Sunday, March 5, 2017

Regulation of Emergence and Ethics of Algorithms



1. Introduction


Algorithms governance is a key topic, which is receiving more and more attention as we enter this 21st century. The rise of this complex and difficult topic is no surprise, since “software is eating the world” – i.e., the part of our lives that is impacted by algorithms is constantly growing – and since software is “getting smarter” every year, with the intensification of techniques such as Machine Learning or Artificial intelligence. The governance question is also made more acute since smarter algorithms are achieved through more emergence, serendipity and weakening of control, following the legendary insight of Kevin Kelly in his 1995 “Out of control” best seller: “ « Investing machines with the ability to adapt on their own, to evolve in their own directions, and grow without human oversight is the next great advance in technology. Giving machines freedom is the only way we can have intelligent control. » Last, the algorithmic governance issue has become a public policy topic since Tim O’Reilly coined the term “Algorithmic Regulation” to designate the use of algorithms for taking decision in public policy matters.

Algorithm governance is a complex topic that may be addressed from multiple angles. Today I will start from the report written by Ilarion Pavel and Jacques Serris “Modalities for regating content management algorithms”. This report was written at the request of Axelle Lemaire and focuses mostly on web advertising and recommendation algorithms. Content management – i.e. deciding dynamically which content to display in front of a web visitor – is one of the most automatized and optimized domain of the internet. Consequently, web search and content recommendation are domains where big data, machine learning and “smart algorithms” have been deployed at scale. Although the report is focused on content management algorithms, it takes a broad view of the topic and includes a fair amount of educational material about algorithms and machine learning.  Thus, this report addresses a large number of algorithm governance issues. It includes five recommendations about algorithm regulation intended for public governance stakeholders with the common intent of more transparency and control for algorithms that are developed in the private sector.

This short blog post is organized as follows. The first part provides a very simplified summary of the key recommendations and the main contribution of this report. I will focus on a few major ideas which I found quite interesting and thought-provoking. This report addresses some of the concerns that occur from the use of machine learning and artificial intelligence in mass-market services. The second part is a reply from the angle of our NATF work group on Big Data. As was previously explained, I find that we have entered a “new world” for algorithms that could be described as “data is the new code”. This cast a different shadow on some of the recommendations from the Ilarion Pavel & Jacques Serris report. As algorithms become grown from data sets through training protocols, it becomes more realistic to audit the process than the result. The last part of this post talks about the governance of emergence, or how to escape what could be seen as an oxymoron. The question could be stated as “is there a way to control and regulate something that we do not fully understand ?”. As a citizen, one expects a positive answer. Other sciences have learned to cope with this question a long time ago, since only computer scientists from Silicon Valley believe that we may control and fully understand life today (these issues arise constantly in the worlds of medicine, protein design or cellular biology for instance). But the existence of this positive answer for Artificial Intelligence is a topic for debate, as illustrated by Nick Bostrom’s book “Superintelligence – Paths, Dangers, Strategies”. To dive deeper into this topic, I strongly recommend the reading of "Code-Dependent : Pros and Cons of the Algorithmic Age" by Lee Rainee and Janna Anderson


2. Algorithm Regulation


First, I should start with my usual caveat that you should read the report versus this very simplified and partial summary. The five recommendations can be summarized as follows:

  • Design a software platform to facilitate the study, the evaluation, and the testing of content / recommendation algorithms in a private/public collaboration opened to research scientists
  • Create an algorithm audit capability for public government
  • Mandate private companies to communicate about algorithm behavior to their customers, through a “chief algorithm officer role”
  • Start a domain-specific consultation process with private/public stakeholders to formalize what these “smart content management services” are and which best practices should be promoted nationally or internationally.
  • Better train public servants who use algorithms to deliver their services to citizens

A fair amount of the report talks about Machine Learning and Artificial Intelligence, and the new questions that these techniques raise from an algorithm ethic point of view. The question “how does one know what the algorithm is doing” is getting harder to answer than in the past. On page 16, the concept of “loyalty” (is the algorithm true to its stated purpose ?) is introduced and leads to an interesting debate (cf. the classical debate about the filter bubble). The authors argue – rightfully – that with the current AI & ML techniques the intent is still easy to state and to audit (for instance because we are still mostly in the era of supervised learning), but it is also clear that this may change in the future.  A key idea that is briefly evoked on page 19 is that machine learning algorithms should be evaluated as a process, not on their results. Failure to do so is what triggered the drama of the Microsoft chatbot who was made non-loyal (not to say racist and fascist) through a set of unforeseen bet perfectly predictable interactions. One could say there is the equivalent of Ashley’s law of requisite variety in the sense that the testing protocol should exhibit a complexity commensurate to the desired outcome of the algorithm. Designing training protocols and data sets for algorithms that are built from ML to guarantee the robustness of their loyalty is indeed a complex research topic that justifies the first recommendation.

