The Ornithology of Epistemology

Philosophy of science is as useful to scientists as ornithology is to birds. 

— Richard Feyman

There is an interesting flock of bird concepts in epistemology, philosophy of science and risk theory:

Russell’s Turkey

Black Swan

Pink Flamingo

Ostrich Effect

Raven Paradox

283px-kaninchen_und_enteDespite decades of intensive ornitho-epistemological research, the bird-ness of Rabbitduck [1] could not been firmly established yet. Every time we think we nailed it, the the rabbit-ness wins over and we are back to square one. More research is needed.

[1] We are sad to report, but there is still a small, but very vocal part of our research community, that insists on the clearly ignorant “Duckrabbit” nomenclature.

We also don’t subscribe to the wishy-washy, cheap consensus of the Copenhagen interpretation, that claims that the creature is both a rabbit and a duck.

Sararīman in the high agency economy and in the roboconomy

1. Rise of the freelancer

Taleb just reprinted his excellent How To Legally Own Another Person essay at Medium.com. It originally appeared at Evonomics, which is very much worth following.

Developed nations are increasingly relying on freelancer / contractor works (40% of America’s workforce will be freelancers by 2020 says Intuit’s report). This puts new demands on the workforce – high agency, independence, risk tolerance. Reputation is already now one of the main currencies (think Uber, Airbnb et al.).

2. Internal and external coordination costs

Is this the death of the poor Sararīman (salaried employee)? How far can this trend go?

Coase’s Theory of the Firm did tap into an important insight from complex systems: Firms grow to achieve a dynamical equilibrium between internal and external coordination costs, between economies and dis-economies of the scale.

The technology-enabled, highly networked world with its remote communication and work tools and reputation-rings shifts the set point of most companies towards a more distributed, freelancer-reliant structure.

Nonetheless there are limits to this shift as coordination cost will never quite reach zero. More importantly, there are limits on the size of the workforce pool, that does have the required levels of agency and risk tolerance.

3. Rise of the robo-worker

There is one more, overwhelming trend that is coming to play. The rise of the freelancer (and the death of sararīman) will co-evolve with automatization. The Uber freelancers can already start competing with their robotic colleagues.

The salaried man will not live long enough to die by glorious Karōshi (death by overworking), but rather die the double death of outsourcing to a freelancer or a robot.

4. The Tale of the Slave

While I don’t want to get into the neo-luddite debate now, I do want to link Taleb’s article with Nozik’s fantastically funny Tale of the Slave. It is a very short read, but very much worth it. The point being, that maybe the salaryman-hood is not such such a great thing to cling too anyway.

5. The Diogenesian or Epictetian future of the salaried worker

Ultimately, the boundary conditions for the “freed” (= kicked-out, redundant) Sararīman are:

  1. living in barrel like fellow slave Diogenes (the neo-luddite scenario)
  2. or flourish like ex-slave Epictetus, becoming a wealthy and self-actualized freeman

Here, I’m more on the cautious optimist’s side.

While not everybody will become an influential philosopher (artist, writer…), my hope is that we will be wise enough to use the surplus generated by automatized technology to make the barrel really comfty (good Wifi and VR googles are included of course).

Wait, did I say optimistic?

 

 

 

Niches between two absurd positions

1.

In many discussions we are drawn to extreme boundary values. Here is a possible dilemma for a potential parent:

“If all comes down to genetics, there is nothing I can do to really affect my child. It is a total gamble. I have no control.”

But imagine it was 100% nurture: every interaction with your child, every word you say shapes its personality – permanently and possibly irreversibly.

Who could bear this kind of responsibility?

2.

This can be extended to everything in your life. (Learned) helplessness is a serious condition, when one believes that everything that happens to him is completely out of his/her control.

The other extreme is Total Responsibility – you are responsible for everything that happens to you, to your current conditions and future prospects.

Again an immense burden.

3.

