Analogy

I don't have a favorite analogy, but looking at other analogies made me realize why a useful analogy is really hard to develop. Most of the analogies on "best of" lists are humorous. I think this is because humor is universal (except when it falls flat), so the insights generated by humorous analogies can be easily understood by anyone. Have a girlfriend and a car (or a boyfriend and a dog)... well you'll understand many of the analogies on this list. For these simple analogies, how much would I need to specify to a computer to get a pleasing result? Here are 23 metaphors of boyfriend is like a dog (http://elitedaily.com/dating/always-wants-bone-23-ways-boyfriend-exactly-like-puppy/939976/), each of them illustrate a different aspect of being in a relationship. How would a computer decide which one is appropriate? It all depends on the goals of the analogizer.

As analogy becomes more complex, with more nodes and edges being activated, this judgement becomes less important because there is less ambiguity in the relationships. Agreeing with Cole, I think computers will be able to create highly practical and accurate analogies eventually. These will be great as cognitive aids. But to be clear, these highly specified analogies are not the norm nor the kinds of analogies that *currently* give humans pleasure or insight. To have pleasure and insight through analogy, the analogizer must choose a simpler graph structure that contains more ambiguity, but then to set up the analogy in a way that the inevitability of the analogy seems obvious.

I'd like to end quoting Nabokov on his process of constructing a particularly difficult chess problem. I think it's pretty clear that chess problems are for him an analogy about how he leads his readers through metaphor and analogy in his own works:
"I remember one particular problem I had been trying to compose for months . . . It was meant for the delectation of the very expert solver. The unsophisticated might miss the point of the problem entirely, and discover its fairly simple, 'thetic' solution without having passed through the pleasurable torments prepared for the sophisticated one. The latter would start by falling for an illusory pattern of play based on a fashionable avant-garde theme . . . which the composer had taken the greatest pains to 'plant' . . . . Having passed through this 'antithetic' inferno the by now ultrasophisticated solver would reach the simple key move . . . as somebody on a wild goose chase might go from Albany to New York by way of Vancouver, Eurasia and the Azores. The pleasant experience of the roundabout route (strange landscapes, gongs, tigers, exotic customs, the thrice-repeated circuit of a newly married couple around the sacred fire of an earthen brazier) would amply reward him for the misery of the deceit, and after that, his arrival at the simple key would provide him with a synthesis of poignant artistic delight. (Speak, Memory 291-2)" (originally quoted in www99.libraries.psu.edu/nabokov/walter.htm)

Re: Do you think computers can successfully create analogies?

Thinking of Microsoft’s Tay and Cole’s discussion on analogy examples in online forums, I certainly think computers can successfully create “good” “appropriate” analogies --or at least, this is reasonably feasible in the reasonably short future. With Tay, that AI was learning to talk like a teenage girl, but by learning from pretty much anyone on Twitter. An AI agent with perfect “morality” is perhaps an unachievable goal because morality is a fuzzy concept, but by learning narrower biases, good or bad or appropriate or inappropriate analogies seem all possible. Tay made some great appropriate analogies appreciable only if you are a terrible person, perhaps. Holyoak discusses that analogies depend on the goals of the analoger and the biases of the audience. An agent may perhaps learn what key features of concepts can be appropriately made into analogy by many examples and learning who/what sources to listen and bias to. (This is a very difficult task, if an agent never encounters the word “nazi” they may not know how to judge it, but lots of heuristics are probably possible).

I think computers may struggle for a long time, however, on creating novel analogies such as “how is X like a house?” when X shares only metaphorical imagined features with a house and the answer isn't online.

I think the happening of "relational shift" in children is because of their understanding and experiences of the world. At first, when they see an object, the things they see are the features, like shape, color, material, etc. When they have more experiences, not only more things to draw from, but also the understanding of the same object, they can make inferences in more dimensions. This is quite similar to computer vision and deep learning. The first several layers kind of represent blobs, edges, shapes. And when getting higher and higher, the model can gradually learn concepts like cat or dog. Therefore, for the components of analogical reasoning, relational generalization is the hardest for computers.

Also, I agree with the above that computers can make decent analogies, if they have the data to draw from. However, understanding what analogies fit what context and audience is relatively hard, since computers have not been able to understand context and human effectively.

