Fabio Colasanti

Earlier this week, this tweet
[https://mobile.twitter.com/random_walker/status/1196870349574623232] appeared
on my feed.

> Much of what’s being sold as "AI" today is snake oil. It does not and cannot
work. In a talk at MIT yesterday, I described why this happening, how we can
recognize flawed AI claims, and push back. Here are my annotated slides: 
https://t.co/iCpyFw5urN pic.twitter.com/pmOTI3vq8p [https://t.co/pmOTI3vq8p]

— Arvind Narayanan (@random_walker) November 19, 2019
[https://twitter.com/random_walker/status/1196870349574623232?ref_src=twsrc%5Etfw]

Prof Narayanan's deck
[https://www.cs.princeton.edu/~arvindn/talks/MIT-STS-AI-snakeoil.pdf] has
reached a relatively sizeable audience and received widespread positive
commentary. After all, anyone who has had any involvement with AI in the last
few years can 100% relate with the catchy title of his deck.
Hey - we're just about starting to see a feeble light at the end of the
blockchain tunnel, right?

But, but...
Something doesn't quite sit well with me. So I'm going to stick my neck out and
respectfully reject, if not the conclusions, the way Prof Narayanan has chosen
to make his point. Or rather, the way he has chosen to make his real point.

I find this deck deceptively coy, as the actual message is not "How to recognize
AI snake oil" but rather "Why I don't want AI anywhere near socially sensitive
topics", which would have been a perfectly valid argument, and made for a
fantastic set of slides, but required a slightly different set of arguments and,
most importantly, to be explicit about it.

At the end of the day ML/AI is about identifying patterns in data and draw
inferences, so of course it can "predict the future". Google Maps does it all
the times when it gives you an ETA for your journeys. In Prof Narayanan's own
examples, bail decisions and border control are not performed by machines, but
based on the prediction of machines. It's not about predicting the future: it's
called decision support, performed at speed that is beyond human capabilities.
Is the way banks and insurance companies make decision on mortgages and premiums
snake oil too? 
Prof Narayanan seems exceedingly preoccupied about "massive transfer of power
from [...] workers to unaccountable tech companies" and points at "enormous
commercial interests" that "we must resist" and "push back". He is of course
completely entitled to give the whole argument a political spin, but I'm not
sure this helps his arguments.

The real issue here is that some problems, and especially socially impactful
ones, carry high personal risks if an inference is wrong (problem 1). Therefore,
in order to minimise that, large amounts of potentially sensitive data may be
required (problem 2) and in most, if not all, cases there is a strong argument
for transparency, accountability and human-in-the-loop that makes such
propositions much more expensive and challenging from both a technical (problem
3) and governance (problem 4) perspective. 
Bias poses no doubt massive challenges (problem 5), not just in its
identification and removal, but also in its very definition, as what turns
common sense into bias is a function of our set of social and cultural values,
which greatly varies in time and space. Should history be changed because it
doesn't fit our Western 21st century view of the world? I don't have an answer,
but at least I'm asking the question.
Also, there is the topic of regulation which can't be ignored in these cases:
who establishes the boundaries? Who is it meant to enforce them? Do we need a
universal legislation, or would locally applicable frameworks suffice? Who makes
a call on what is an acceptable risk vs potentially beneficial outcomes -
supranational bodies, governments, individuals, etc.. ? (problem 6)

So the better discussion to have is around the risks involved in deploying AI in
socially sensitive fields, what is the minimum level of benefits it must bring
about to make it worth us facing the challenges above, and how do we go about
having that conversation.

Just branding the whole thing as snake oil is sidestepping the issue and largely
unfair, because it's bypassing the actual conversations to be had and going
straight for the outcome that Prof Narayanan - in arguably a biased opinion -
has decided is what's best for us all.

AI is a core discipline at Data Language. We have an expert team skilled in machine learning and deep learning, and importantly we also have the engineering and deployment skills to deliver AI at scale for your production systems.

AI is a core discipline at Data Language: Our expert team is skilled in machine learning and the engineering skills needed to deliver AI at scale.

Artificial Intelligence delivered at scale - our core discipline.

Artificial Intelligence

We need to talk about AI

(Update @ 2020: "Tagmatic" has been rebranded as the Text Classification service
in our "Text AI Services
[https://datalanguage.com/products/datalanguageai/text-ai-services]" product
suite).

It may be argued that customer satisfaction is the only and ultimate KPI that
really matters when judging the success of a product. In which case Tagmatic
would pass with flying colors, given the fantastic feedback we are receiving
from clients. 

