Understanding understanding – could an A.I. cook meth?

What would it take to say that an artificial system “understands” something? What do we mean when we say humans understand something? I asked those questions on Twitter recently and it prompted some very interesting debate, which I will try to summarise and expand on here.

Several people complained that the questions were unanswerable until I had defined “understanding”, but that was exactly the problem – I didn’t have a good understanding of what understanding means. That’s what I was trying to unpick.

I know, of course, that there is a rich philosophical literature on this question, but the bits of it I’ve read were not quite getting at what I was after. I was trying to get to a cognitive or computational framework defining the parameters that constitute understanding in a human, such that we could operationalise it to the point that we could implement it in an artificial intelligence.

So, rather than starting with a definition, let me start with an illustration and see if we can use it to tease out the parameters that characterise understanding, especially the difference between understanding something and just knowing something or being able to do something.

I know it when I see it

Here goes (and if you haven’t watched Breaking Bad, I can only apologise for my choice of example):

Jesse Pinkman knows how to cook crystal meth. But Walter White understands the process. Jesse can follow the steps of the protocol. He knows that he should do step 1, then step 2, then step 3 – he can carry out the algorithm. Walter knows why they do step 1, then step 2, then step 3. Whatever constitutes that understanding, it is easily recognisable to us because he invented the protocol, because he can teach it to someone else, and because he can modify it as needed to scale it up or if various ingredients become hard to obtain.

So, what’s the difference? What parameters differ between Jesse’s brain and Walt’s brain when it comes to their knowledge of cooking meth? Why does Walt’s state constitute understanding, while Jesse’s does not?

Jesse’s knowledge is specific, isolated, fragmentary. He understands that when you add substance A to substance B you produce substance C. Walt’s knowledge, by contrast is situated in a much wider and deeper context. He understands how substance A reacts with substance B to produce substance C. That’s because he knows the properties of these substances that drive their reactivity. So, he has a level of knowledge that is more fundamental and more general, to which he can relate more specific facts.

And, beyond that, Walt understands why those substances have those properties. He has even more fundamental levels of knowledge about types of substances – not just those particular ones – and about their physical structures at a deeper level that produce their chemical properties. That’s why he could come up with the protocol in the first place and that’s why he can change it as circumstances demand.

So, we have a hierarchy of understanding – that, how, why. At each level we are situating some facts in the context of a wider body of knowledge, but also, crucially, a deeper level of knowledge. But not one that is reductionist – that is bogged down in the details of a lower ontological level. Instead, it is one that sees the important principles at that lower level and how they determine the properties at the higher level.

The simple accumulation of knowledge – the addition of extra facts – does not confer deeper understanding. What is required is the abstraction of general principles from all that knowledge, so that new facts can be situated within a logical, coherent framework – one that crucially entails causal relationships.

Responses from the hive-mind

This ties in to a lot of the responses I got to my questions on Twitter. These mostly converged onto ideas of abstraction – of categories, principles, causal relationships – and abduction, or abductive reasoning – forming a hypothesis from a set of observations, or thinking about why some things were observed or some facts hold true. It’s basically guessing, but, here, in a way that is informed by wider knowledge and experience.  

But there were a lot of other parameters suggested, and I have organised those below in a kind of ascending order. They start with those needed for a very simple kind of understanding, which might be programmed into an artificial intelligence. (Indeed, A.I. is exceptionally good at some of them).

And they ascend to properties that might be thought of as necessary for a higher kind of understanding, which may ultimately depend on internal models and representations, consciousness, awareness, and our own histories of embodied experience. And, who knows, maybe those properties could be built into an artificial intelligence too and maybe they are exactly the type of properties that would be required for what some people might consider “real” understanding, as opposed to a convincing simulation of it.

