Symbolic Emergence
Thrilled today to have anticipated around two years prior a paper that DeepMind just released. This is validation that my approach to general intelligence. Here’s the said paper:
Quoted from the paper “Our definition of a symbol draws on the work of the philosopher Charles Sanders Peirce. Peirce outlined three categories of relation — icons, indices, and symbols — whose definitions illuminate the role of convention in establishing meaning.” The authors conclude that “The views we’ve presented here seem to run counter to the past and current AI zeitgeist surrounding symbols, which emphasizes rigid mechanisms and particular representational forms.”
I’ve been a big proponent of C.S. Peirce formulation of evolutionary logic. When one first encounters Peirce, one finds his formulation of signs where he defines icons, indices, and symbols. The DeepMind paper is inspired by their understanding of this slice of Peirce’s work. However, Peirce semiotics is actually much richer than the pedestrian understanding of semiotics.
It’s relieving to find confirmation that one’s approach in fact runs ‘counter to past and current AI’. This may explain why I’ve my approach has difficulty gaining traction in the mainstream AI community. This is also a blessing because to separate oneself from the crowd, one must not be running with the crowd.
The key points of this paper are what the authors describe as symbolic fluency: receptive, constructive, embedded, malleable, separable, meaningful, and graded. Allow me the liberty to explore these 7 characteristics of symbolic fluency so as to mine newer insights.
Let me first point out that the authors allude to an enactivist approach to cognition. An enactivist approach implies that an agent learns about this world by being embodied, embedded, extended and enactive with its environment.
- Embodied involving more than the brain, including a more general involvement of bodily structures and processes.
- Embedded functioning only in a related external environment.
- Enacted involving not only neural processes but also things an organism does.
- Extended into the organism’s environment.
24 years ago, Horst Hendriks-Jansen wrote a book on AI “Catching Ourselves in the Act” where he describes an enactivist approach of AI revolving around 3 core ideas: situated context, interactive emergence, and evolution. Hendriks-Jansen defines interactive emergence as “Patterns of activity whose high-level structure cannot be reduced to specific sequences of movements may emerge from the interactions between simple reflexes and the particular environment to which they are adapted.”
It’s very rare that new ideas originate from a complete vacuum. Peirce semiotics is over 100 years old and there have been many deep thinkers who have evolved their ideas in the past. Enactivism also has a long history, perhaps going back to the work of Chilean cognitive scientists Maturana and Varela. However, it is safe to argue that such literature is not to well studied by most AI practitioners.
David Marr described 3 layers of explanation for neuroscience. A very important one that is often overlooked is the ‘computational layer’. The computational layers explain why a brain does what it does. It is its intrinsic motivation. We can in fact plot the different AI approaches in a graph with a horizontal axis describing epistemology (i.e. how is knowledge arrived at) and a vertical axis describing ontology (i.e. models of the world). The two main areas of metaphysics involve ontology (i.e. what is reality) and epistemology (i.e. how do we know what we know)
The meta-physics underlying deep learning is in fact one that is enactivist in nature. However, the ontological nature of its model is mathematically inspired rather than biological. Said differently, one based on a previously invented formula rather than from observations of biology.
First I can’t discern if there’s a logical order in how these seven are presented. So allow me to re-order them in a manner that makes sense to me. That is from an Enactivist perspective.
Receptive — The ability to ‘appreciate’ existing conventions, and to receive new ones. Most Deep Learning approaches enable this. I favor another term for this, I would call this intuitive. In fact, it is counterintuitive to realize that our language ability is a consequence of System 1 (i.e. Intuition) thinking. I explain this further in my exploration of reasoning.
Graded — This is a very good insight that has not been explicitly expressed enough. An agent’s ability for symbolic thinking covers a wide spectrum. I argued once that being diagnosed for being infected with a virus isn’t a binary measures (i.e. positive versus negative). It is important to know how much viral load had been exposed to. In the same way, symbolic fluency has varying degrees of competence. It is not a single behavioral condition to that one has access like Chomsky’s Merge capability. Rather, it is constructively developed through education and experience. One’s ability for symbolic thinking is related to one’s reasoning ability.
Malleable — Hofstadter would characterize this as conceptual slippage. The Bongard problem illustrates the importance of conceptual slippage.
Separable — “A symbol user should be able to demonstrate partitioning of their understanding of a substrate and symbol”. I call this Symmetry Breaking.
Constructive — The ability to form new conventions by imposing new meaning on effectively arbitrary substrates. This implies a kind of compositionality that creates new meaning. It is indeed interesting to notice that GPT-3 in its ability to generate text seems to have an intuitive grasp of the compositionality of natural language sentences. However, it does not derive the construction of these sentences from an internal model of meaning.
Embedded — The ability to derive meaning from the context of which symbols are used. This relates to Wittgenstein’s exploration of language games.
Meaningful — I don’t think it is an easy task in defining how semiotic processes create meaning. Usually you find a circular definition, that is a semiotic process is a meaning-making process. The question for symbolic interpretation is how a process interprets a sequence of symbols and maps it into a higher dimensional space. Does the mapping convey the meaning of the words? Wittgenstein proposed that language interpretation required a picture created in one’s mind.
DeepMind’s paper ‘Symbolic Behaviour in Artificial Intelligence’ is certainly a good first step towards a framework of understanding of how to get to general intelligence.
I would like to end with an important quote from Alan Turing’s 1938 Ph.D. dissertation:
“We have gone to the opposite extreme and eliminated not intuition but ingenuity, and this in spite of the fact that our aim has been in much the same direction.”
What he describes as ingenuity is what we describe today as symbol manipulation. Computers didn’t exist in 1938, but Turing had the intuition that symbol manipulation could indeed be automated. However, he was not as certain as to how intuition could be automated.
Kudos to the DeepMind folks ( Adam Santoro, Andrew Lampinen, Kory Mathewson, Timothy Lillicrap, David Raposo ) for articulating a promising path towards general intelligence. It is good to know that I’m not the only ‘crazy’ person out there that recognizes this path. Hendricks-Jansen coined the neologism ‘interactive emergence’, I will follow in the same spirit and coin this as ‘symbolic emergence’.