Remarkably enough, we found that sheer inferencing performances are often at the
edges of our clients' considerations: this is partly due to the intrinsically
tailored nature of Tagmatic's models, but mostly attributable to a plethora of
invaluable qualities that go beyond the naked F1 scores: being able to quickly
spin up a model on your own content and vocabulary, and then interact with it
via a minimalistic HTTP interface, zero maintenance, on-the-fly learning of new
classes, low latency, security, scalability and high availability out of the
box, are crucial factors to enterprise-scale deliveries. 

Your Intellectual Property lies in your content, your data, your data models,
and how you use the classifications. It is not in commoditized compute
solutions.
Your resources are thus more effectively spent developing and exploiting your
market differentiating IP, rather than spending way too much time and money on
trying to replicate the engineering complexities of off-the-shelf MLaaS
solutions which may not be your technical team's core competence.

This is where Tagmatic comes in. The hard work and complex engineering has been
done, just bring your content!

The datasets
RCV1 is arguably the most popular dataset for multi-label classification,
comprising 800,000 newswires by Reuters tagged with 103 classes. Oddly enough
for such a totemic dataset, the breadth of the vocabulary is anachronistic when
compared to some of our clients, modern global publishers who routinely maintain
taxonomies with thousands of topics.
The dataset is available here
[http://www.ai.mit.edu/projects/jmlr/papers/volume5/lewis04a/lyrl2004_rcv1v2_README.htm] 
although unfortunately not in its original XML format with full HTML markup,
which we would normally leverage as part of the classification task. Also the
corpus comes already tokenised and train/test datasets pre-canned. Besides
taking away part of the fun, this has also negative implications on the
vectorisation logic Tagmatic implements, making it largely ineffective
unfortunately. For the purpose of this benchmark, we have used the full train
set (23,000) and a reduced test set (100,000).

AAPD (Arxiv Academic Paper Dataset) is a collection of research papers covering
a number of subjects organised in several classification systems. The dataset 
here [https://www.kaggle.com/neelshah18/arxivdataset] is annotated with a number
of classification schemes, of varying level of quality. We have carved out a
subset corpus in the Computer Science domain, comprising of 32,000 papers and 39
subjects in total.

The benchmark
We have considered a number of models amongst the most widely employed for
multi-label classification tasks. The reference evaluation metric is
weighted-F1.

RCV1 AAPD One-vs-Rest - LinearSVC 0.78 0.71 One-vs-Rest - SGD 0.74 0.62 
One-vs-Rest - XGBoost 0.75 0.64 Classifier Chain - LinearSVC 0.79 0.71 Label
Powerset - MultinomialNB 0.55 0.55 Label Powerset - RandomForest 0.60 0.55 
Tagmatic 0.80 0.74 
Tagmatic offers out-of-the-box improvement over a number of basic
implementations. Granted, a team of Data Scientists might possibly come up with
better optimised flavours of such models, pushing performances up. But at what
cost? And by how much? And what then? There is still the much bigger task of
bringing that model to fruition in a real world environment, requiring
non-trivial development effort. Is that really worth it? 
Unless text classification is either your USP or a strong market differentiator,
then the answer ought to be a resounding no. 

Tagmatic is production-ready, enterprise-grade, and available to you right now -
see our Text AI Services
[https://datalanguage.com/products/datalanguageai/text-ai-services] product.

Lastly - Using Tagmatic (Text AI Services) on Hansard
We have also benchmarked Tagmatic against Hansard data: a corpus of ~200,000
parliamentary interventions for the 2005/6 period , classified with ~7200
categories, scoring a whopping 0.83. More on this on the next blog post. 
Make sure you sign up for our newsletter [ https://datalanguage.com/contact-us] 
and follow @datalanguage [https://twitter.com/datalanguage] on Twitter for all
the latest.

News and product updates about our Text AI Services. Read about how our Text AI products are being used in innovation projects, and by our clients.

Read the latest on our Text AI Services. Projects, Innovation initiatives, and new developments for AI features in specialized use cases.

News and Product Updates about our Text AI Services

Text AI Services

Benchmarking Tagmatic

For Data Science to truly live up to expectations, it must embrace software craftsmanship.