We hear a lot of conflicting opinions about the threat of missing the train of AI development in Europe or in France, compared to the US or China. The topic is amplified by the huge amount of hype around AI and the enormous investments made in the last few years, while at the same time there seems to be a “race to open source” from the most notorious players. The authors propose three scenarios of AI development. In the first scenario, the current trend of sharing dominates and produces “algorithms as a commodity”. AI becomes a common and unified technology, such as compilers. Everyone uses them, but differentiation occurs elsewhere. The second scenario is the opposite where a few dominant players master the smart systems (data and algorithms) at a skill and scale level that produces a unique advantage. The third scenario focuses on data ecosystems but recognizes that the richness and regulatory complexity of data collection make it more likely to see a large number of “data silos” emerge (larger number of locally dominant players, where the value is derived more from the data than the AI & ML technology itself). As will become clear in the rest of this blog, I see the future as the combination of 2 and 3 : massive concentration for a few topics (cf. Google and Facebook) that coexists with a variety of data ecosystems (if software is eating the world and tomorrow’s software is derived from data, this is too much to chew for a single player, even with Google’s span).

A key principle proposed by the authors is to “embody” the algorithm intent through the role of “chief algorithm officer”, with the implicit idea that (a) algorithms have no will or intent of their own, that there is always a human behind the code (b) companies should have someone who understands what the algorithm does and is able to explain it to stakeholders, from customer to regulators. The report makes a convincing case that “writing code that works is not enough”, the of “chief algorithm officer” should be able to talk about it (say what it does) and prove that it works (does what is intended). There is no proof, on the other hand, that this is feasible, which is why the topic of algorithm ethics is so interesting. The authors recognize on page 36 that auditing algorithms to “understand how they work” is not scalable. It requires too much effort, will prove to be harder and harder as techniques evolve, and we might expect some undecidability theorems to hit along the way. What is required is a relaxed (weaker) mandate for algorithm regulation and auditing: to be able to audit the intent, the principles that guarantee that the intent is not lost, and the quality of the testing process. This is already a formidable challenge.

3. Data is the New Code


This tagline means that the old separation between data and code is blurring away. The code is no longer written separately following the great thinking of the chief algorithm officer and then applied to data. The code is the result of a process – a combination of machine learning and human learning – that is fed by the available data. “Data is the new code” was introduced in our NATF report to represent the fact that when Google values software assets for acquisition, it’s the quantity and quality of collected data that gives the basis for valuation. The code may be seen as the by-product of the data and the training process. There is a lot of value and practical expertise with this training process, which is why I do not subscribe to the previously mentioned scenario of “AI as a commodity”. Smart systems is first and foremost an engineering skill.

A first consequence is that the separation of the Chief Data Officer from the Chief Algorithm Officer is questionable. The code that implements algorithms is no longer static, it is the result of an adaptive process. Data and algorithms live in the same world, with the same team. It is hard to evaluate / audit / understand / assess the ethical behavior of data collection or algorithms if the auditor separates one from the other. Data collection needs to be evaluated with respect to the intent and the processes that are run (which has always been the position of the CNIL) and algorithms are – more and more, this is a gradual shift – the byproduct of the data that is collected.

Data ethics is also very closely related to algorithm ethics. On page 29, the report tells that bias in data collection produces bias in the algorithms output. This is true, and the more complex the inference from data, the more complex tracking these biases may be. The questions about the ethics of data collection, the quality and the fidelity of the data samples, are bound to become increasingly prevalent. As explained before, this is not a case where one can separate the data collection from the usage. To understand fairness – the absence of biases - , the complete system must be tested. Serge Abiteboul mentioned in one of his lectures the case of Staples, whose pricing mechanism, through a smart adaptive algorithm, was found to be unfair to poorer neighborhood (because the algorithm “discovered” that you could charge higher prices when there are fewer competitors around). I recommend reading the article “Discovering Unwarranted Associations in Data-Driven Applications with the FairTest Testing Toolkit” to see what a testing protocol / platform for algorithm fairness could look like (in the spirit of the first recommendation of the report). The concept of purpose is not enough to guarantee an ethical treatment of data, since many experiments show that big data mining techniques are able to “find private pieces of data from public ones”, to evaluate features that we not supposed to be collected (no opt-in, regulated topics) from data that were either “harmless” or properly collected with an opt-in. Although the true efficiency of the algorithms of “Cambridge Analytica” are still under debate, this is precisely the method that they propose to derive meaning full data traits from those that can be collected publicly.

