My Own Adventures in Causality

Everyday on LinkedIn, more posts mention causality. Causality is required for AGI, causality would underpin XAI, and lesser publicised uses of causality could improve performance of computer vision systems in fairly technical (mostly CVPR) papers… I’m not even mentioning the “correlation does not imply causation” statistical meme, to which I have already dedicated an article (albeit in French). We even see the rise of AI start-ups focussing on causality as their raison d’être (e.g. causalens , causality-link ...)

Yet, when we invoke causality in an AI context, do we fully realise the history, ramifications, and incredible complexity of the concept, from its very definition to its multiple applications? Reflecting upon the many facets of what we subsume under causality, I looked back to my own previous work, which has taken me through various combinations of AI and human-centred systems.?

Bertrand Russel did not think much of causality, which he characterised as “a relic of a bygone age”. Not the most helpful statement from Wittgensetin either, when he purported that causality “is not a law, but the form of a law” (TLP 6.32)? … Sic transit the logicians: we’ll touch upon this issue later when discussing how causality can be embedded in logical formalisms.?

If I would then be allowed an oversimplified history of causality, the prevalent view is that the four types of Aristotelian causality have, through the ages, been reduced to the only ‘efficient causation’ version, largely via Suarez, Descartes and Leibniz (as brilliantly explained in Vincent Carraud’s ‘Causa sive Ratio’), until it found its contemporary formulation through David Hume. With so many other philosophers involved, including Vico, Kant, and Zubiri, the prospect of a unified view became a rather remote one, not least because of the later autonomisation of scientific disciplines pertaining to causality, from Psychology to Physics, framing their own requirements and derived visions of causation. I shall not even dare venturing on the latter aspect (Physics), other than referring to the corresponding discussion in Bernard D’Espagnat’s ‘On Physics and Philosophy’...?

One useful distinction introduced by Galavotti is between ‘event causality’ and ‘property causality’: the former describes relationships between events in the physical world, while the latter accounts for causal inference from properties or variables. Galavotti’s distinction could be considered as following a user-centric perspective: conversely, Pearl’s mathematical framework for what he calls the ‘causal revolution’ includes: graphical models; structural equations; and counterfactual and interventional logic.?

Faced with my own inability to provide a synthesis, I have referred in previous discussions to the tuple {Michotte, Granger, Pearl} to account for the main theorists of causal (event) perception, causal inference between variables, and graphical models. This tuple is mostly a heuristic that has proven useful to highlight the diversity of problems, and should probably not be taken too literally: for instance, there is much more to Pearl’s work on causality than graphical models; the same problem (e.g., functional connectivity in Neuroscience via neural signals causality) can be tackled by Granger-like approaches, but also Bayesian models like DCM, even if they exhibit differences from other graphical models; and finally, the tuple only partially captures the role of counterfactuals in causal reasoning.??

Causal Perception

Humans exhibit a propensity to attribute causality to co-occurring events in the physical world (‘post hoc, ergo propter hoc’), and it was the seminal work of Belgian psychologist Albert Michotte, who established the physical and temporal conditions under which causal perception takes place (and, in Galavotti terms, this is clearly event causality, through single event pairs, rather than event repetitions suggesting regularity). While Michotte used simple animation sequences featuring geometrical shapes, we took advantage of more modern technology to stage experiments in 3D (with the help of Marc Buehner, a cognitive psychologist expert in causal perception), also using heuristic search techniques to explore a wider range of more or less realistic ‘effects’ to be associated to priming events based on their semantic properties.?

I-DID (Interactive Dynamic Influence Diagrams)

This work, led by my former colleague Yifeng Zeng, provides a causal framework for understanding the behaviour of multiple agents. Owing to DID association to Bayesian systems, this is the closest I got to what is often seen as the mainstream endeavour on causal inference (the ‘Pearl’ in the tuple). However, there is much more to Pearl’s work on causality than his Bayesian publications, and he has produced plenty of recent and enlightening papers on topics as diverse as clinical trials, or criticisms of contemporary Machine Learning.?

Causality in Formal Logic

It is generally considered that representing causality in standard (e.g. first order) logic is particularly challenging, which also limits its potential use in causal inference. On the other hand, Jean-Yves Girard’s Linear Logic has been described as a logical formalism encompassing both resource consumption and causality. Using again a narrative context (the formal representation of plot structures for Flaubert’s Madame Bovary), this work explores how causality can contribute to the formalisation of narrative events, within an overall (linear) logical framework.

Causality and Narrative Understanding

Studying causal understanding in children is an important topic in development psychology, and narratives provide an ideal experimental context to explore the issue. However, staging experiments can be challenging if we want to precisely control parameters (or cues) supporting causal inference. With the help of Paul van den Broek, one of the major experts in causal understanding, we have redirected our work in plan-based interactive narrative to develop bespoke virtual narratives in which the deployment of causal cues can be controlled to support experimental design.?

Causal Analysis of Neural Signals

Fairly traditional neuroscience methodology, establishing functional connectivity between brain regions in the search for neural mechanisms (in this case, subjects’ response to human or robotic conversational partners, a line of research at IMT in Marseille, France). Convergent Cross-Mapping (CCM) is a more recent method than Granger causality, which is based on state space reconstruction, and applicable to nonlinear signals, slowly gaining popularity in Neuroscience.??

In addition, neural signals causal and connectivity analysis is a topic from which we have seen much recent interest in applying Deep Learning methods, with a surprising number of papers appearing recently at AAAI or NeurIPS. Several techniques have been explored including GNN, GAN and Echo State Networks (ESN). Interestingly, in the latter case a parallel has been drawn between ESN and Granger [Duggento et al., 2021], while in principle ESN’s relation to Taken’s theorem would make them closer to CCM. When researching previous work for this research, I came across some quite original, and even surprising applications of these methods, originally developed for time series, to textual analysis (see for instance https://aclanthology.org/R15-1091.pdf).?

In hindsight, it is difficult to say whether I have been subconsciously drawn to multiple facets of causality through my own work, which has varied through the years with my own evolving interests, research collaborations and involvement of various PhD students (Jean-Luc Lugrin, Ross Conroy, Cameron Smith, and Caio de Castro Martins for the above topics/papers), or whether this simply reflects the ubiquity of causality in agent perception and reasoning.?I hope that, short of having provided any overarching framework, I managed to illustrate the ubiquity and diversity of causality, in a way that is even loosely connected to current AI debates, whether these focus on perception, reasoning, or explanation.?