Sunday, 4 June 2017

The death of the “Data Scientist” or the return of the programmer




The “Data Scientist” is dead; long live the programmer! The world will soon not need “data scientists”, but instead programmers will rule the world.

This brutal assessment comes from the World Economic Forum (WEF), and I must say that their arguments make a lot of sense.

In this age of fake news, credentials are important; so what is the WEF? Founded in 1971, it is committed to improving the state of the world, is the International Organization for Public-Private Cooperation (1). Its board of trustees includes world leaders such as Jim Yong Kim – president of the world bank, Christine Lagarde – Managing Director of the IMF, but also Jack Ma – founder of Alibaba Group, Mukesh Ambani – Chairman of Reliance Group among others. (2)

Furthermore, the “centre for the fourth industrial revolution” is one of their major undertakings (3), and one of the focus areas is artificial intelligence and machine learning (4).

So this is not fake news, neither is it flimsy covfefe gossamer; sad!

In an article entitled “You’ve heard about it, but do you understand?  Everything you need to know about Machine Learning”(5), the WEF gives a very easy to understand convincing description of what machine learning is, how it can be used, and what is the way forward.

The article covers the history of Machine Learning, broad classes of machine learning (supervised – the machine is told what the different groups are, unsupervised – the machine is left alone to find out the different groups, and reinforcement machine learning – after finding out the groups it is told whether it is right or wrong this feeds back into the learning cycle.

In illustrating the concept of reinforcement ML, the WEF shows a clip where a machine learns to play Atari Breakout – where a horizontally mobile paddle at the bottom of the screen is controlled and the aim is to reflect a ball back to destroy an obstacle made of bricks at the top of the screen. The results are amazing (6).

But even more astounding is the commentary:
“The most important thing to know is that all the agent is given is sensory input (what you see on the screen) and it was ordered to maximise the score on the screen”
“No domain knowledge is involved! This means that the algorithm doesn’t know the concept of a ball or what the controls exactly do”

After 10 minutes the machine is clumsy. After 120 it is playing like and expert. After 240, it has developed its own strategy!


Isn’t this amazing?

So you will ask, what does that have to do with the death of the “data scientist”?

Well, if you, like me, think that Drew Conway was on to something when he came up with the basic “Data Scientist” skill diagram (8):

 

In this diagram, data science requires substantive expertise, or “domain knowledge”.

What the WEF agreeing with is that Machine Learning does not require “domain knowledge”. Furthermore, there is no mention of the “data scientist” in the WEF view point, but rather the article highlights “the programmer”:  ML “needs a programmer to tell it what to do when it is fed with data.” Not a “data scientist”, a “programmer”.

But does that mean the death of the “data scientist”?

Yes it does.

Think about it, if an problem can be solved without “domain knowledge”, then ML is all you need. 

And what does ML need? Not pompous “data scientist”, but the humble “programmer’ who had been cast aside and finally makes his/her triumphant return.

There is a wave washing away the veneer off the “data scientist” and restoring that of the programmer. An example argues that coding is not fun, but it is technically and ethically complex (9). 
What I find most interesting is the acknowledgement of the complexity of coding (I am sure most people would agree that it takes serious skill to develop industrial grade code), but also the introduction of ethics into the mix.

To quote from the article “As well as being highly analytical and creative, software developers need almost superhuman focus to manage the complexity of their tasks. Manic attention to detail is a must; slovenliness is verboten. Attaining this level of concentration requires a state of mind called being ‘in the flow,’ a quasi-symbiotic relationship between human and machine that improves performance and motivation.”

To push the idea further, the latest triumph of machine over man was Alpha-go. Alpha-go was built by the same organisation that developed the machine playing Atari Breakout above. And the same principle was used: “no domain knowledge”, but lots of Monte Carlo Simulation. (basically the machine learns which move is best by, from the current position, randomly trying different plays and evaluating the end result, then assigning a value to that play and picking the one that gives higher change of winning/lower chance of losing, and repeats the process at every turn)

No domain knowledge

No data scientist

One of the greatest triumphs of machine over man was enabled by programmers, not data scientists.(9)


The “data scientist” is dead, long live the programmer!
  

 Further readings/references:

  1. https://www.weforum.org/about/world-economic-forum
  2. https://www.weforum.org/about/leadership-and-governance
  3. https://www.weforum.org/center-for-the-fourth-industrial-revolution/
  4. https://www.weforum.org/center-for-the-fourth-industrial-revolution/areas-of-focus
  5. https://www.weforum.org/agenda/2017/05/what-is-machine-learning
  6. https://www.youtube.com/watch?v=V1eYniJ0Rnk
  7. http://drewconway.com/zia/2013/3/26/the-data-science-venn-diagram
  8. https://qz.com/987170/coding-is-not-fun-its-technically-and-ethically-complex/
  9. https://www.tastehit.com/blog/google-deepmind-alphago-how-it-works/
Posted by at No comments:

Sunday, 21 May 2017

Collecting DNA, Insurtech, and the death of the insurance industry



I was dozing off in front of the TV recently when I thought I must have been dreaming about 1984(1). But then I realised it was real.

