Robustness in trustworthy AI

Simplicity ad nauseam.

When discussions turn to artificial intelligence, there is an expectation of complexity. Its implementation, however, is all about simplicity, about making something complex so simple that even a computer could run it.  

Actually, everything a computer does can be explained by a simple courtesy rule in social interactions: if everyone is silent, you have to talk, as silence is embarrassing (especially, for non-Finns), otherwise you should be silent (definitely a Finnish type of rule). 

OK, let me clarify the rule:  with an input of two zeros, the output is one, all the other three two-input combinations of ones and zeros must produce a zero. This is called a NAND (not and) logic gate, and it is universal, meaning that all logic can be built out of it, just by connecting NAND gates to each other in a multitude of different ways. This is robust, explainable, and transparent. The errors computers make in gate level are extremely rare – this means that their behavior is also reproducible. We have to conclude that artificial intelligence in computers is really trivial and only artificially intelligent. 

Are all robust rules robust?

In discussions about algorithmic decision making, a concept of rule-based systems have emerged. This is used to bring forward the idea that especially simple rule-based decisions are proper, explainable and acceptable. But, unfortunately, the truth is that simple, robust rules can have complicated consequences that are not simple, not even robust.

As an example, let us consider a robust, algorithmic, rule-based decision that is important and has been used for a century in Finland: you start your elementary school in the autumn of the year in which you turn seven years old. (“Oppivelvollisuus alkaa sinä vuonna, jona lapsi täyttää seitsemän vuotta.”) It is not essential if humans or machines are “running” this algorithm. So, let us say that we are agnostic, and do not care. The accuracy of the decision is what really matters.

However, it has been found that this rule has consequences that are discriminatory based on age, not because of the year seven, but because those born early in the year fare better at school versus those born later in the year. This relative age effect lasts across your life, even after it should not matter anymore.

These unintended consequences of a robust rule will now, when the effect is known, lead to second degree complications when people are knowledgeable of it and start to tune the birth dates of their progeny, initiating a ripple of effects echoing in the society. Also, I am waiting for the first law suits requiring lower entrance criteria to universities for late in the year born. They have a case. Currently, the law is definitely discriminating and against fundamental human rights that particularly forbids discrimination based on age. 

One has to ask is this rule robust after all, as its consequences were so unexpected. There is indeed a dichotomy between simple, easily explainable systems, and their performance. Simple robust rules are too simple. We have to start tuning these, adding exceptions and remedies. 

Robustness

Simplicity of the rule is no guarantee of the robustness of the algorithm in real world. The real evaluation comes for practical deployments, and from real data of its consequences. 

Therefore, the decisive test on rules have to be made in realistic conditions with deployed systems using sand-boxing or in medical terms in-vivo. We cannot reliably predict the consequences ex-ante. Definitely, shooting-in-the-dark-regulation and legislation should be avoided.

The essential question in robustness is about consistent performance. Does the AI deployed make better decisions than humans, in practice, does it save human lives in traffic?

In an algorithmic sense, we should find a solution that is as simple as possible, pruning the number of rules to the minimum, from billions to millions, from millions to thousands, and compartmentalize the algorithm, so that millions of rules could represent a function that humans intuitively understand as a single entity, like recognizing cats.

In a security sense, we should make sure that integrity of the system is preserved against malignant actions like adversarial attacks, hacking, breaches of privacy, or just power outages.

In machine learning there is ever more a realization how much more data we need to make it robust, so that a cat with elephant skin will not be an elephant, or a wolf in a city would not be a dog. There are ways out, but work is needed. Self-supervised learning helps a lot to build robust priors from the plentitude of unannotated data. Training algorithms are also becoming more efficient, and can distill the essential from less data.  

These problems do not really require a particular AI ethical code, detached from those that we already have in the usual state of the art practice in SW and HW engineering. 

Ethical robustness

In an ethical guideline sense, we are looking at the consequences and the means. Does the deployment of a system double the workload of doctors, and lead to extra deaths?  Will the use of a keyboard as an input device lead to worse learning results at schools? Is data gathered in an appropriate way that does not produce dire consequences to the people who have given consent on the use of the data? Are the means in balance with the goals? Are the systems robust against tendencies of misuse? 

What is particular in robustness of artificial intelligent is its ability to automate areas in an unprecedented scale and level that were beyond the reach of automation before. This makes the range of consequences much more varied than we get with earlier phases of digitalization. Will the platformisation of industries lead to unwanted changes in the balance of powers between employees and employers?  To what extent should our society be robust to these changes or should it be? 

New ethical principles and etiquettes

A final act of simplification could be the one that produces, with simple local rules, the desired large-scale characteristics by using emergent complexity. This is the way life creates complexity for flocks of birds or schooling fish. Now, one could use computers to find these local, simple rules, say, for traffic lights that would make the flow of traffic smooth in a city.

C. Gershensonfrom National Autonomous University of Mexico has done just that. He claims that the best way to solve the problem is to give traffic lights of each intersection the ability to follow the local situation, and then give a priority to the side that has the heaviest traffic. This does sound sensible, doesn’t it? 

This solution would not require centralized, large scale, city level models, but just a reactive, simple traffic light rule that self-organizes the traffic flow of the city to platoons in the emerging green light zones. Everyone can understand the modus operandi, and the result is still a city wide, interactive solution that is claimed to beat all others.

However, these simple solutions for complicated problems are seldom and far apart. In most situations we have to face the true complexity of complexity.

Currently, the ethical question of bad consequences of actions with good intentions is present more and more. The development of AI in the digitalised society is giving us the possibility to understand the society better, and to find out which actions have desirable consequences, which not. So, one can not only ask what ethics can do for making AI deployment good, but also what AI can do for figuring out what is ethical – those simple rules that self-organize life and society to flow in a complex, optimal and good way.

Author: Leo Kärkkäinen