As it is often the case, the truth is in-between: Genetics does determine most of the variance in traits (even complex ones) and its effect depends on time and overall socio-economic status (hereditability increases with both). It seems that nurture doesn’t do much on long term – but at least something short term. Beyond partner selection, one is largely (but not completely) absolved of metaphysical responsibility (definitely not the case for physical responsibility – shelter, food, protection).

For life outcomes it is similar – any outcome is a mixture of elements you control (your skills, resources you invest in it etc.) and random noise. The mix depends on the task/decision on hand, going from pretty much all noise (playing roulette), to much more controlled environments (but never completely without noise). You have some responsibility for the outcome, not total.

We are caught somewhere between two absurd positions, reminding me of:

Doubt is not a pleasant condition, but certainty is absurd.

— Voltaire.

 

Possibiliy of an injection attack on self-driving deep neural networks

Image processing deep neural networks can be catastrophically confounded by imperceptibly small perturbations in the input image, as demonstrated by Szegedy et al. 2013.

Nguyen et al. 2014 used genetic algorithms to purposely evolve abstract images that well trained neural networks confounded with real objects.

guitar_classification

Nguyen et al. 2014: Image evolved so that a neural network miss-classifies it as a guitar. “Swerve left” command could in principle be evolved in a similar way.

 

Using these techniques it could in principle be possible to construct artificial images (or video sequences) which when injected into the visual field of a self driving car could cause unwanted, possibly dangerous behavior (such as sudden swerve into opposing traffic).

It is theoretically possible (but likely practically very hard), to create adversarial images that would have the same catastrophic effect, even if covering only part of the visual field of the car (e.g. by holding up a printout of such an image at roadside).

Speaking of injections – an older, “fun” idea are SQL injections on licence plates as a way to mess with automated traffic surveillance systems (plate gets OCR-ed, and written into the database – which possibly triggers a drop table if unguarded). This is a special case of injection attacks – the adversarial data payload is a code snippet (a so called “code injection”).

licenceplatecamerasqlinjection

A “Licence plate” with an SQL injection attack as a way to fight back traffic cameras.

(I discuss “psychology” of deep learning networks also here).

David MacKay passed away

Prof. David MacKay passed away yesterday. He was an excellent teacher, author and scientist. I’ve learned about his terminal cancer earlier this week. He chronicled his treatment on his blog.

mackay-xlarge_translupkpzg5rbvobgsrh7n6pegtdwnmx8g9nfaaaf9gsxk

I’ve learned about Prof. MacKay’s work through his fantastic, free, book Information Theory, Inference, and Learning Algorithms – a book that is scientifically proven to be better than Harry Potter. The gist of the book is also captured in the series of lectures here.

But it was his second fantastic, free, book Sustainable Energy – without the hot air that made a personal impact on my life – I decided to work in the energy sector, because it seemed as a very important topic plagued by misguided incentives, wishful thinking, but also full of exciting opportunities for a better future. Prof. MacKay summarized my motivations wonderfully:

I love renewables, but I’m also pro-arithmetic.

I even mentioned the book during my interview (of course nobody heard about it, which puzzled me at that time…).

Two years later I was working on a project of optimizing home heating and I needed the some additional data from the book. I wrote to Prof. MacKay and he kindly sent me the data. Looking at the date of the email, it was less than 2 weeks after he learned his terminal diagnosis.

Prof. MacKay taught us a lot, not only about information processing, machine learning or energetics, but especially about scientific integrity, curiousness and independent thought. For all that I’m deeply thankful to him.

Interpreting 23andme data (mostly for free)

Introduction

23andme is a personal genome mapping service. You send them your saliva sample and get a (partial) map of your SNP genotype. We have tried the service and the whole experience was very smooth and satisfactory. The results took about 10 weeks (compared to standard 6-8 weeks), but we received a notification email apologizing for the delays.