Data Science is about building software

Earlier this month I was chatting with my old friend Pierluigi Casale
[https://www.linkedin.com/in/pierluigi-casale-41271430/] (Data Scientist at
TomTom, AI lecturer and researcher) about what seems to be the question of the
century: "What is Data Science"? Tongue in cheek, I offered my very own answer:
"Data Science is that thing you get when someone with a PhD plays with a Jupyter
Notebook, and a technical architect refuses to let it anywhere near production
because it doesn't have a single unit test". 
Although sporting a PhD himself, Pierluigi agreed that my slightly caustic
definition was indeed something he observed himself, and that he too was wary
of. He mentioned how he often concludes the many talks he gives at conferences
around the world with a diagram from the notorious paper "Hidden Technical Debt
in Machine Learning Systems
[https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf]
" by D. Scully et al.


There is indeed a widespread assumption that Data Science and Machine Learning
belong into a hazy limbo somewhere between data analysis, analytics, and R&D.
Within this mystical space, a set of practices otherwise regarded as inefficient
seem to be accepted as the norm: manual mining and manipulation of data; one-off
transformations virtually impossible to replicate; casual scripting of code with
total disregard of any development standard; artefacts laying around and code
running here or there...
When Scully talks about "glue code" and "plumbing", he is pointing at the right
problems but possibly underplaying the issue by being too humble about it. So
forgive my bluntness and let me put that straight: Data Science is about
building software.

Whether the context is academia or enterprise, research or commercial alike, the
adoption of Software Development best practices, tools and methodologies is the
key for efficient delivery of outcomes. At Data Language we have worked with a
number of clients at different levels of "enlightenment", and witnessed first
hand how injecting software craftsmanship into Data Science can lead to
profoundly transformational effects. We can also testify how the development of
our Tagmatic [https://datalanguage.com/tagmatic] service has been indeed 10%
Machine Learning and 90% software engineering: not much we could have done with
our shiny model without things like containerisation, high availability, high
throughput, load balancing, scaling, automation, security, governance,
monitoring, and many others. 

The adoption of established best practices in Software Development will boost
productivity, shorten the feedback cycles, and produce more robust code that is
easier to change, maintain, and ship.

Six things we at Data Language do and you should too
1. Ditch the darn notebooks
I'm fully aware this is highly controversial bordering sacrilegious for many,
but the truth is these type of editors foster bad habits and are holding you
back. They are like comfy, safe, cute trikes you got used to riding when you
were three; but if you actually want to go somewhere other than your garden,
you’ve got to move on and get yourself a bike. 
At this year's JupyterCon in NY, the talk widely regarded as best of the lot was
"I don't like notebooks.
[https://conferences.oreilly.com/jupyter/jup-ny/public/schedule/detail/68282]"
by Joel Grus. I wholeheartedly invite you to check it out.

2. Version control
A common refrain in favour of notebooks, is that they enable collaboration.
Oddly enough, when further engaged on this topic and asked about which version
control tool they use, what's their branching strategy, how often they merge,
etc... it all unravels. 
Collaboration is mighty important, so make sure you get on top of your version
control strategy. (And ditch the darn notebooks)

3. Test automation and CI
Arguments for Test Driven Development and Continuous Integration are widely
available and universally recognised as the single crucial factor that enables
faster, safer, better code. The very same arguments apply to Data Science. The
ability to move faster with your code enables quicker feedback and shorter
iterations: you will be able to experiment and prove ideas quicker, boosting
productivity. You will also end up with code that is designed better, and it can
be transitioned into a production environment more easily.

4. Infrastructure as code
In order to take full advantage of the ongoing democratisation of computing
power, it is crucial to write code that is portable and can be deployed wherever
is faster to run it. In a typical development setting, we want be able to
provision computing resources in the cloud, deploy our code there, execute it,
collect outputs and dispose of the remote resources. And we want to be able to
do this in a fully automated fashion, at the push of a button.
Many ways to skin this cat, but if I were to single out one thing you should do,
and learn to do well, is containerisation. Plenty of great Docker learning
resources around: get stuck in!

5. Data management
As in any other scientific setting, it is mighty important for experiments to be
repeatable under the same conditions. We want to make sure immutability of
inputs is preserved through transformations. Versioning and artefacts management
tools play an invaluable role in the development of Data Science solutions. This
goes hand in hand with practices discussed in points 2) and 3), and falls
naturally from there.

6. Metadata management
Often the journey is just as valuable as the destination. As much as possible
(and to the extent that is useful, of course!) we want to be able to retain key
information about the process at hand, so that we can better interpret, analyse
and enhance its outputs. The spectrum here is broad, from basic logging to
systematic capturing and analysis of metadata. The point is: we want to be able
to fulfil this function in the way that maximises the insights offered by the
data being collected. 
Metadata management is at the core of many activities that are either routine
(eg compare performance of models), or required for production (eg backout
strategy for a failing data pipeline), or further reaching into the realms of
ethics and accountability (eg quantitative justification of a Machine Learning
prediction). It is important we do it well.

