Defining our Terms

This post, and all posts published on this site, will be constructed using words. I imagine this seems reasonable enough (though there may be hope for some pictures as well!) – words are the foundation of communication across our world. But how well do we actually understand the nuance of this tool? In this post, we’ll take a deeper look at what actually happens when we use words, with a goal of revealing how words are much more abstract, inexact, and oftentimes blunt tools than we generally perceive them to be. We’ll start with a brief history of communication (covering many of the same ideas as in this post on intelligence), leading up to the use of language. We’ll then take a look at the essence of a word – what allows it to mean something to you; and more importantly, what allows it to mean the same (or just very similar?) things to different people. Afterwards, we’ll explore the implications of these findings for artificial intelligence, and put some thought into what is really required for a computer to understand language. With any luck, the words found below will manage to twist around and give you deeper understanding and appreciation of themselves.

The concept of communication varies quite drastically across different hierarchical levels, leading up to the highest level we know of (and the most common concept) – language. While it is relatively easy to define the types of communication commonly engaged in by humans, a more holistic definition is a bit harder to capture, making it important to cover a brief history of how communication as we know it arose to ensure we’re all starting off on a similar conceptual footing. The list below offers one way of looking at “levels” of communication, and is ordered by both complexity and evolution. There are some pretty significant step changes between these levels (theoretically could lay out a list numbered in the 10’s or 100’s or even 1000’s, but likely a much less interesting read!), but I hope the general progression comes through clearly – for example, while the first two items (especially the first) may seem out of place, they’re important to consider as part of laying the foundation for the latter conceptualizations, which are more aligned with our normal way of thinking about communication.

1. Communication of matter (e.g. gravity, charge, etc.)
2. Communication of simple organisms with no brain (e.g. E. Coli)
3. Communication of organisms with a small brain (e.g. C. Elegans)
4. Communication of organisms with larger brains (e.g. from ants to apes)
5. Communication using symbolic representation (e.g. human language)

1. Taking a very basic hierarchical view of the universe, all we see are atoms (for purposes of this post, will focus on atoms – but you could take a step down and look at subatomic particles, or a step up and look at molecules) interacting in different ways according to certain laws. These atoms and the actions they take have no meaning or intentionality, but they do take actions, and do influence the actions of other atoms. These interactions can be viewed as a form of communication – when an Na atom loses an electron and becomes Na+, and then bumps into a Cl- atom and attracts it, there’s a form of communication going on – the Na+ atom is communicating to the Cl- atom that the Cl- atom is to follow its trajectory, and vice versa. The medium of communication here is the electromagnetic force, and both atoms “know” how to send their “message” and “hear” the “messages” from other atoms (i.e. the atoms follow the laws of physics). Further explanation of how the atoms “know” or “hear” is outside the scope of this post (and might just get you a Nobel prize!), but the key point is that the atoms can be viewed as engaging in a form of communication, albeit meaningless (i.e. devoid of intention).
2. At the next step in the hierarchy, we find simple organisms communicating with each other using chemical signals. Looking at E. Coli, we see the bacteria engaging in quorum sensing – essentially, all bacteria release a certain chemical, and so the concentration of this chemical allows each bacteria to “know” when it is in a high density area, and adjust its behavior accordingly. It’s not too difficult to imagine the evolution of this behavior (assuming you’ve successfully imagined the origin of life, which I’ll leave for another post – or read Terrence Deacon’s “Incomplete Nature”) – the organism likely was already releasing a number of chemicals as by-products, and one got “picked up” to drive this beneficial reaction. While each bacteria is astronomically more complex than a single atom, the pattern of communication seems to parallel that laid out in level 1 – the bacteria play the role of atoms, and the laws modulating quorum sensing play the role of the laws of physics. The communication is still devoid of meaning to the bacteria, but is getting significantly closer to our hierarchical level.
3. Next, we see organisms communicating with each other in a slightly more flexible way, due to the abstractions made possible with a brain (as mentioned in the opening paragraph, this post provides some helpful context on this idea). While each organism’s brain is capable of forming only a very rudimentary model of the world, this model contains very simple representations of the other members of its species, allowing the organisms behavior to be influenced in certain ways by its peers. For instance, C. Elegans can exhibit swarming behaviors, which occur under certain circumstances when each worm “realizes” that its fellow worms are moving towards a certain area and follows suit. This communication is still mostly meaningless, as the worms sending the signal don’t have any realization that they’re doing so (they’re just moving according to their own drives) – but we do see the first steps towards emergence of meaning for the worms receiving the signal, as this information is incorporated in their very rudimentary model of the world (note that we can still view these phenomena in an analogous way to level 1, i.e. with the worms as the atoms and the laws governing their environment and neuron activations and synapse strength as the laws of physics, but the far more natural view is to attribute agency to the worms. This attribution of agency is an important step – we couldn’t do that in the same way for E. Coli. as there was no abstract “model of the world” being formed [which requires a brain-like artifact that can establish isomorphisms]).
4. Here’s the big jump. As brains gain greater ability to model the world, the representations of other organisms become increasingly accurate, and in turn a larger variety of signals are able to be picked up (allowing for the more complex behavior of organisms like ants and bees). However, the represented organisms also become increasingly complex (as their brains are evolving in tandem), and the representation is always playing catch-up. To give an example, an ant’s brain is able to create a much better model of the world than a worm’s brain, and so an ant will create a more accurate representation of a worm than a worm can – but an ant’s representation of another ant won’t be quite as complete, due to the complexity of the other ant. As discussed in this post, as organisms formed increasingly accurate models of the world, these models began to include the organism itself, leading to the formation of a sense of “I”. In parallel, the representations of other organisms morphed into a type of “agent” representation – which attributes the same “I”-ness and control over actions to other organisms. This was the pivotal step for the next level of communication – where the communicator can be said to be sending the signal for the purpose of sharing information with the communicatee (i.e. the communicator’s model of the world incorporates the fact that the communicatee will respond to the signal in a particular way, and that response is the reason for sending the signal). This allows more particular situations to be addressed (such as “food here”, “predator there”, “shelter below”, etc.) as certain types of actions can be learned (whether through evolution, as in ants, or associative learning, as in mammals) to be interpreted as signals for these concepts. 

