Unpacking Understanding

• Understanding: unpacking another suitcase By Andrés Monroy-Hernández Society of Mind, Spring 2006 Professor Marvin Minsky

Introduction In this paper I try to unpack Understanding using the framework and style for the unpacking of Consciousness presented in Chapter IV of the Emotion Machine [1]. The relevance of this analysis for Artificial Intelligence is exemplified by the fierce debate over John Searle’s Chinese Room Argument [2]. Similarly, this topic is also the focus of study of educators, particularly those interested in a pedagogy centered on teaching for understanding, as eloquently stated by David Perkins. My goal is to bring together different perspectives on the subject to better dissect it or, as Minsky would put it, to unpack the suitcase of Understanding.

The nature of Understanding “To understand is to perceive patterns.” — Plato “Understanding is joyous.” — Carl Sagan “Peace cannot be achieved through violence, it can only be attained through understanding.” — Albert Einstein “Understanding is the reward of faith.” — Saint Augustine While not as mysterious as consciousness, understanding is yet another everyday word that Minsky would identify as a “suitcase word”. It used for “many types of processes, and for different kinds of purposes”, we apply it to “feelings, emotions and thoughts” [1]. Definitions are not useful either. For example, we find short definitions like “mental grasp” [3] being too broad, while others too specific that do not fully

describe all its uses: “a psychological process related to an abstract or physical object […] whereby one is able to think about it and use concepts to deal adequately with that object” [4]

Opening the suitcase Similar to the analysis of consciousness, one can wonder what kinds of creatures possess “understanding”. Recent studies in animal behavior suggest that some animals have more advanced behavior that previously thought. For example, a recent study [5, 6] suggested that wild bottlenose dolphins “know each other’s names”, that is, they seem to share the human ability to develop “an individually distinctive signature […] which appears to be used in individual recognition”. Analogously, if we were to find a non-human being with “higher level” cognitive skills than us, the concept of “true” understanding could be challenged. Minsky: “A Martian, or some alien machine that had a trillion of parts might consider us to be simple machines without any genuine understanding-because of being a trillion times simpler than them. They might just regard us as very cute toys” [7] Also, like with Consciousness, one could challenge the idea that understanding is an “all-or-none” trait. Educators have encountered detrimental effects in learning when using this way of thinking about understanding. For example, when the approach is that people either “get Math or they don’t” the results are a lot poorer than with an incremental approach where a learner is thought to be able to perform their understanding in variety of different in ways. This is in line with some of the AI work that identifies understanding as the ability to represent something in multiple ways.

How do we recognize Understanding? In the same way that “consciousness” gets detected by a “conscious recognizer” resource [1], I hypothesize that our minds also have an “understanding recognizer” resource. However, people experience that often times the mind incorrectly detects understanding, something that does not happen with

consciousness. For the most part, people are always able to tell with certainty if they are “conscious”. For example, let’s look at this imaginary conversation: Charles: “Joan, are you conscious now?” Joan: “Yes, I am.” Charles: “A few seconds ago you blinked. Did you do that consciously?” Joan: “No.” An hour later…. Charles: “Joan, an hour ago you told me you were conscious. Was that assessment correct?” Joan: “Yes, of course!” Charles: “You also told me that you had blinked unconsciously. Was that assessment correct?” Joan: “Yes.” Typically, Joan’s ability to detect her state of mind as conscious or unconscious does not change over time. On the other hand, let’s look at this other imaginary conversation: Teacher: “This is how you multiply 15 x 14…” The teacher shows the student the process on the chalkboard starting by multiplying 4 x 5, and so on… Teacher: “Did you understand what I did?” Student: “Yes” (student honestly believes he does) Teacher: “Could you multiply 16 x 17?” Student tries but fails… Teacher: “You have said before that you understood. Was that assessment correct?” Student: “No, now I realized I had not really understood. I was just able to follow the steps you did and I recognize that I had that in my memory, but I cannot generalize it to other numbers.” This caricaturized version of a teacher-student interaction represents something we all have experienced at some point in our lives. Why does our “understanding recognizer” create false positives (and sometimes even false negatives) more

often than our “conscious recognizer” does? I think this can be explained once we have a model for “understanding”.

