I'm in no position to store the raw text. That would require over a hundred times more storage, probably closer to a thousand times the storage costs, making it fiscally untenable for me to even build the database let alone a product with it. Any idea if release 2 in October will change this restriction and give people who aren't sitting on GB of unstructured text, like myself, a chance to apply.
The techniques of today are actually much more elaborate than decision trees. The key is gathering evidence automatically (i.e., learning) from medical literature - that requires parsing the text through ML techniques that have been invented just recently.
I agree that adoption is a challenge but the tech is also really challenging. Getting a system to a level of accuracy that gives doctors enough confidence to use it is no small task.
A "second opinion" service would have a lot of potential legal pitfalls.
Not necessarily directed at you, but one big frustration of mine while leading a project team was when someone would list off 'potential issues'. If the concept is compelling and the quick estimates look like positive ROI, go get the lawyers/marketers/buyers/executive team to do their job and push forward or come back with logical and specific reasons!
What happens in the situation when something Watson has "studied" gets proved wrong or inconclusive say 1 year later. Does Watson have a method to discount that and all subsequent chains based on that research?
My understanding is that if you feed the results of the research into the system then it would be there when the algorithms are run in future and from that point on, the results would be taking the research into account and it's answer would change.
Yes, the system works in a way similar to the Jeopardy Watson. It collects evidences, and weighs them based on their date and possibly source quality. So new results will displace older results at some point.
As awesome as it was watching it play Jeopardy, it feel like that is all it has ever done. So many press releases boast about the areas it could improve, however the real-life tangibles seem disparate at best. It would be great if the PR department touted its actual, real-life, in-the-field results rather than hypotheticals. Next up, Watson considered for role in improving food production!
What a terrible, terrible article. Why is this blog-able information, there is nothing substantial here. Everything is future possibilities and no concrete details.
There have been many people working on similar projects to this in academia and otherwise. "Identifying the proteins that modify p53" is incredibly easy with the database information in Medline and doesn't require a "Watson" to do so. That being said, finding connections within medical literature could be interesting, but only if it is combined with other public databases that describe protein interactions and gene information. This blog post could be much more detailed in this sense.
I'm reaaaally getting tired of IBM's PR department plastering Watson press releases all over the place in the guise of "news". Especially when "Watson" is more of a brand name, and there's next to no evidence that whatever "it" is is actually better or more cost-effective than off-the-shelf techniques (let alone the frontier of research) at any particular machine learning application.
"Ford sees broader role for Taurus in aiding transportation networks of the future!" Of course they do.
+1. IBM must regain control over their PR people, speak less and do more, or they will lose any credibility with developers.
Maybe Watson's inference engine could be reworked into an automated PR generator and put them out of a job.
Maybe it already has...
Watson may be overhyped, but it isn't a replacement for off the shelf machine learning algorithms. Watson is more like a search engine that takes advantage of a lot of natural language processing. It is for answering questions in natural language and searching through huge amounts of literature automatically.
So, like Solr?
I don't mean to be glib; and I do understand that "answering natural language questions" is distinct from "search engine". But I have the feeling that it's a capability distribution that is much more impressive in the context of a full-court IBM sales pitch to an executive who saw it on Jeopardy once and wants to get in on this machine learning thing that everyone's talking about, than as a tool used by experts to do a task or set of tasks better. Their shameless press release marketing isn't exactly dispelling that notion.
Nice Try....
As someone who works in the biological sciences and who has played around with various machine learning algorithms (ann, k-nn) this sounds like an advertising gimmick. I am all for using computers to process and analyze data, but frankly machine learning has not reached a point yet where algorithms can "identify" new hypotheses by sifting through the scientific literature.
sorry, but, ... playing with KNN/ANN and claiming insight into ML is like playing with a nerf gun and claiming knowledge of advanced tactics.
Other companies focused on genes related to extending human life span are Google's Calico http://en.wikipedia.org/wiki/Calico_(company) and http://genopharmix.com and http://www.humanlongevity.com/
I am very interested in the IBM Watson platform but so far I am still on the waiting list for being considered for early access.
Hint to IBM: at least make all documentation, example code, etc. available for perusal.
It will happen on October 7th.
Thanks for the information. I put October 7 on my calendar as a reminder.
The article is scarce in details but this is a real product with real applications. The software uses NLP and ML techniques to sift through large amount of scientific documents, i.e., patents or publications, and extracts relevant entities, e.g., proteins, drugs, genes and their relationships, e.g., how a protein acts on a gene, and lets the user visualize all this and query it using natural language.