With that last step change, we established a concrete basis for meaningful communication, building up to organisms representing other organisms as agents and interpreting their actions as intentional signals regarding the state of the environment (with environment here including the signalers internal state). This level of communication is commensurate with the general definition found in dictionaries (“a process by which information is exchanged between individuals through a common system…” – Merriam Webster) – but the complexity of language still feels a far way away. What is so different about language? The primary difference between language and the types of communication observed in level 4 is that all the level 4 communication is about something directly present in the environment (and is therefore tied directly to immediate action), while language has the ability to refer to something only abstractly present in the environment (in the respective brains of the communicator and communicatee). One way of looking at it is that all organisms operating at level 4 are only able to understand that their signals serve as an input for the actions of the receivers, while level 5 requires understanding that your signals serve as an input for the thoughts of the receivers (e.g. the signals can / will create certain patterns in their minds). An ape only sends a signal because of the behavior it expects to elicit in the receiver (getting another ape to share food, or to groom it, or to run away, etc.) – while a human can send a signal because of the patterns they expect to elicit in the receiver’s brain (e.g. sharing knowledge about the world). This ability requires understanding concepts (e.g. that all trees share a certain “treeness”), understanding the fact that concepts can be shared (i.e. that other members of your species also see that trees share a certain “treeness”), and being able to associate particular concepts with particular signals (i.e. the class of objects with “treeness” is what is being referred to with the sound signal “tree”) – certainly not straightforward steps to take. The evolutionary process which allowed for this advance is not fully known, although there are a number of interesting hypotheses – for the purposes of this post, I’m going to skip over the theories and instead jump to a deeper examination of the outcome – looking at what words are, how they’re represented, and how they work. Perhaps adding more structure to this concept will get some thoughts bouncing around in your head about how they might have originated!

“Dog”. What happens when you read that word, or when you hear it spoken? The simple explanation is that it causes your concept of dog to be activated (more on that in a bit) – bringing to mind an image of four legs and hair and a wet nose and floppy ears and a bone and a game of fetch and a golden coat and the name “Rover”… at least, that’s what my concept consists of. Was your concept the same? How is it that we’re still able to communicate clearly without identical conceptions? Well, for a concept like dog, there’s enough overlap in our concepts that for most purposes there’s no ambiguity – in forming our concepts (over the course of our lives, but mostly while young) we’ve seen many of the same types of animals labeled as dogs doing many of the same types of things, and so we’ve been able to parse out a “dogness” concept that accurately represents what animals of a specific sort (dogs) look like and do. Our concept formation started the first time we asked a parent “what’s that?”, and got the answer back of “dog” – to start, whatever that first animal looked and acted like would have been the basis for our conception, but over time this conception got adjusted based on additional experiences of dogs. How exactly our concepts get encoded is an open question in neuroscience, but the general principle is that they’re encoded as an associative, hierarchical web (which gets updated as we learn). When we hear the sound “dog”, there’s an associative connection with the concept dog, which is built up from connections between other object-type concepts such as “legs” (which is built up from “straight” and “object” and…) and “hair” (which is built up from “thin” and “many” and…) and is also connected to more abstract concepts like “run” (which is built up from “legs” and “move” and…) and “friendly” (which is built up from “smile” and “happy” and…) and… There’s no image of a dog sitting inside your brain, just this network of connections which represents the pattern dog – and this network of connections is constructed over time based on experiences in the world. This web works the other way around, too – when you see a dog, the word “dog” pops into your head (along with many of its associations, based on the context of the experience) – and even when you see a more removed object such as a bone, it may also activate the concept dog, thereby activating the symbol “dog” (again, along with many of its associations). Whether or not this happens depends both on the context of the situation, and also on how exactly your concept of dog is wired up with the other concepts in your brain (specifically in this example with the concept of bone). Although people’s concepts of “dog” differ in this way, from a usage perspective everyone’s concept of dog is nearly identical – this is because we experience many examples of dogs in our life to refine our concept, and so the activation patterns of our brains are similar (however, still possible to run into issues, e.g. does your concept of dog include both animals below?). 