Conclusions
I fully appreciate some of the recommendations above may feel daunting to the
more traditional Data Scientists, due to such activities not (yet!) being
typically associated with their profession. To some extent, we witnessed
something similar in the past with network and infrastructure specialists, now
evolved in DevOps that design and build infrastructure solutions through code,
fully embedded in the fabric of development teams from the onset, sharing tools
and methodologies. 
Similarly, the key for enabling the evolution of Data Science from R&D practice
to fully fledged enterprise function, lays in bringing data scientists and
developers closer: before long the distinction will blur, and specialist
knowledge will thrive on a common substrate of working practices. 

At the end of the day, remember, Data Science is about building software.

We are specialists in the design, implementation and deployment of Scalable Data Science solutions. We deliver practical data science implementations for both our own SaaS products, and for our clients' bespoke platforms and products, for the analysis of data and media at scale.

We are specialists in the design, implementation and deployment of Scalable Data Science solutions. Find updates, News and insights on Data Science here.

Data Science - our Work, Features, News and Updates

Data Science

Technical Debt is slippery, no doubt about it.

Its very definition often misunderstood; its formal quantification elusive; its
causes/effects all too frequently brandished as the ultimate weapons of blame
allocation (or indeed self-justification). Perhaps it's time we stop a second
and actually take the time to understand what Technical Debt is, how it
typically manifests itself, and reflect on viable, sustainable strategies to
control it. In fact, by the end of this post, I will indeed argue the case for
its rehabilitation.

As a short preamble to the problem space, I need to introduce the notion of
conflicting interests. Far from me subscribing to the toxic notion of Product
Team vs Technical Team, it must be recognised though that in any sizable
endeavor (especially at enterprise level) there are different interests,
concerns, roles and responsibilities within the delivery team. I'll spare
everyone the obvious paragraph about technically savvy POs/PMs, commercially
astute technical architects and business focussed developers... how they
actually exist and how can all work together in harmony. What I'm interested in
here is to define two groups of interests (not people) which I'll lazily call 
Product Interests (PI) and Technology Interests (TI). Each one embodies
different and sometime conflicting needs, aspirations, drivers and objectives.


What it is and how it comes about
The expression "Technical Debt" stems from an analogy with the financial loans
market, first attributed to Ward Cunningham
[https://www.youtube.com/watch?v=pqeJFYwnkjE] in the early 1990's. Put simply,
producing and shipping artefacts of sub-optimal quality is the equivalent of
contracting a debt: opportunities are levereged, the immediate cost is deferred
on a promise of repayment, and interest become payable until settlement. In the
analogy, interest are constituted by a cumulative decreased agility of
intervention, and the principal repayment equals to the replacement of the
temporary, sub-par solution with a better one.

Access to credit enables the realisation of opportunities otherwise
unattainable. As such, it is crucial to recognise that Technical Debt is a
choice dictated by the pursuit of an advantage, as part of the execution of a
precise strategy. It has very little to do with a set of malpractices that are
often conflated in the mainstream notion of Technical Debt: missing
documentation, scarce test coverage, non-existing refactoring, lack of basic
standards, and general disinterest in quality have more to do with poor
technical craftsmanship than with the gain of advantage through deferral of
costs. These are part of core TI's concerns, and a zero-tolerance policy can be
an appropriate: developers are expected to carry these out without PI's prompt,
authorisation, or sign-off on a backlog ticket.
Equally, there are examples of reckless ingerence of PI in TI (eg
vendor/technology selection, refusal to innovate, distrust in industry best
practices, distrust in experts, etc..), that shouldn't be classed as Technical
Debt either. Causes for this sort of idiosyncrasies are rooted way deeper in the
fabric of the organisation and are beyond the topics of this post.

Technical debt is a choice commonly associated to a lack of time, resources,
experience, or knowledge - in pretty much any combination.
In the age of Agile and Lean delivery, it's not infrequent for TI to have a
stake in the fast release of features with the main purpose of acquiring
precious feedback from both the users and the surrounding environment. Proof Of
Concepts are prime examples of this. Iterative processes inheritedly carry some
degree of redundancy which is accepted by PI, but in itself represent
progressive enhancement of a solution that is initially sub-optimal. As new
learnings are acquired, experience will contribute to better designs for the
next iteration and faster shipment of better products.