As for how analogical thinking can be applied to HCI, I think a successful example is the use of metaphor in interface design. Using existing metaphors in the real world can reduce our cognitive load at understanding something new. We can also design interactions in the interface, beyond similarity of objects, that can help us make high-level analogies when transferring the knowledge to something new, such as data model, interaction flow through the use of visualizations and workflows.

Re: What are different areas/ways in which our understanding of how analogical thinking works can be leveraged/applied to advance HCI?

In HCI, specifically in Learning Sciences, I think analogies can be a powerful tool to help understanding and learning of difficult topics. I am not familiar with the literature in this field, but many analogies from physics are coming to my mind: the horizontally stretched elastic fabric that bends when a heavy object is put on it (to represent the bend in space-time by heavy bodies such as the sun), or how a watermelon can be an atom and the seeds are the electrons wondering around. On the other hand, analogies seem hard to find for certain domains. For example, from various interactions with teachers and researchers who work in the algebra domain, it seems that there is "no good" analogy to linear equations. "The balance" principle is the most used one, which cannot be used with the "negative terms" and does not distinguish between the variable and constant concepts.

Re: (Just for fun) What is your favorite analogy?

I was taking a storytelling course in college, and I went to meet my professor because I was having trouble with one of the essay assignments. Specifically, I was not sure how to even start the essay, but after that I kind of knew what I wanted to say. My professor (great one by the way), after listening to my concern, called in his office a random person from the hallway. He turned to me and said "This is XXX". I immediately stood up to shake hands with this random person and introduce myself " HI XXXX, I am Franceska". That's when I realized that what my professor wanted to convey to me was "Introduce your topic", for my initial concern of how to start the essay. I still remember this as one of the best (and most spontaneous!) analogies I have ever heard or experienced.

Responding to whether computers can create analogies, I believe that there has been significant progress in this area depending on how you are qualifying whether an computer has been "successful". Finding mappings between words or semantic concepts or even images has been accomplished by a number of computational models of metaphor and similarity. I've encountered some language-based models that utilize large repositories of knowledge (eg: wikipedia or the internet as a whole) in order to look for conceptual links and to be able to draw some form of analogy from these databases (not exclusively by repeating ingested data). There is a robotics speaker presenting this week on context-free grammers for manipulation. This is an extension(or analogy if you will) of the analogy creation paradigm where any given manipulation motion can be drawn known submotions as is relevant based on how the context of the problem maps to knowledge of applications of submotions.

Another interesting problem with the computational aspect of analogy production is that to a degree we are treating machines as adult humans and thus many times they are expected to reproduce a singular human ability (analogical reasoning) with adult proficiency at runtime. To some degree one might consider more developmental models that would be able to perform some graded version of analogical reasoning depending on acquired experience, and as knowledge increased, analogical reasoning ability may increase similarly. These models may tend to demonstrate similar developmental patterns as children such as the "relational shift" as a consequence of knowledge representational change over time.

On the computational analogy-generating question, there has been some work at Georgia Tech (among other places) on developing an AI that uses conceptual blending to engage in co-creative pretend play with a child [1]. So, it certainly seems possible that more complex analogies could be created by an AI (although the existing system mentioned only does entity and predicate substitution, and not a more substantial restructuring), though the source and target domains would likely need to be well-structured. Even for humans, though, unprompted spontaneous retrieval is an issue, as seen in the radiation example. Crucially for a learning science approach that leverages this work, for an expert, the abstract schema is more readily accessible when confronted with novel problems that share its structure. So, it would seem essential that any ITS hoping to support expert analogical reasoning would help students develop abstract schema and move away from the more easily perceptual concrete (surface) features.

One area of this I thought was under-addressed was the use of analogy as a rhetorical device for persuasion as opposed to reasoning. It would seem like, for President Bush, for instance, creating an analogy between the Iraq War and World War II that placed him in the role of FDR or Churchill in opposition to Hussain-as-Hitler would be beneficial for his war-time goals, but would not be particularly useful for reasoning about the rationale to go to war. Although analogies can certainly be generative, they might also lead the listener towards a seemingly "inevitable" conclusion when used for malicious purposes.





[1] www.aaai.org/ocs/index.php/AIIDE/AIIDE13/paper/viewFile/7433/7650