The authors of the report are well aware of the rising importance of emergence in algorithm design. On page 4, they write “one grows these algorithms more than one writes them”. I could not agree more, which is why I find the fourth recommendation surprising – it sounds too much of a top-down approach where data services are drawn from analysis and committees versus a bottom-up approach where data services emerge from usage and collected data. In the framework of emergent algorithm design, what needs to be audited is no longer the code (inside of the box which is becoming more of a black box) but the emergence controlling factors and the results:
  • Input data
  • Purpose (intent) of the algorithm
  •  “training” / “growing” protocol
  •  Output data

This brings us to our last section:  how can one control the system (delivering a “smart” experience to a customer) without controlling the “black box” (how the algorithm works) ?

4. How to Control Emergence ?


The third recommendation tells about the need to communicate about the way algorithms operate. Following the previous decomposition, I favor the recommendation on communicating about intent, with the associate capability (recommendation #2) to audit the loyalty (the algorithm does what its purpose says). On the other hand, I do not take this literally to explaining how the algorithm works. This was perfectly achievable in the past, but emergent algorithm design will make it more difficult. As explained earlier, there are many reasons to believe that it may simply be impossible from a scientific / decidability theory view point.

This is still a slightly theoretical question as of today, but we are coming fast to a point when we will truly no longer understand the solutions that are proposed by the algorithms. Because AlphaGo is using reinforcement learning, it has been able to synthetize strategies that may be qualified as deceiving or hiding its intent to the opponent player. But humans are very good at understanding Go strategies. In the case of the recent win of AI in poker tournaments, it is trickier since we humans have a more difficult time at understanding randomized strategies. We have known this from game theory and Nash equilibriums for a long time. Pure strategies are easier to understand but mixed strategies are often the winning ones. Some commentators assess that the domination of the machine over human is even more impressive for Poker than for Go, which to me reflects the superiority of the machine to handle mixed (i.e. randomized) strategies. As we start mixing artificial intelligence with game theory, we will grow algorithms that are difficult to explain (i.e., we will explain the input, the output, the intent and the protocol, not what the algorithm does). If one only uses a single AI or machine learning technique, such as deep learning, it is possible to still feel “in control” of what the machine does. But when a mix of techniques is used, such as evolutionary game theory, generative AI, combinatorial optimization and Monte-Carlo simulation, it become much less clear. As a practitioner of GTES (Game Theoretical Evolutionary Simulation) for a decade, it is very clear that the next 10 years of Moore Law will produce “smart algorithms” with deep insights from game theory that will make them able to interact with their environment – that is, us – in uncanny ways.

I have used the “backbox” metaphor because a systemic approach to control “smart algorithm” is containment, that is isolate them as a subsystem in a “box of constraints”. This is how we handle most of the other dangerous materials, from viruses to radioactive materials. This is far from easy from a software perspective, but there is no proof that it is impossible either. Containment starts with designing interfaces, to ensure what the algorithm has access to, and what outcome/ suggestions it may produce. The experience of complex system engineering shows that containment is not sufficient, because of the nature of complex interaction that may appear, but it is still a mandatory foundation for safe system design. It is not sufficient for practical reasons: the level of containment that is necessary for safety is often in contradiction with the usefulness of the component. Think of a truly great “strong AI” in a battery powered box with no network connection and a small set of buttons and lights as an interface. The danger of this “superintelligence” is contained, but it is not really useful either. The fact that safety may not come solely from containment is the reason we need complex / systemic testing protocols, as explained earlier.
Another possible direction is to “weave” properties into the code of the emergent algorithm. It is indeed possible to impose simple properties onto complex algorithms, that may be proven formally. 

The paradox is that there are simple properties of programs, such as termination, which are undecidable, while at the same time, using techniques such as abstract interpretation or model checking, we may formally prove properties about the outputs. For my more technical readers, one could imagining weaving the purpose of the algorithm using aspect-oriented programming into a framework that is grown through machine learning. This is the implicit assumption of the scifi movies about Asimov’s laws that are “coded into the robots” : they must be either “weaved” into the smart brain of the robot or added as a controlling supervisor – precisely the containment approach, which is always what gets broken in the movie. The idea of being able to weave “declarative properties” – that capture the intent of the algorithm and may be audited – into a mesh of code that is grown from data analysis is a way to reconcile the ambition of the Ilarion Pavel and Jacques Serris report with the reality of emergent design. This is a new field to create and develop, in parallel with the development of AI and machine learning in software that is eating the world. This will not happen without regulation and pressure from the public opinion.


These are not theoretical considerations because the need to control emergent design is happening very soon. Some of these concerns are pushed away by creating divides: “weak AI” that would be well controlled versus “strong AI” that is dangerous but still a dream, “supervised machine learning” that is by definition under control, versus “unsupervised learning” which is still a laboratory reseach topic. The reality is very different: these are not hard boundaries, there is a gradual shift day after day when we benefit from more computing power and more data to experiment with new techniques. Designing methods to control emergence requires humility (about what we do not know) and paranoia (because bad usage of emergence without control or foresight will happen).

 
Technorati Profile