Four Decades, Four AI Papers
At the turn of a decade, it’s especially tempting to reflect on one’s life and career and the passage of time. For me, as we enter the 2020s, this led to the following unsettling thought. All being well, at some point in the next year or so, I will be on the author list of a newly published paper in artificial intelligence. My first paper was in 1987, so that means I will have published AI papers in five different decades. So, what have I accomplished? If I had to pick four representative AI papers from my own publications list, one for each of the past four decades, which would I choose, and why?
Why should this vanity project be of any interest to anyone else? Well, my achievements may be modest compared to some of my peers, but my publication record has one unusual feature; it spans several disciplines (computer science, philosophy, neuroscience, dynamical systems), and within my home discipline of artificial intelligence, it covers multiple paradigms. Moreover, because my own intellectual journey reflects the way the field has changed over the years, my four chosen papers serve as a framing device for some commentary on where we’ve come from, where we are, and where we might be going. I’ve only selected publications in artificial intelligence (rather than related disciplines), and I’ve stuck to those that made a technical contribution with substantial engineering and/or theoretical content (avoiding reviews or methodology papers).
1980s
Shanahan, M. (1989). Prediction Is Deduction but Explanation Is Abduction. Proc. IJCAI 1989, pp.1055–1060.
In the 1980s, symbolic AI was very much in the ascendant (although it was also the decade that saw a revival of connectionism, largely thanks to the publication of Jay McLelland and David Rumelhart’s Parallel Distributed Processing). Expert systems were the most prominent application domain, and the aim was to encode knowledge in a suitably precise representational formalism and devise algorithms for rule-based inference over that knowledge. In symbolic AI’s purest incarnation, championed by the likes of John McCarthy and Pat Hayes, the knowledge representation formalism and inference mechanism were derived from mathematical logic. Or at least, it was mathematical logic that conferred on them a formal semantics.
Logic-based knowledge representation was the area I worked in back then, and this 1989 paper was my first to garner any attention from the field. It concerns the problem of how to represent actions and their effects, and specifically how to formalise the default assumption that things don’t change unless acted upon (which is related to the notorious frame problem). This was thought to be a foundational problem, and all proposed solutions at the time had one flaw or another. The problematic scenarios I investigated involved reasoning backwards in time from effects to causes (explanation), and I showed that if this is done with abduction rather than deduction, then the problems that beset other approaches don’t arise.
1990s
Shanahan, M. (1996). Robotics and the Common Sense Informatic Situation. Proc. ECAI 1996, pp.684–688.
In the late 1980s and early 1990s, roboticist Rodney Brooks published a series of highly influential papers criticising symbolic artificial intelligence. His central complaint was that symbolic AI views intelligence as a disembodied phenomenon, whereas in nature, intelligence has evolved to help embodied creatures get by in complex, dynamic environments. So Brooks advocated a re-orientation of AI research towards robotics, emphasising the importance of sensorimotor coupling with the world and downplaying the relevance of representation and inference.
For me, Brooks’s critique struck a chord. I’d been working on knowledge representation for a few years, and I was becoming disillusioned by the lack of progress. The field seemed obsessed with formal problems and was increasingly disconnected from the ambitions of the founders. I wanted to build AGI — artificial general intelligence — although in those days it was just called AI. I thought Brooks was right about the fundamentally embodied nature of cognition. But I couldn’t get on board with his radical proposal to jettison representation. I agreed with the philosopher Stevan Harnad, whose paper on the symbol grounding problem appeared in 1990; the problem isn’t with representation per se, but with representations that are not connected to the world through perception and action.
This 1996 paper (which won the ECAI best paper award) showed that it was possible to build a (very primitive) robot “whose architecture is based on the idea of representing the world by sentences of formal logic and reasoning about it by manipulating those sentences”, an approach that was dubbed cognitive robotics. One of the paper’s main contributions was to formalise the idea of assimilating sensor data as a form of logical abduction, thereby conferring a degree of groundedness on the resulting representations. The idea that perception can be seen as abduction has echoes in contemporary approaches that view perception as inverse graphics, such as Geoff Hinton’s capsule nets.