The elements of understanding

1.     Categorisation – building a kind of internal database of types of things, with knowledge of the properties that define each type. These necessarily have a hierarchical nature. So, Rover is a golden retriever, which is a subtype of dog. Dogs are characterised by A, B, and C – these collective properties form the schema or concept of “dog”, while properties D, E, and F characterise the subtype of golden retrievers. And, of course, dogs are themselves subtypes of mammals, which are subtypes of animals, and there are schemas for those levels as well.
2.     Generalisation – categories allow you to generalise. When you see a Corgi for the first time you may still be able to recognise it as a type of dog and make some predictions about its behaviour, because you understand some general things about dogs.
3.     Abstraction – the activity that enables you to build categories. You have to be able to see which are the properties that are essential to define a category at each level, and which are incidental.
4.     Compression – a related way to look at abstraction. How much detailed information can you throw away when moving from one level to the next, while retaining the important properties? Being able to see the forest for the trees depends on extracting trends or patterns, while ignoring a lot of the details, much of which may be noise.
5.     Pattern recognition – again, this is related to abstraction and compression. The ability to detect statistical regularities in a set of observations, but hopefully without over-fitting noise.
6.     Abduction – drawing hypotheses about the state of a system. Now we’re starting to go beyond just the properties of objects, to consider the properties of the relations between objects.
7.     Causal reasoning – this is where we really get into the meat of it, in my opinion, as we begin to understand systems, not just things or types of things. If I understand a system, I should be aware of the causal relations between its elements and the causal dependencies of system behaviour on those relations. I should have an understanding of causality both within levels and between levels.
8.     Counterfactual reasoning – one way to demonstrate such understanding is to be able to consider counterfactuals and their likely consequences. If such and such were NOT the case, how would that affect the operation of the system? To perform such reasoning, you have to know which details are important for the causal dynamics of the system. 
9.     Prediction – this is an extension of counterfactual reasoning. If you really understand a system you should be able to predict what would happen if you changed something about it, or predict how it would behave in some new scenario.
10. Manipulation – if we can manipulate and control a system with predictable outcomes, many would say that demonstrates understanding. It certainly demands knowledge of the causal relations at the level of operation one is aiming to control, though it is possible to achieve without a deeper understanding of the underlying principles at play.  
11. Invention – this is an even deeper level of understanding. It requires not just knowing the causal relations in a given system but understanding the more abstract principles underlying those relations. Those then become elements that can be reconfigured into new arrangements to carry out novel functions or operations.
12. Analogy – again, this is a question of abstraction of deeper principles. First, seeing that some particular causal relation between two things is a relation between two types of things. Then seeing that that particular relation is itself a type of relation with more general properties. This allows you to go beyond knowledge or understanding of a particular system to understanding of systems in general.
13. Mathematical description – some would argue you haven’t really understood a system until you can express it in mathematical terms, the most abstract level of reasoning. I think that goes too far – most of our understanding is intuitive or logical, without involving formal mathematics. However, mathematical expression can reveal deep correspondences that might not otherwise be apparent, such as the correspondence between Shannon information and thermodynamic entropy, or between predictive inference in perception and Bayesian statistical reasoning. 
14. Awareness? – Do you have to know you understand something to understand it? I don’t see any reason why you would, but, again, some people might have this as a criterion.
15. Articulability? – Being able to explain something to someone else or teach them how to do something can certainly be a good test of understanding, but I don’t know if it is a necessary criterion – maybe just an additional way we recognise it. 
16. Embodied phenomenological experience? – Do you have to have lived experience of something to truly understand it? This feels very vague, but hints at the idea that one thing that may limit artificial intelligence is that its knowledge is not gained by active exploration in the world. Perhaps disembodied knowledge will never reach a human level of understanding that comes with being embodied agents in the world. Like I said, it’s vague but links to the question of how information comes to have meaning, and whether meaning is required for true understanding. 
17. Perspective? – does understanding necessarily entail some subjective perspective? This is really just the context of the life history of the organism or agent that is doing the understanding. Clearly it can be a barrier to shared understanding, when there are implicit assumptions of wider context, prior positions, or values that are not themselves shared.

I am sure there are other criteria or properties that could be considered, but I think those capture most of the responses I got on Twitter. The key thing, to my mind, seems to be embedding knowledge at one level in the context of knowledge of another level, in particular knowledge of the causal relations of a system.

Now, the question is whether that discussion gets us any closer to knowing what we’d need to build into an A.I. to make it capable of real understanding.

Artificial intelligence

Some people would argue that A.I. is already capable of some kind of understanding. For example, DeepMind’s incredibly impressive engine AlphaZero knows how to play chess. But does it understand it? It certainly seems to, if you measure it by its success, and also by the kinds of moves it makes. Indeed, some people argued that its style of play – the “beauty” of some of its moves and strategies – shows more understanding than any human has ever demonstrated.