Four-leveled Model of Understanding "I hear and I forget. I see and I remember. I do and I understand.” – Confucius “Data is not information, Information is not knowledge, Knowledge is not understanding, Understanding is not wisdom.” – Cliff Stoll “The improvement of understanding is for two ends: first, our own increase of knowledge; secondly, to enable us to deliver that knowledge to others.” – John Locke “Whatever you cannot understand, you cannot possess”.— Johann Wolfgang von Goethe If we are to design a machine that does what our minds do, we must identify what are the elements that constitute human understanding. David Perkins and other educators, in their interest to develop better pedagogies, have identified different constituents of understanding. Perkins starts by distinguishing knowledge from understanding: “A comparison between knowing and understanding underscores the mysterious character of understanding. […] The mystery boils down to this: Knowing is a state of possession […] But understanding somehow goes beyond possession. […] we have to be clearer about that ‘beyond possession.’” [8] To go beyond possession one must show enablement, the capacity to do something with that knowledge, be it observable or not by others. In Chapter IV the Emotion Machine, when unpacking consciousness we are presented with the story of Joan almost hit by a car when crossing the street. Minsky then points out that instead of asking if Joan was conscious, we should look at what Joan did. Similarly, instead of dwelling on what understanding is we should look at what are the things we do when understanding something. Perkins

calls these things “understanding performances”. For example, for a statement like “Joan understands multiplication”, we could have the following performances: 1. Repetition. Joan repeats the multiplication tables and the memorized definition from the dictionary. 2. Paraphrasing. She uses different words to say the same definition. 3. Execution. She uses the multiplication algorithm to multiply any pair of numbers. 4. Explanation. She explains with her own words what multiplication is for her. 5. Exemplification. She gives examples of multiplication at work. For instance, she can say that in order to know how much she earns a year she can multiply her monthly salary by the number of months in a year. 6. Application. She uses multiplication to explain something not yet studied by her. For example, she could think that in order to calculate the combinations in which she can dress using her favorite t-shirts and pants she can use a multiplication. 7. Comparison. She notes the similarities between multiplication and addition and how division is the inverse of division. 8. Contextualization. She explores the relationship of multiplication with the broader world of Mathematics. For example, she can say that the area of a quadrilateral figure can be expressed in terms of the product of two contiguous sides. 9. Generalization. She constructs in her mind the concept of multiplicity as a model that represents concepts beyond Mathematics. These performances hint for an organized multilevel structure of understanding based on the type of performances shown by a human or a machine. Perkins identifies the following four:

1. Content. This level refers to knowledge stored in multiple representations. In computational terms this would refer to data, the relationship between data (i.e. RDBMS, semantic web, common-sense knowledge) and the procedures that can be performed on data (i.e. algorithms, computer programs). 2. Problem solving. This level is the ability to solve specific problems applying the knowledge stored in memory. This can include problem solving strategies from Minsky’s model six such as Reasoning by Analogy or Crying for Help. 3. Epistemic. This level is the ability to generate justifications and explanations. I believe this relies on Panalogies that can provide us with basis to justify or explain something from different mental realms. For example, Joan can explain the reasons why she crossed the street based on different realms. 4. Inquiry. This level is about self generated questions, hypothesis and challenges of other people’s ideas. I think the use of Panalogies also plays an important role in this level. For example, a child can come up with his own design for building a bridge with LEGO blocks based on ideas from the Physical realm (the two blocks at the bottom support the top one), Social realm (this design will give me attention from others), Emotional (I will be proud to have built the bridge), and so on.

Multirepresentation "Understanding means seeing that the same thing said different ways is the same thing." --- Ludwig Wittgenstein The ability to represent knowledge in different representations is the core of the first level and it is essential for the second, third and fourth levels. This allows us to retrieve the knowledge we have under multiple circumstances that perhaps were not the same as when it was stored: when solving problems (level 2), explaining the why’s of the world (level 3) or creating new hypothesis (level 4).

Without this multirepresentation, the knowledge stored would be too specific and not applicable to any other situation.

