Two IMFD research studies stand out at NeurIPS

An innovative architecture for graph neural networks and a tool that allows users to specify explanations for artificial intelligence systems are two IMFD papers that will be presented at NeurIPS (Neural Information Processing Systems), one of the most relevant conferences in computing and computer science held internationally, which this year will take place virtually in December.

With a focus on artificial intelligence and machine learning, thousands of research centers, institutions, companies and universities from around the world work at the forefront of these topics and present their best work at this annual event. One of the reports that stands out in the 35th edition of the conference -which was originally conceived by figures such as Ed Posner, one of NASA’s most renowned scientists- is “Graph Neural Networks with Local Graph Parameters” .

In the study, the researchers propose and test an architecture that allows for a trade-off between simple and complex graph neural networks.

Neural networks are the most popular architecture today in machine learning. However, studies have found that these networks have a weakness: they make very little use of the underlying structures of the data, explains Pablo Barceló, director of the Institute for Mathematical and Computational Engineering of the Pontificia Universidad Católica de Chile and Deputy Director of the Millennium Institute for Foundational Research on Data (IMFD): “What these types of networks learn are correlations, and very little of the structure, which makes them less intelligent and less effective at detecting complex patterns in the data”. This means, he explains, that the network may need too much data to learn, or that it may not fully understand what it is learning, so that allowing the learning of this underlying structure would make training these systems much more effective and efficient.

 

In many cases, this underlying structure of the data is specified as a graph; for example, we can think of a social network with connections between participants, or another example could be public transportation routes: these are networks that have nodes and connections between those nodes. “There are a lot of structures there that traditional neural networks do not make much use of; graph neural networks were designed to exploit that structure and be able to learn those patterns more effectively”. However, since graph neural networks are more powerful, they are also more computationally expensive to implement since they require more resources, as their architecture is more powerful. “What we try to do in this work is find a trade-off: I am going to try to have the maximum computational power while trying to minimize the cost that this implies, because there are tasks that become impossible to carry out if the model is too complex”.

Within graph neural networks, there are simpler models and more complex ones. “The complex ones, at least theoretically, seem to be quite useful. We tried to find a good trade-off between the two, because although these are quite useful, they are not implementable. They cannot be put into practice because they are too complex. And the simple ones are too simple. So, we tried to find something that would allow us to combine the best of both worlds. We also tried to understand what their real computational power was, something that stood in the middle of these two worlds and that was simpler to implement”.

This architecture provides a middle ground that allows certain things to be detected: for example, cycles of a certain size, what are called cliques, which are subgraphs whose nodes are all connected to each other, without incurring the computational cost of more complex networks. “It seeks a trade-off: I need to detect this, the system says, well, detect it, pay upfront with a type of preprocessing, but do not keep paying for it afterward. It allows for the minimum possible cost to detect what one wants to detect without falling into the computational cost of more complex networks”.

This paper was co-authored by Pablo Barceló along with Floris Geerts and Maksimilian Ryschkov, from the University of Antwerp, and Juan Reutter, a professor at the Department of Computer Science of the Pontificia Universidad Católica de Chile, professor at the Institute for Mathematical and Computational Engineering UC (in a shared position), and researcher at IMFD.

“Foundations of Symbolic Languages for Model Interpretability” is the title of the second paper accepted at the NeurIPS conference, and its authors are Marcelo Arenas, Daniel Baez, Pablo Barceló, Jorge Pérez and Bernardo Subercaseaux. Arenas, who directs the Millennium Institute for Foundational Research on Data and is also a professor at the Institute for Mathematical and Computational Engineering UC (in a shared position) and at the Department of Computer Science UC, explains that the paper addresses fundamental challenges that currently exist in areas such as machine learning and artificial intelligence.

In both areas, he notes, applications and platforms have been emerging that operate quite efficiently in tasks such as image recognition or language translation. However, one of the problems that persists is how to explain the answers they generate. “Let us look at image classification systems. It is very impressive how they work and how they are able to recognize, for example, whether a photo shows a cat. But if you try to understand why that result was generated, you do not know”.

Arenas -whose specialties include database theory, logic in computer science and the semantic web- adds that the problem is that “as applications begin to make more sensitive decisions, one needs to have explanations”. In that sense, if a system says that “there was a dog instead of a cat, it does not matter, nothing is going to happen. But in the case of an automated hiring system, résumés are going to arrive and that platform is going to decide who to hire or who to make an offer to. We can also talk about a banking system that is going to determine whether I will be given a loan, for example”.

Eventually, a person could even demand clear explanations to find out why they did not get a job or why they were not given a loan. “There, obviously, one cannot just say ‘well, the machine told me no’ or ‘given the machine’s experience, it gave me a zero’. Furthermore, in these automated hiring systems it has been seen that there can be discrimination. Experiments have been done with exactly the same résumés, but “when you change, for example, a parameter like male to female, you see differences in terms of classification”, notes the academic. As systems make increasingly sensitive decisions, the situation becomes more critical: “If I have a car that is moving autonomously, I want to understand why it recognizes things in a certain way. If it does not recognize something, I need to understand why. If it did not recognize that there was a person at this intersection, I need to know the reason, because that could end in a catastrophe”.

This field is known as “Explainable AI” and there are many strategies to address it. Arenas notes that both he and the report’s co-authors chose a different path than the one usually chosen by those who work in this area. “Instead of saying ‘to explain, let us use this measure’ or ‘this other one’, what we said was ‘let us give the user the ability to explore the system’. What we wanted to do for that kind of exploration was to provide a tool where the user could go about asking the system things, like ‘I want to understand why it gave me this result’ or ‘if I had changed this feature, what would have happened’. Those are some possibilities, but there could be others, such as saying ‘how many features of the person do I need to change to get a positive classification?’ and have the system tell me ‘you need to change four’”.

That is precisely the central theme of the paper written by Arenas and the other researchers. The paper proposes the design of “a general language, where one can represent this type of properties that speak about a machine learning model”. In other words, what they propose “is a language that allows for making these queries with which one can build the chain” of explanations for, say, why a person’s loan application was rejected.

This language is based on a type of logic called FOIL, which, as Arenas explains, is a well-known logic that “has a good balance between expressiveness and evaluation complexity. We studied in detail what its characteristics or functionalities were, until we designed a logic in which we knew that everything could be evaluated efficiently, from a computational standpoint”.

The academic clarifies that the paper does not imply that the authors currently have a system where someone can sit down and start evaluating their model: “What we wanted to see was that the algorithms we had actually worked well. We managed to build a system with a reasonable interface where things could be handled in a reasonable amount of time, but there is still a very long way to go before it becomes a system that I could download right now, install on my computer and start using”.

Even so, the potential of the study is so high that the NeurIPS organizers decided to include it in the section reserved for the most outstanding papers, known as “spotlight”. Of the nearly 10 thousand studies that the event receives each year, only 20.1 percent are accepted, and just 3 percent of that total is considered for the section featuring the most relevant papers. “This is the most important conference in the field of machine learning. It is very important for us that the paper was accepted at this extremely competitive event, where all the artificial intelligence groups take part”, says Arenas.

“You come across all the major universities, companies like Google, all the big labs. The conference is so large that there is very little space for presentations, and these papers that are placed in the ‘spotlight’ are the ones that will get more time to be presented, because they are considered the most interesting”, says Arenas. An important aspect of the study’s acceptance process at NeurIPS was the discussion it sparked in the conference’s program committee: “It was quite heated. There were people who supported it, who said ‘this has to be included’, and there were other people who said ‘I am not convinced, this is not the way to do it’. That is exactly what you want to happen, instead of indifference”.