Most of us have watched CSI(2) or other police dramas, and we ‘know’ at least a couple of things about DNA:

1 DNA is quite unique

2 Identical twins have identical DNA (basically they come from the same zygote that was split)

I was thus very intrigued by this advert that showed a pair of twins, who received different ‘health’ advice based on their DNA. Of course they could be fraternal, but then why use twins? In most people’s mind, when they think twins, they think identical.

So how would it be possible to have different advice to 2 people based on their DNA when they share the same DNA?

It was a ridiculous ad, from my point of view.

I was fully awake by the time I realised that the organisation getting people to gleefully share their DNA was an insurance company. This wasn’t dystopian 1984(3), this was a brave new world(4); we aren’t forced to do things that are probably bad for us, we do them more than willingly.

I am very concerned about data privacy, and allowing someone to get access to your DNA is a huge breach of your privacy. Someone was arrested recently(5) for refusing to divulge passwords to his mobile devices argued that “With a digital strip-search, it's not just your possessions that are searched, it's your personal and professional life”. I would say that “a strip-search of your DNA it’s not your present that is searched, but your past, that of your blood relatives,  and your potential future, one you may not even be aware of(6)”

Even today, in order to get the coverage on many insurance policies, you are required to fill up a form truthfully, and are reminded that your coverage would be invalidated if you were found to have been economical with the truth. Also you are required to report any changes to your answers. For example, for health coverage, the medical history of your family is usually part of the questionnaire. It makes sense; some illnesses are hereditary, and in order to properly evaluate your risks, such knowledge is invaluable. It also usually results in an increase in premiums paid.

Now what happens if an insurance company takes a peek at your DNA, and increases your premium based on some markers that you weren’t even aware of?

Or worse...

What happens if the insurance company denies your claim because while you thought you have no relatives who had cancer, your DNA clearly shows indicators of hereditary cancer (and you should have known of relatives who had it)?

That’s something I would try to resist as much as I can. Out of curiosity, I decided to take a deeper look at the product at their website (7). The insurance company has outsourced all the non-insurance bits to a third party, Prenetics(8); basically while the data is being collected by the “reputable” insurance company, the DNA data is not held by them, nor is the advice provided by them. The third party guarantees(9) that the DNA data will not be made available even to the organisation that provided the test to the individual (here the insurance company). Phew...

However they do reserve the right to send you marketing information, campaigns, third party ads based on your cookies, perform R&D...

I personally think that this would be something that Pharmaceutical companies would find very useful. If I were them, I’d get Prenetics to target people with specific DNA profiles, for different drugs to test their efficacy; or better still make the ads appear while you are surfing. Then based on the purchase where I collect personally identifying data, I could easily track the effectiveness of my drugs, at very low cost and you would know nothing about it; a true blind test.

Merger and Acquisitions are also a thorny subject. I am in no way suggesting that Prenetics is planning to break its own rules as set out in its privacy statement. However these rules do not necessarily apply to anyone who would choose to acquire Prenetics.

Prenetics is a Hong Kong based company. In a recent case when Facebook acquired whatsapp, the FTC wrote to both companies a strongly worded note(10) that stated: “We want to make clear that, regardless of the acquisition, WhatsApp must continue to honor these promises to consumers. Further, if the acquisition is completed and WhatsApp fails to honor these promises, both companies could be in violation of Section 5 of the Federal Trade Commission (FTC) Act and, potentially, the FTC’s order against Facebook”. There is no guarantee that the Hong Kong Authorities would take a similar stand; after all Hong Kong is known to be one of the freest markets in the world.

Imagine, being able to understand what strengths someone’s DNA shows, this pool of data would be manna for insurance companies, being able to select lower risk individuals. Most insurance up to now has been based on the idea of pooling risks. Being able to segment the pool into smaller, finer pools of similar risk would allow insurance companies to, at the same time, reduce premiums, reduce risk, and increase profits. Who wouldn’t want that?

However, think of the people who, through no fault of their own, are in the pools that carry high risks; either their premiums would shoot up, or they would become uninsurable. And out goes the ‘solidarity principle’.

And to me that’s one of the fundamental questions that insurtech asks: are we ready to have less solidarity, be more selfish? Big Data certainly allows more precise evaluation of risks, hence it is relatively easy to create segments of lower or higher risk, and price them accordingly. Is that something we, as consumers, want? May be for behaviour driven risk such as driving, but may be not for the lottery of genetics such as health. Or “do you feel lucky?”(11)