On the other hand, its knowledge is all at one level. It doesn’t know chess is a game – it doesn’t know what a game is. It doesn’t know it’s a metaphor for war. It doesn’t relate its knowledge of chess to its knowledge of anything else, because it doesn’t know about anything else.

Maybe all it would need to develop this understanding is to be exposed to lots more information. It clearly has the computational power to recognise patterns in massive amounts of data and to make predictions from massive amounts of prior experience. But is that how humans develop understanding? Or do they do something different? Do they actively extract rules and principles with less data? Are they wired to make abstractions and draw general inferences? Is our neural architecture particularly attuned to causal structure?

Perhaps it is the hierarchical architecture of the cerebral cortex that fosters the development of understanding, with each level being able to abstract more general principles by integrating across multiple units at the level below. This gives more opportunity to see emergent causal relations, to draw more distant analogies, to derive deeper principles.

Indeed, the expansion of the human cortex is characterised not by the expansion of individual areas, but by the addition of more areas, particularly of association cortex – the bits that integrate information from lower levels. We don’t just have more raw brainpower – evolution has extended the hierarchy of our neural architecture to higher and higher levels.

Of course, A.I.’s like AlphaZero have a hierarchical structure, with multiple layers, but it is all devoted to one thing. Maybe if we hooked a bunch of them up in parallel, each doing different things (say, playing different games), and then added another layer on top, that layer could extract more general principles (like aspects of game theory in general).

I don’t know whether the preceding discussion is really any use in thinking more precisely about how to operationalize and implement understanding in artificial systems, but maybe some people from that field will chime in. (This excellent blog on the interactions between the A.I. field and neuroscience is relevant).

Understanding in neuroscience

The discussion above also resonates with another question I asked recently on Twitter. I had been browsing the articles in the latest issue of Neuron, which span all levels of neuroscience, from the molecular up to human cognition, and wondered what would it take for a single person to really understand all of them?

This is one of the flagship journals of our field, yet most neuroscientists, myself included, can only really understand a small slice of the papers in each issue. This reflects the history of the field, which is, in fact, a loose agglomeration of many traditionally distinct disciplines – molecular and cellular neuroscience, developmental neurobiology, genetics, animal behaviour, electrophysiology, pharmacology, neuroanatomy, circuit and systems neuroscience, cognitive science, computational neuroscience, neuroimaging, psychology, psychiatry, neurology…

The reason the discussion of understanding is particularly relevant to neuroscience is that this field is by its nature hierarchical. These different disciplinary approaches are not defined solely by their methods but by their objects and levels of analysis, from single cells, to microcircuits, to extended systems and brain regions, to the ultimate emergent level of mind and behaviour.

The key challenge for any neuroscientist is to see across these levels. To understand how the dynamics of molecules within a cell determine its electrophysiological properties, in ways that determine its role in information processing within a microcircuit, which is itself a subcomponent of a larger circuit, the activity of which has some mental correlates, and so on. This is a challenge that one would hope could be met by new generations of students who do not have the baggage of having been educated in one of the historical silos.

But it is very easy to get bogged down in the details at each level. What is important is to abstract the general principles at play, which are required to understand the emergent functions of the next level up.

In fact, I think that’s not just the approach that we need to take to understand the nervous system as scientists. I think it’s the approach that the nervous system itself takes. From neuron to neuron, from region to region, from level to level, information is abstracted from all the noisy details, and this information has meaning that is understood in the context of lots of other information.

Perhaps it’s anthropomorphic to say that each neuron is trying to understand the neurons talking to it or that each level is trying to understand the one below it. But, then again, perhaps not – perhaps there is a deep mathematical correspondence between understanding at the psychological level and even the simplest elements of information processing at the neural level.  

With thanks to all those who engaged on Twitter - you're the reason it's such a great platform for scientific and philosophical discussions!



Popular posts from this blog

Undetermined - a response to Robert Sapolsky. Part 1 - a tale of two neuroscientists

Grandma’s trauma – a critical appraisal of the evidence for transgenerational epigenetic inheritance in humans

Undetermined - a response to Robert Sapolsky. Part 2 - assessing the scientific evidence