PGM-class and MRF parameter learning

I’m taking Stanford CS 228 (a.k.a. pgm-class) on Coursera. The class is great, I guess it provides close to the maximum one can do under the constraints of remoteness and bulkness. The thing I miss is theoretical problems, which were taken aside from the on-line version because they could not be graded automatically.

There is an important thing about graphical models I fully realized only recently (partly due to the class). This thing should be articulated clearly in every introductory course, but is often just mentioned, probably because lecturers consider it obvious. The thing is there is no probabilistic meaning of MRF potentials whatsoever. The partition function is there not only for amenity: in contrast to Bayesian networks, there is no general way to assign potentials of an undirected graphical model to avoid normalization. The loops make it impossible. The implication is one should not assign potentials by estimating frequencies of assignments to factors (possibly conditioned on features) like I did earlier. This is quite a bad heuristic because it is susceptible to overcounting. Let me give an example.

For the third week programming assignment we needed to implement a Markov network for handwriting OCR. The unary and pairwise potentials are somewhat obvious, but there was also a task to add ternary factors. The accuracy of the pairwise model is 26%. Mihaly Barasz tried to add ternary factors with values proportional to trigram frequencies in English, which decreased performance to 21% (link for those who have access). After removing pairwise factors, the performance rose to 38%. Why has the joint model failed? The reason is overcounting evidence: different factor types enforce the same co-occurrences, thus creating bias towards more frequent assignments, and this shows it can be significant. Therefore, we should train models with cycles discriminatively. 

One more thought I’d like to share: graphical model design is similar to software engineering in the way that the crucial thing for the both is eliminating insignificant dependencies on the architecture design stage. 

Next Primal Chef Event Sunday 5/20

Gil Butler has been working on a television show called Primal Chef, where he invites local chefs to make creative dishes from a list of Paleo ingredients, in a designated amount of time.  The format is reminiscent of Iron Chef.  The food is judged afterward by figures in the Paleo community.  Robb Wolf was a judge on the first episode.

Gil has invited me to be a judge on the next show, along with Sara Fragoso and Dr. Tim Gerstmar.  The next day, Sunday April 20th, Gil is organizing a catered Primal Chef event in Seattle, with Paleo dinner, speakers, entertainment, prizes, and a screening of part of Paleo Chef episode 1.  You can read the details and sign up here.  I won't be speaking because I don't have time to put together another talk right now, but I will be attending the event. 

What if Kimmer was a real 100 lb weight loss success story?

direct image link

For those not familiar with "Kimmer", allow me to fill in with a little background.  Kimmer was the screen name used by one Heidi Diaz who started a subscription weight loss site KimKins.com.  The website -- version 3.0 -- lives on.  KimKins is essentially several low carb diets that are also low fat and low calorie.  I think the most severe of them is around 300 cal/day.  Kimmer used pictures of others for fake "after" pictures while remaining obese herself and promoting KK as a successful weight loss plan.  You see, here is the before and apparently a real intermediate pic of Kimmer, and her "after" shot.  Below that we have a montage of before/afters used by Heidi.  It always amazed me that folks were apparently unquestioning of these shots, but maybe it's just me that would look at any of them and at least wonder whether this was the same person or not.  Below right is a split screen of Heidi's before picture and the shot taken by a private detective.

Read more »

Drs. Westman & Vernon on Gastric Bypass and Diabetes

So I was getting my morning news fix the other day and of course the new study mentioning gastric bypass as a cure for diabetes was discussed.  One of the many mainstream press articles about two recently published articles in NEJM can be found here.  The full texts of both studies are available online free:  Bariatric Surgery versus Conventional Medical Therapy for Type 2 Diabetes - Mingrone et.al. and Bariatric Surgery versus Intensive Medical Therapy in Obese Patients with Diabetes - Schauer et.al.  I've discussed some other studies previously here.  Lastly on the links for now, I have not followed all contained in this post, but from a surgery center blog we have:  Gastric bypass to ‘cure’ diabetes goes mainstream (you are forewarned that the last one is a "medical tourism" clinic in Colombia).