I can give more details if anybody is curious. (Disclaimer: I work for Watson)
It sound very impressive and i'm very curious.
A few questions:
1. It feels a lot like literature-based-discovery. Is it ? if so , how is watson better than current literature based discovery methods that did create some scientific hypothesis AFAIK?
If not please share more details(if you can of course) about how it reaches hypothesis).
2. Does it work well in non biological literature ? Are there any examples ?
1. I am not super familiar with the space (I am not responsible for this specific product) so I may not answer your question exactly. The discovery product is an interactive tool more than a static tool generating hypothesis. It lets you search, navigate, visualize, query in natural language the literature and the extracted entities/relationships.
It uses a lot of pretty advanced tech to do that: * State-of-the-art statistical annotators * Some deep domain knowledge understanding of the entities themselves (e.g., how chemical entities are composed) * Some nifty visualizations * Many component from the Jeopardy stack (you can read the Jeopardy papers for more insights) - deep question parsing, candidate generation techniques, scorers, etc. * Down the line it'll integrate more reasoning techniques along the lines of http://www.research.ibm.com/cognitive-computing/watson/watso...
2. Our first version was focused on life sciences and intelligence, we are expanding it to finance, law, and education.
Thanks you!
Is there an API for Watson? Is it possible as a developer to obtain access?
More abstract: Watson, based on this article (http://www.wired.co.uk/news/archive/2013-02/11/ibm-watson-me...), can already diagnose cancer better than a human doctor. Will we see more of this functionality rolled out publicly? Or will it be reserved for use by healthcare organizations?
Platform 2.0 is coming in October. It will have many new APIs, the docs and code examples will be public, and we will start opening out for beta testing to a much larger set of developers. The goal ultimately is to offer all these capabilities through APIs to anybody.
Thank you!
I went through the paperwork pile to apply to round one access of the Watson APIs and discovered the shocking restrictions/requirement that the system apparently "needs" unstructured text, and if your dataset is more structured than unstructured they won't accept you into to the program.
Needless to say I was a bit annoyed, because I was already using fact extraction in the system I wanted to test Watson's query 'skill' on. I'm in no position to store the raw text. That would require over a hundred times more storage, probably closer to a thousand times the storage costs, making it fiscally untenable for me to even build the database let alone a product with it. Any idea if release 2 in October will change this restriction and give people who aren't sitting on GB of unstructured text, like myself, a chance to apply.
We could already diagnose cancer better than a human using decision trees in the 1960's.
The breakthrough with computational medicine is actually using it.
The techniques of today are actually much more elaborate than decision trees. The key is gathering evidence automatically (i.e., learning) from medical literature - that requires parsing the text through ML techniques that have been invented just recently.
See https://www.youtube.com/watch?v=8lGJ0h_jAp8 for what the Watson medical diagnosis product looks like - it's not just about diagnosis, it's about evidences.
That's definetly progress, but still , the point made by spitfire is true: The real challenge is adoption, not tech.
We even see it with the low acceptance of Watson in medicine and IBM's move into africa.
I wonder thought: why hasn't IBM offered a "second opinion" service directly to consumers ? It could might have sped adoption.
I agree that adoption is a challenge but the tech is also really challenging. Getting a system to a level of accuracy that gives doctors enough confidence to use it is no small task.
A "second opinion" service would have a lot of potential legal pitfalls.
Not necessarily directed at you, but one big frustration of mine while leading a project team was when someone would list off 'potential issues'. If the concept is compelling and the quick estimates look like positive ROI, go get the lawyers/marketers/buyers/executive team to do their job and push forward or come back with logical and specific reasons!
Sorry for the late response - i'm across the Atlantic.
I wonder, with the API , would it be possible for a startup to build such "second opinion service" , with reasonable costs ? will you block it?
What happens in the situation when something Watson has "studied" gets proved wrong or inconclusive say 1 year later. Does Watson have a method to discount that and all subsequent chains based on that research?
My understanding is that if you feed the results of the research into the system then it would be there when the algorithms are run in future and from that point on, the results would be taking the research into account and it's answer would change.
Yes, the system works in a way similar to the Jeopardy Watson. It collects evidences, and weighs them based on their date and possibly source quality. So new results will displace older results at some point.
As awesome as it was watching it play Jeopardy, it feel like that is all it has ever done. So many press releases boast about the areas it could improve, however the real-life tangibles seem disparate at best. It would be great if the PR department touted its actual, real-life, in-the-field results rather than hypotheticals. Next up, Watson considered for role in improving food production!