Another theme of the paper that still has contemporary relevance is common sense, which remains one of the biggest obstacles to achieving AGI. To make the process of perception through abduction work, a set of background axioms was needed that formalised aspects of time, action, change, space, spatial occupancy, and shape. At that time, my thinking on this was very much influenced by Pat Hayes and his classic naive physics manifesto papers. Today’s reinforcement learning agents still lack the foundational understanding of the everyday world that Hayes described (and which Elizabeth Spelke calls core knowledge). Rectifying this is a preoccupation of my current work, although I long ago abandoned the view that the way to do this is with formal logic.
2000s
Shanahan, M. (2006). A Cognitive Architecture that Combines Internal Simulation with a Global Workspace. Consciousness and Cognition 15 (2), 433–449.
The early 2000s were the depths of the 2nd AI Winter, and I was losing faith in the paradigm of symbolic AI altogether. My own proofs of concept didn’t scale beyond toy systems, and almost nobody seemed to be working towards AGI anyway. It was as if the field had given up on the grand vision of its founders, with half its practitioners migrating to applications and the other half withdrawing into theory. I decided to go back to basics and study the one exemplar of human-level intelligence we have: the human brain.
Abandoning a research area in which you have spent over a decade building a reputation is not recommended as a career move. Relentlessly ploughing the same furrow is a much better strategy. Nevertheless, I gave up on symbolic AI, and immersed myself in the neuroscience literature. In parallel, I pursued a longstanding interest in the philosophical questions that naturally arise when you think about AI, questions about mind, language, and consciousness. This combination — neuroscience and philosophy — is how I became interested in Bernie Baars’s global workspace theory, one of the leading contenders for a scientific theory of consciousness.
One of the things that first drew me to global workspace theory was that its basis is a cognitive architecture with a long pedigree in AI. Originating with Oliver Selfridge’s 1950s pandemonium architecture, the basic blueprint cropped up again in the blackboard systems of the 1980s, and its descendents are found today in mixture of experts models, as well as in Daniel Kahneman’s opposition of system 1 and system 2 thinking. Global workspace architecture comprises a set of parallel, specialist processes that compete (and sometimes co-operate) for the privilege of broadcasting information back to the full cohort of specialists. According to global workspace theory, local processing carried out in the parallel processes is unconscious, while information that is broadcast is processed consciously.
No trace of symbolic AI is to be found in this 2006 paper, and I present as a full-blown connectionist. The paper introduces a brain-inspired cognitive architecture based on global workspace theory that incorporates a recurrent, internal sensorimotor loop to realise a form of look-ahead (which I rather fancifully dubbed “imagination”). The architecture controls a very simple simulated robot in a very simple 3D environment. The paper describes an experiment in which the system predicts the sensory input expected to follow from the robot’s actions, and uses that prediction to modulate action selection. The architecture was implemented as a collection of recurrent neural network modules, each of which realised a form of attractor dynamics reminiscent of a Hopfield net. There was little learning involved, and the neurons were quite different from those we find in contemporary deep learning architectures. Nevertheless, I remain attached to many of the ideas in this paper, and I suspect they may still have relevance today.
2010s
Garnelo, M., Arulkumaran, K. & Shanahan, M. (2016). Towards Deep Symbolic Reinforcement Learning. ArXiv preprint, arXiv:1609.05518.
The early 2010s saw a series of breakthroughs in the application of deep neural networks, partly thanks to the availability of large datasets and plentiful computing power in the shape of repurposed GPUs. The big moment came for me in late 2013 when my friends and colleagues at DeepMind dropped their first deep reinforcement learning paper on arXiv. Their paper described DQN, a system that learned to play a large suite of Atari games from scratch, from raw pixel input. By then, my amateur foray into neuroscience had become something more professional. Most of my research involved looking at the brain as a complex dynamical system. I was particularly interested in brain connectivity and how it influences dynamics. I was still teaching AI, and it was always at the back of my mind — around that time I encountered Nick Bostrom’s and Eliezer Yudkowsky’s existential risk arguments, which exercised me a fair bit — but I was no longer doing much research in the field.