Read more »

Once More with Feeling - Johnson and Johnson Found Guilty, Fined $1.1 Billion

The latest legal black eye for giant pharmaceutical and device company Johnson and Johnson appeared in an Arkansas court room.  As documented by Bloomberg, via NJ.com,

Johnson & Johnson was ordered to pay $1.1 billion by a state judge after an Arkansas jury found the company’s officials misled doctors and patients about the risks of the antipsychotic drug Risperdal.

Jurors in Little Rock yesterday said the company’s marketing campaign also violated consumer-protection laws. The panel deliberated about three hours before finding J&J and its Janssen unit engaged in 'false or deceptive acts' by sending a 2003 letter touting Risperdal as safer than competing drugs to more than 6,000 doctors across the state.

The prosecutors had alleged a variety of kinds of deceptions about Risperdal,

Along with contending that J&J and Janssen defrauded the Medicaid program by failing to properly outline the antipsychotic medicine’s risks, Arkansas officials alleged J&J officials deceptively marketed the drug as safer and better than competing medicines.

The state also argued the companies marketed the drug for 'unapproved uses, including various symptoms in children and the elderly' after being warned by federal authorities to halt such sales.

In summary,

Arkansas Attorney General Dustin McDaniel said in an e- mailed statement that he sued because residents in the state deserved to be protected from 'fraud and deceptive practices.'

He said that jurors found 'Johnson & Johnson and Janssen Pharmaceuticals lied to patients and doctors because they cared more about profits than people.'

Bloomberg noted that

It’s the third jury verdict against J&J, the second-biggest maker of health products, in cases where states alleged it hid Risperdal’s risks and tricked Medicaid regulators into paying more than they should have for the medicine. Louisiana and South Carolina juries also found the company’s Risperdal marketing violated consumer-protection laws.

In fact, this is just the latest in a remarkable string of legal cases suggesting an ongoing pattern of unethical and illegal behavior by this very large health care corporation. As we wrote recently, this included
- Convictions in two different states in 2010 for misleading marketing of Risperdal, as noted above
- A guilty plea for misbranding Topamax in 2010
- Guilty pleas to bribery in Europe in 2011 by J+J's DePuy subsidiary
- A guilty plea for marketing Risperdal for unapproved uses in 2011 (see this link for all of the above)
- Accusations that the company, which makes smoking cessation products, participated along with tobacco companies in efforts to lobby state legislators (see post here)
- A guilty plea to misbranding Natrecor by J+J subsidiary Scios (see post here)
- More recently, in 2012, testimony in a trial of allegations of unethical marketing of the drug Risperdal (risperidone) by the Janssen subsidiary revealed a systemic, deceptive stealth marketing campaign that fostered suppression of research whose results were unfavorable to the company, ghostwriting, the use of key opinion leaders as marketers in the guise of academics and professionals, and intimidation of whistleblowers. After these revelations, the company abruptly settled the case (see post here).
- Most recently, there are reports that the company is in negotiation with the US Department of Justice to settle other lawsuits about the marketing of Risperdal, perhaps for as much as $1.8 billion (see this BusinessWeek story.)

Meanwhile, as we also discussed recently, Johnson and Johnson seems to have lost the ability to manufacture high quality products. It has had to make 30 separate product recalls since 2009. The latest was Liquid Infant Tylenol. (The current WSJ Health Blog list of recalls can be found here.)

So the real question here is how many instances of manufacturing of defective, mislabeled, contaminated, or harmful products, and how many court cases showing unethical or illegal behavior will it take until the leadership and governance of this company improves (and by extension, the leadership and governance of other large health care corporations with similar bad records improves).

As we posted not long ago, the CEO who presided over most of these messes will be allowed to retire with a huge golden parachute.  His devotion to "making the numbers," that is, hitting short-term revenue targets ahead of all other goals, probably had something to do with a company once thought of as a paragon of corporate behavior turning so bad.  Yet he became extremely rich doing this, and neither his riches nor his legal status have ever been challenged.  His successor apparently will be another former sales representative who may be just as devoted to "making the numbers,."  Thus has the business school dogma that short-term financial results, prettied up as "shareholder value," is the only thing that should matter to corporate leadership (look here) poisoned health care.

How many more doctors and patients have to be deceived, how many products have to be contaminated or defectively made before we demand better leadership of health care organizations?