When a person uses the word “dog” in communication (or any word), they’re relying on the fact that the listener’s / reader’s conception of dog is similar enough to theirs to get the point across. For simpler sentences, such as “The man walked the dog”, the vast, vast majority of the population does share similar conceptions, and so the point is communicated effectively. However, when venturing into topics involving more abstract concepts such as “abstract”, “belief”, “concept”, “divine”, “emotion”, “free will”, “government”, etc. the deviations between the conceptions of both sides can grow significantly larger. The concept dog has numerous observable instantiations we can use to align our definitions, but words like these have none – we’re primarily reliant on other words to help us build up our concepts. A definition of “abstract” might be “existing in thought or as an idea but not having a physical or concrete existence” – but for this definition to help, we must have prior concepts of all the words in it – and it seems equally likely for differences to pop up when considering words like “existing” or “thought” or “idea” or “physical” or “concrete” (or the words in any of their definitions!). Luckily, we’re able to more effectively refine our internal concepts of these types of words by observing how they’re used by people we talk to, in books we read, or even in blog posts about words – but people have significantly more variation in these types of experiences, leading to far less congruent representations. This is why communicating can be so hard! The communicator needs to not just understand what they’d like to communicate in terms of their own internal concepts, but also in terms of the concepts of their communicatees – and the concepts of the communicatees can vary greatly! The best way to overcome this limitation is to make communication a two way street – and so as we discuss some of these more complex topics on this blog, I ask you to please point out when I’m not communicating effectively – after all, the primary goal is to get my brain’s thoughts over into your brain. 

On Artificial Intelligence

The above discussions on concept formation (and particularly the dog images) may have reminded you of some of the recent advances in artificial intelligence with regards to image classification – we’ve trained algorithms to recognize dogs in pictures even better than we can, so does that mean computers have a more robust concept of “dog”? A quick overview of the operations of these computers may be helpful before answering that question. The vast majority of image recognition progress in computers has come by running a particular type of algorithm, known as a neural network (more specifically for image recognition, a convolutional neural network). This network takes as an input an image, and based on the pixel values of an image, gives an output which states what type of image it is (e.g. an image with a dog or without a dog). The network is structured in a hierarchical manner – there are a number of layer 1 nodes (“neurons”) whose values are based on equations which take as input the pixel values (with each node’s value calculated based on multiple pixels), then there’s a second layer taking the first as an input (with each layer 2 node connected to multiple layer 1 node), and so forth. This website https://www.cs.ryerson.ca/~aharley/vis/conv/ provides a helpful visualization of a network designed to recognize handwritten digits – but it’s important to note that this network has already been trained! All networks start in an untrained state, where all the connection equations between neurons (and between neurons and the input pixels) are randomly initialized (more accurately, the weights between neurons are randomly initialized, but the specific equation remains constant and takes those weights as an input) – so at the start, these networks are not very good at identification (no better than chance). Training happens by letting the network run on a number of labeled examples, and keeping track of what can be thought of as an error function. This error function allows you, after a number of examples have been run, to identify in what direction you need to move the connection equations (weightings) between neurons to more accurately label the images. You can then update the weightings in this direction and run additional rounds of training, making the algorithm better and better at recognition. One interesting emergent property of this approach is that the different layers are trained to identify different hierarchical levels of the desired object – so taking a dog, for instance, layer 1 might pick up hair and skin textures, level two might pick up a nose, legs, and ears, level three might pick up a dog’s face and body, and level four might identify the entire dog (these designations are vastly simplified, but similar behavior is exhibited in real neural networks). So back to our original question – does this type of network have a robust concept of a dog? It certainly has some type of concept – as mentioned before, a concept is really a recognition of certain repeated patterns, and here the computer is able to recognize the repeated pattern of the image of a dog. Adding credence to this idea, research has shown that the human visual system actually has much in common with the hierarchical structure of neural networks (also structured in a layered manner, with more complex features identified at the higher levels). However, there seems to be something shallow about the computer’s concept. It can recognize the patterns of pixels constituting a dog, but it can’t do anything with that concept – it doesn’t have any knowledge about the concept itself. This knowledge comes from the web of concepts looked at earlier – and understanding how that web of concepts is encoded will be critical for furthering artificial intelligence.

Author’s Note: If you enjoyed this post, please consider subscribing to be notified of new posts 🙂

Name*

Email*