Then along came DQN, and that got my attention. For the first time, it seemed, someone had built an AI system with a degree of real generality. For sure, the domain of Atari games is limited. But since DQN is a viable reinforcement learning system whose input is raw pixels, you can, in principle, confront it with more or less any challenge that might be presented to an animal, and it will give it a shot. As soon as the source code was released, we had it up and running in the lab at Imperial College. It was fascinating to watch the system learn to play space invaders, starting as a jittering incompetent, but ending up (after many hours) as a player with superhuman skill. However, the excruciatingly slow pace of learning got me thinking about the differences between the way deep reinforcement learning works and the way a human would learn the same game. It also got me thinking about symbolic AI again, a paradigm I believed I’d left behind forever.
I already remarked that abandoning a research area in which you have built a reputation is not good for your career. Well, doing so twice is surely madness. But that’s what I did. My work on neurodynamics and brain connectivity was starting to get noticed, and I was publishing in top neuroscience journals. But I couldn’t resist the pull of artificial intelligence, now that the field was getting interesting again. This 2016 paper, co-authored with Marta Garnelo and Kai Arulkumaran, was the outcome. The paper has a critical portion and a constructive portion. In the critical part, we highlighted the shortcomings of contemporary (2016 vintage) deep reinforcement learning systems: they required large amounts of training data; they learned policies that didn’t transfer well, even to similar tasks; and the trained networks at their core were hard for humans to interpret. We pointed out that classical symbolic AI, despite its own weaknesses, has strengths that complement those of deep learning: its representations are abstract, compositional, interpretable, and lend themselves well to high-level reasoning. We also discussed the need for common sense, and the importance of causal inference, issues that few people in deep learning were addressing at that point.
But we didn’t want to stop at critique. We wanted to build something that would back up our argument. In the constructive part of the paper, we proposed a hybrid neuro-symbolic reinforcement learning architecture, in which a deep learning component extracts a representation of raw pixel input in terms of objects and relations, while a symbolic component learns high-level rules that encapsulate the agent’s policy. We then built a prototype agent based on our architecture — the engineering credit here goes to Marta Garnelo — and tested it on a simple foraging game, using DQN as our baseline. Even though it was a pretty rough first attempt at such an agent, we discovered several scenarios in which it dramatically out-performed DQN, thanks to the level of abstraction in the rules it learned, and its consequent ability to generalise.
Today there is a great deal of discussion in the AI community about the relative merits of the deep learning and symbolic paradigms, and the need to adopt the best of both is widely recognised. But very few people were voicing such opinions even as recently as 2016. There was a fine paper from Brendan Lake et al. (that scooped ours by a few months), and — as he often points out — Gary Marcus has been making related arguments for many years. A small community of enthusiasts had been pursuing neuro-symbolic hybrids for years, but they hadn’t yet addressed deep reinforcement learning, and the deep learning mainstream was paying them little attention. So our paper was timely. Shortly after we put it on arXiv, DeepMind offered jobs to myself and Marta Garnelo, where today we pursue a similar approach to AGI, albeit with more emphasis on differentiable components and end-to-end learning.
Well, there you have it: four decades, four AI papers. If I attempt the same exercise in 2030, I will be approaching retirement age. (Of course, many professors never really retire; I suspect I will be one of them.) I wonder how things will look then. Perhaps we will be in another AI winter. But I don’t think so. Perhaps we will still be arguing about the right way to overcome the conceptual obstacles that have dogged the field since its inception. That is quite likely. But perhaps a rudimentary form of artificial general intelligence will have been achieved (hopefully with safety and benefit to humanity in mind), in which case it will be fascinating to look back and see which ideas worked out.