Most often though, Technical Debt is associated with a different scenario. In
order to keep up with changing market conditions and new customers demands, take
full advantages of windows of opportunities, or even gain first mover advantage
(all very typical PIs), ambitious delivery schedules are set, and these soon
become the root cause of shortcuts and sub-optimal solutions. The choice to
contract a Technical Debt is both explicit and often understandable.
Unfortunately, running interest and total debt are often neither acknowledged
nor honoured with the same determination, giving way to the undesirable side
effects most commonly associated with Technical Debt.


The effects
Whilst increased speed of delivery, associated learnings and market advantages
are all arguments in favour of a responsible Technical Debt strategy, when badly
executed it can have catastrophic effects on an enterprise. Just like financial
debt, accumulation of unconsolidated debt can bring companies to their knees:
the weight of interest as a self-inflicted handicap on the ability to adjust and
evolve, has the potential of bringing organisations into a deadlock of forced
stasis. There are countless examples of badly managed Technical Debt that,
blinded by the short term advantages, have neglected its management resulting in
lengthy and costly next iterations, innovation more and more difficult to
implement, PI and TI engaging in open conflict with mutual allocation of
responsibility, and people losing enthusiasm, interest, and ultimately jumping
ship.

Like productivity and total cost of ownership, Technical Debt too suffers the
undeniable difficulty of quantifying it. The loan analogy comes comes all but
short here: how do you measure exactly how much is due and by when, what the
interest rate is, and whether in mutating conditions the borrowed capital is
ever repaid at all.

Interestingly, when it's in both the TI and PI advantage of cutting corners,
there's often an implicit understanding of the boundaries of the Technical Debt
extent, its duration and repayment terms. It's probably this very alignment of
intents and objectives that have made "alpha versions" and "first iterations"
slip through the nets of Technical Debt classification.
There is therefore a strong argument for this very aspect to become the first
ingredient in the recipe for Technical Debt management.


Managing Technical Debt
Share objectives and motivations
All too often the tension between PI and TI stems from the former calling the
shots and the latter being at the receiving end (frequently being left to mop up
the resulting mess too). It's not by chance that in highly cohesive teams where
PO/PM are deeply embedded in the development team and technologists are involved
in commercial strategy, Technical Debt is virtually not-existent. Or rather its
perception is. The same hard choices are made and some degree of shortcuts
become accepted because it's the team as a whole that has a stake in it.

Call it, own it
For effective accountability to be enforced, it is necessary to openly recognise
Technical Debt and explicitly own it. Something as simple as a registry can have
a massive positive impact on the perception of the debt, and the credibility of
the borrower.

Stay on top of it
We already discussed how effective measurement of Technical Debt can be
challenging, but so are many other indicators for which frameworks and
methodologies have been put in place (productivity, velocity, cost of ownership,
cost of change, etc..). Once recognised and owned, it should be easier to openly
refer to it as any other aspect of product delivery. Time must be allocated to
the restitution of borrowed capital in the same way we would commit to a
mortgage repayment plan. It just something that needs budgeting for, it's part
of a delivery strategy, and as such must be managed openly and consciously.

Don't borrow beyond your means
When contracting a debt to enable immediate advantage, a longer term view must
be held to assess our capacity to sustain the debt (and its interest). Technical
Debt inevitably implies a loss of agility in implementing changes: innovation
becomes harder, lengthier and more costly. The temptation to face these
challenges through additional borrowing can lead to disastrous outcomes:
organisations becoming slaves of their own inability to evolve with the changing
market conditions, every adjustment so costly that there's no room for error
(hence for learning), compounding the problem of stiffness one round of
borrowing after the other.

Pay back, and make a big deal of it
Countless example of irresponsible borrowing and missed repayments have dented
the credibility of loanees to the point where perfectly sensible strategies are
hidden away from the team, in the implicit assumption that they would be
misunderstood, undermined, and rejected. It's almost like a long string of
written-off debts have led to the average delivery manager having a terribly
poor (assumed) credit score.
Every time Technical Debt is duly repaid, it should be recognised, celebrated
and shouted out as loudly as possible. At least until it becomes the norm.


Conclusions
Too long a string of failed executions has already twisted the very definition
of Technical Debt, confusing the perception of undesiderable (but avoidable)
outcomes with its original intent as enabler. Tech Debt must be treated as a
strategy in itself - not a synonym of failed strategy as is typically the case.

Technical Debt is ultimately a problem of trust. Incurring Technical Debt is a
perfectly viable, sustainable practice that can lead to competitive advantage,
but it's vital for trust to be won back through open ownership, accountability
and timely repayment.

Blogs by Fabio Colasanti