What is even more pertinent, how much say will we have in this?
Incentive wise, people who think they are lower risk will be attracted to lower premium pools, forcing the rest to be charged higher premiums. And organisations will have the incentive to devise ways of identifying the lower risk individuals since they can offer them lower premiums and make more profits at the same time.
The beauty of things is that, pushed to its extreme, as more detailed accurate data is collected, enabling the size of the segments to drastically fall, and offers become more personalised, at the limit, your hospitalisation insurance premium should be slightly higher than the actual cost of your hospitalisation over time, after all “data science”/”Big Data”... are not precogs(12).
If the costs of the premiums you would pay for insurance would be slightly higher than hospitalisation over time, why would you need insurance in the first place?
The main attraction of insurance if the costs of the premium over time are about the same as the cost of what the insurance is covering is that there would be no need to worry about finances when the risk hits (no double whammy). However, financial institutions might think of offering loans that would deal with cash flow issues if/when hospitalisation happens... Then why would you waste money on say hospitalisation insurance? You might be better off finding ways of making your savings grow in case of a rainy day.
Hence, to me, as “Big Data” and insurtech makes the cost of insurance closer to the actual cost of the risk (say hospitalisation), then there is even more reason to focus on growing savings/investments for a rainy day: investments as a form of insurance.
But until then, it may be a good idea to think twice before surrendering your DNA to a third party. That's a piece of data you need to be really careful about...

References:

Posted by at No comments:

Tuesday, 28 February 2017

PWC and the Oscars: what are you paying for and why?



The incident at the Oscars, where the wrong winner was announced in the category of “Best Picture” is very interesting. The results were vetted and made watertight by PWC, but I guess moonlight consists of smaller molecules than H2O. So the question is, if you hired PWC, what did you pay for?
First of all, the 2 PWC employees who oversee the whole process are stars in their own right; and show their process with personal attention to detail[1]. They even have their glamourous pictures on the red carpet[2]
The job of PWC is simple, ensure the votes are correctly counted, and the results delivered to the correct people at the right time to make the right announcement.
Sounds like a simple big data problem actually, you know: right offer, right person, right time...
However, PWC chose a very analog method, you can actually follow the Oscar votes briefcase across America! Perhaps in line with the rumours of hacked vote counting machines this is not a bad idea, but the execution must be there.
PWC is not the only one to have had major failures, a quick search on google with the name of a big 4 and the word “failures” shows: PWC[3], EY [4], Deloitte [5], KPMG [6].
And for people in Singapore, who can forget the Barings Bank collapse? Barings had not one but 2 Big 4 Auditors[7].
I am not against PWC, it's just that the incident triggered this blog.
While I looked at auditors, I could have as easily looked at the large IT consultancies... IBM[8] or Microsoft[9], Google [10]... In sum, even highly reputable companies do make 'big mistakes'.
So what are organisations paying for? Is it to ensure things are right or is it someone to blame?
If it is to ensure things are right, are you getting value for your money? This brings me to my 2 current bug bears:
1 When organisations are employing external experts, are they getting value for their money in terms of quality? Or are you paying for layers of management? This article from 2011 questions why large organisations were finding it so hard to change[11].
As a result, interesting business models that offer independent expert services have sprung up such as expert360[12] and alphazetta[13], this phenomenon has been explained in a McKinsey article on the gig economy [14]. This also ties in to my earlier argument that most organisations could use resources much better than hiring full time data scientists [15] (experts in the field of Analytics/'Data Science')
2 Why aren’t organisations such as auditing firms taking advantage of “Big Data”? I am not privy to the detailed workings of auditing firms, but I doubt they actually make full use of “Big Data”. Most auditing is done purely looking into the client organisations focused on looking at internal discrepancies. Very often, discrepancies are easier to detect if you have context, especially if the context is outside the control of the client organisation.
Let me give a very simple real life example in the domain of risk. Most banks have a risk rating to each of their clients, and again, these risk ratings tend to be based almost exclusively on the accounts of the client, with some variation for health of the sector in general.
However, it is easy for a bank to understand the relationships between various organisations.
http://assets.teradata.com/resources/aofa/dollar-diadem2.html
http://assets.teradata.com/resources/aofa/dollar-diadem2.html

This art piece from Teradata[16] illustrates what can quickly be done to understand the relationships between organisations. Now imagine what it means for an organisation whose main customer happens to be highly risky. Even if the organisation has healthy balance sheets, if the major customer collapses, it poses high, at least short term, risk (and opportunity depending on your risk appetite). But if you look at organisations independently, however deeply, you will miss this, and systematically underestimate the risk.
Similarly, looking at the Barings collapse, having a context against which to compare the activities of Mr Leeson would have been critical in understanding the risks. This is more into the area of employee activity, dealing with internal threats. Similarly in the recent case of 1MDB where a lot hinged on an unauthorised letter vouching for the worthiness of the accounts by an ex-Goldman Sachs dealmaker [17] this could easily have been spotted when looked at in context.
The real question is, did the players really want to look, or were they content to close one eye and hope for the best as their tills kept ringing?
I believe that the advent of “Big Data” and the gig economy will change or even is already fundamentally changing the provision of expert services. Organisations are likely, going forward, to use independent experts that know how to make use of “Big Data” on an as-needed basis, paying directly for the expertise, rather than engage large companies that are not evolving to adapt to the new realities and find themselves laden with too many overheads.


Posted by at No comments:
Subscribe to: Posts (Atom)