Once your profile is analysed you get on-line access to your reports. Unfortunately, after a FDA ruling, 23andme can’t provide health related information to its customers. Therefore you only get basic information, your maternal haploid group (parental too if you are male) and a few other ancestry related test.

I would not recommend purchasing the test just for this information. Fortunately, you can download your raw data for further interpretation.

So thanks to FDA’s infinite wisdom, instead of having your genetic data in a single place you end up sending it nilly-willy to dozens of web services trusting their security and ToS agreements. Sounds like a good plan.

Most services can either directly connect to your 23andme profile or you can upload your data set (just read the instructions, some want the zip file and some the plain txt file).

Some of the services are paid, many are free. Here is a quick rundown/mini-reviews. The services are ordered in a rough, subjective utility.

Obviously – all disclaimers related to health data, security and privacy apply. You are on your own responsibility.

Codegen.eu: free

codegen is a free service and I recommend it as the first station. You get a quite exhausting report with human readable annotations thoroughly linked to SNPedia.

The focus is mostly health related – so this is your main station for all the “good” news about your cancer, diabetes, obesity, Alzheimer’s/Parkison etc. risk factors.

The site is in beta, so not everything is quite top-notch. But from the services we tried this was the best one.

Foundmyfitness: free

Rhonda Patrick is one of the most helpful people on the planet. Her nutritional / nutrigenomic reports are worth to read and the podcast is also highly recommended. She also launched a free 23andme interpretation service with a focus on nutrigenics (it seems micronutrient and fat metabolism are currently covered in most detail). The services is in early phase, but already now is very useful. You’ll also receive some nutrition advice along the interpretation.

Geneticgenie: free

Geneticgenie provides a free methylation and detox report. The methylation report is quite detailed, but the detox one is just a table with no interpretation. For now I only saved the result, because they will require a bit more work.

There is a significant overlap here with foundmyfitness and livewello (below).

Interpretome: free

Interpretome is a free educational javascript application (so – for a change – your data won’t leave your computer). The focus are educational applications, but there are interesting diagnostics for type 2 diabetes and the “longevity gene”.

There are also several specialized pharmacology related screens – potentially interesting if you are taking the pharmaceuticals listed here. For example, they provide a Warfarin dosage calculator based on your health data (age/health/weight) plus your genetic profile. This gives a taste of the near-future, genetically personalized medicine.

livewello: 20$ one time payment

livewello was the only commercial service we tested. The fee is 20$ one time, for a battery of reports or a monthly subscription for personalized health report emails. The focus is again nutrition.

Compared with the other services the design of the pages is nicer, but I found the reports lacking – there was large overlap esp. with foundmyfitness and geneticgenie, yet a few important markers were missing at livewello.

It’s likely that the service will improve with time, but at the moment I’d say it is not really worth the investment.

Untested

I’ve researched a few other services, but didn’t test them yet. Here are my research notes.

Promethease: 9$ one time payment

For a one time payment of 9$ you seem to get a very detailed interactive report from Promethease. Based on the preview video, its academic pedigree is very obvious – the tool seems to be quite technical and very detailed.

I’m fairly certain that I’ll get this report in the future, but for now decided to wait a bit, since I have quite a lot to chew on already.
But definitely have a look.

Enlis: free

I also wanted a nice visualization tool for the data and preliminarily Enlis seems to fit the bill. This is mostly for educational purposes, not for direct interpretation per se.

I haven’t tried it yet, and I plan to explore a few more options. The downside here is again the need to upload the data. There has to a local vizualization tool, right?

Atheligen

Atheligen does “athletic profiling”. They have their own kit (really expensive), but can also take 23andme data. I haven’t tried the service yet, because the page didn’t work for me. Will try later.

Summary

I definitely recommend 23andme. I think we received some important information for the future. The situation with the interpretation is a bit unfortunate and takes a bit more work, but it is also educational and you can learn a lot.

As always – please remember the statistical nature of the data and the non-genetic factors!

Do you have any favorite services or tool?