Complex Declarative Learning

So I'm taking Applied Machine Learning with Carolyn Rosé, and I think my learning about this new domain has many of dimensions of monotonic knowledge acquisition as well as as some of the qualities of non-learning responses that resist non-monotonic acquisition. The size of my knowledge has increased: I know more procedures and terms. I can combine those procedures and terms into simple patterns that I can apply to a variety of problems. The complexity of my ML schema has become greater as I've learned these new concepts, and the granularity has increased as I've been forced to do practice problems. I don't think the connectedness of my knowledge nor the abstraction level has greatly increased, perhaps those are more advanced dimensions that require a minimum size and complexity of knowledge before they increase. I'm not sure if my vantage point has changed, though Carolyn I feel would like us to think about how a ML algorithm sees the world. But I don't think she has put it in those words!

Because my knowledge of the domain is neither highly connected nor existing at multiple levels of abstraction, I think it is easier for me to assimilate and evade contradictory knowledge. Because a piece of information will contain multiple connections, with my sparsely connected knowledge graph, I can find a way to place the knowledge without it forcing me to confront how that piece of information contradicts my mental model. In a classroom, where I'm receiving a large stream of information, I can also evade contradictory information by focusing on the knowledge that supports my mental model. For a beginner, that might be a good thing.

I think the question of how does this apply to our own learning is an illuminating one. As my short narrative points out, these dimensions are not just not independent, they may be hierarchical or causal. I'd like to read more about when in the learning process the different dimensions are activated.

I was interested that the author mentions multiple anecdotes of scientific progress and meta-research on scientific progress such as how a french chemist Lavoiser’s conceptions of chemistry changed on the path to developing a major theory.

I struggle a lot with nonmonotonic knowledge while doing research in an area where, on multiple levels of abstraction, there is very little consistency or agreement among sources. After months of studying my research topic, I still become confused by what to believe. Reading a research paper in my area, much of it is simple, known monotonic information that strengthens what I’ve learned before, but I struggle to recall of all the intricacies of disagreement between sources and I struggle keeping track of which sides of these disagreements I can agree with in trying to build a reasonably unified concept of my own research. I’m interested how this work on complex declarative learning fits with a meta understanding of research, and the act of forming schemas/themes out of knowlege with unknown, chaotic, or unstable relationships.

In response to:
1. Think of a domain of complex declarative knowledge in which you have recently experienced change in your knowledge representation network beyond the accretion of new concept nodes (change in size). Identify whether that change was monotonic or non-monotonic, describe the dimension of change you experienced, and explain whether your reading this chapter shed light on that process of representational change.

I'm currently taking Graduate Ai and at the very beginning I thought (and was told by others) this was a course about search+decision making mostly. The search part was true however the decision making part was not, this became optimization. Later on it became almost obvious that decision making is really an optimization process, however without having clear this new concept, it is hard to see that optimization techniques can be used to solve decision making problems using for example linear programming. Notice how in the first place I didn't really have "Ai" knowledge nonetheless I had a pre-setup Ai network for which initially had only slight connection to optimization, later optimization's role in the network became central. This could be seen as a monotonic change since the connections of an existing node increased, however it could also be non-monotonic since decision making became almost replaced by optimization.


2. Knowing what we know about learning and non-learning responses to contradictory information, how might we better motivate people to engage in non-monotonic change for knowledge representations for which they might have a personal, perhaps politically or religiously motivated investment (e.g. climate change or “anti-vaccination”)?

A way this could be done is through Cognitive Behavioral Therapy (CBT) which basically consists on correcting a person's belief model. This correction usually replaces negative thoughts by more positive ones. Straight from wikipedia "CBT helps individuals replace "maladaptive... coping skills, cognitions, emotions and behaviors with more adaptive ones", by challenging an individual's way of thinking and the way that they react to certain habits or behaviors ". The therapy process seems not as straightforward and involves many sessions (fast version of CBT takes at least 12 hours which is about 12 sessions). However, CBT was designed to threat serious mental disorders and may not work for changing political or religious views.

4. Why do we have all these barriers to nonmonotonic change? Why isn't it the case that we can just update our knowledge smoothly in the face of more (or better) information?

(Speculation) Because more and better information is a relative measure. If there is a clear measure of better (i.e proves scientifically vs not), maybe people who understand the value of "proven scientifically" can more easily update their knowledge. Or maybe, because we choose to believe something is right, and we get confused if that is not. Confusion can lead to frustration (in learning in general), so maybe it is a subconscious act to refuse confusion. I am thinking here in terms of learning subjects, but also in culture change.

5. Do you see examples of these kinds of knowledge changes in your own line of research? Are there instances where your research community has struggled with nonmonotonic change (or monotonic change for that matter)?

I am too new to my field to know much about this kind of thing, but if it counts, I can bring an example from Physics When Einstein came up with the general theory of relativity, a lot of Newtonian Physics fans (most of them actually) were against it. It took a while (cite: Einstein's biography) for the field to change and accept the new theory.

Question 3:
I didn't quite grasp the division between procedural and declarative knowledge. For example, I think that learning how to play a sport is procedural knowledge (I know 'how' to defend against this play). But what if I'm just learning to recognize different plays. Has that now moved into declarative knowledge? (I know 'that' is a zone defense) If that's a proper distinction, then I'm more interested in how procedural knowledge and declarative knowledge relate to each other, how learning one helps learn the other.

But perhaps it's different when we are talking about knowing how to physically do something instead of just mentally knowing?

It’s exciting to see Design’s “problem reframing” claim shows up in Cog mini by the name of “Non-monotonic change”. In the non-learning stage, people examine and calibrate the knowledge landscape; in learning, the assumptions are challenged and point of views shifts.

In response to Q4. Why do we have all these barriers to nonmonotonic change? Why isn't it the case that we can just update our knowledge smoothly in the face of more (or better) information?
I think partly it’s because not all new information instantly fits into the existing mental model/knowledge landscape. However, I also doubt fitting into it should be the goal learning/education pursues. Does nonmonotonic change necessarily lead to “learning”, or merely an updated point of view? Do nonmonotonic changes hurt the “fresh eye” of the learner (aka creativity, cross-domain analogy, etc)?

So after reading the other papers and finding that schemas, examples, and analogies help with ideation, this paper got me thinking about how we might want to use the properties of learning and knowledge to help with that process. If it is true that knowledge is represented as domain specific semantic networks and that connectedness can increase among that network, would trying to blend domains be more likely to create cross-domain schemas?

I remember reading a few years ago that when you studied flash cards for classes, you shouldn't do a session with just Math cards and a session with just Physics cards, etc. Having sessions across subjects (supposedly) would help you learn how to transfer knowledge from one domain to another since you were encoding information together in time. I wonder if that is a good idea for research as well. If I read my machine learning papers right after my cog mini papers, will I be more likely to have an insight about how machine learning architectures resembles some cognitive models? What about the other way around?

On the other hand, what happens when they conflict? If my computer vision algorithm behaves in a way that is dissimilar to a biological model of vision, that doesn't seem like a problem, but it sure as hell bugs me.