False positives As mentioned before, it seems that often times our “understanding recognizer” makes erroneous detections because sometimes it changes previous verdicts. For example, sometimes after reading a few pages from a book, we think we have understood them, but later we change our mind and realize we had not really understood. A few reasons for this can be: 1. Our “understanding recognizer” resource takes time to be developed. Perhaps the “understanding recognizer” only develops as much as it’s needed and no more. Therefore in childhood, when we are just starting to experience the world, we might experience false positives more often. 2. The “understanding recognizer” incorrectly decides which levels from the understanding model it will need to detect to accomplish a goal. This is the idea that the “understanding recognizer” looks for the activation of different levels depending on a predicted goal. For instance, for a student used to only memorize data and processes to accomplish his goals, his or her “understanding recognizer” might be triggered with just level one being active. For a different student, with previous experiences that require the four levels to accomplish goals, his or her “understanding recognizer” will only be triggered when the four levels are active. This might depend on the familiarity with the topic, previous experiences and age. This is why children and novices in a field might have similar experiences related to detecting false positives. Perhaps, one of the goals of education should be to teach which levels of understanding to turn on at any given point.

The Chinese Room Argument “The computational theory of the mind, strong AI, has two wonderful features. First of all you can state it simply: ‘the mind is a computer

program, the brain is computer hardware’ […] and furthermore, not only can you state it simply, you can refute it in a few minutes […] It takes five minutes to state the refutation.” – John Searle on the Philosophy of Mind [10] Some people might contend that the analysis of understanding would not be complete without mentioning one of the arguments that have inspired intense debate in the Artificial Intelligence community: the Chinese Room Argument. This argument is a Gedankenexperiment, or thought experiment, proposed by John Searle based on five well known ideas: 1. Turing Machines, the mathematical representation of a symbolmanipulating device. 2. Algorithm, the mathematical description of a program. 3. Church’s Thesis, the idea that any algorithm can be implemented in a Turing machine. 4. Turing’s Theorem, the idea that a Universal Turing Machine is capable of simulating any Turing Machine. 5. Turing Test, originally proposed by Alan Turing “in order to replace the emotionally charged (and for him) meaningless question ‘Can machines think?’ with a more well-defined one.” [10] With those ideas in mind, Searle asks us to think that “a monolingual English speaker who is locked in a room with a set of computer rules for answer questions in Chinese would in principle be able to pass the Turing Test, but would not thereby understand a word of Chinese. If the man doesn’t understand Chinese, neither does any digital computer.” [9] A first reaction to this experiment could be to note the weight Searle places on the Turing Test as a way to determine if a machine can understand Chinese. However, to his defense, later he states the formal structure of his argument as follows:

A. Programs are syntactical B. Minds have semantic contents C. Syntax is not sufficient for semantics D. Therefore, programs are not minds. There are a number of replies to the Chinese Room Argument, some of them, I believe, are unconvincing. For example, the “robot reply” proposes that if we replace the room with a robot that walks around and interacts with the environment, like a human does, it would be showing understanding for the things he does. Other responses, like the “systems response” are more convincing. The “systems reply” states that it is not the man inside the room that understands Chinese, but the system as a whole. It is important to note that Searle does not deny the possibility that machines can do what human minds do. What he does not believe is that a computer program or for that matter a Turing Machine, can achieve it. He says: "only a machine could think, and only very special kinds of machines, namely brains and machines with internal causal powers equivalent to those of brains And that is why strong AI has little to tell us about thinking, since it is not about machines but about programs, and no program by itself is sufficient for thinking.” [2] As stated before, one of the biggest issues with this type of discussion is the use of the word understanding. This leads to a lengthy discussion over a word that encapsulates multiple other processes without looking at each one of them separately. Unfortunately, he also puts emphasis on yet another suitcase word: intentionality. He says: “Any attempt literally to create intentionality artificially (strong AI) could not succeed just by designing programs but would have to duplicate the causal powers of the human brain.” [2] On the other hand, I agree with Searle’s on his following ideas:

1. Passing the Turing Test does not prove understanding. Perhaps even Turing himself would agree with it given how the test was originated. Also, because, as mentioned before, understanding is a collection of multiple cognitive processes and a single test cannot probe for all of them. 2. An individual program does lack semantics. 3. Only very special kinds of machines could do what our minds do. What I think is Searle’s main faults are: 1. The use of words and concepts such as “understanding”, “causal powers” and “intentionality”. 2. The focus of his argument on the power of a single Turing Machine or a computer program. He severely underestimates the power of a large, distributed and decentralized number of programs enriched with a large repository of knowledge. I believe that the model of understanding presented in this paper shows that the programs are just one of the many components of the processes that embody what we call understanding.

Conclusion The model of understanding presented in this paper gives importance not only to the programs, but to the knowledge and the relationships between elements in the four different levels. It is my hope that it also hints on the quantity of those elements needed in order to achieve each level. The Chinese Room Argument debate help us conclude that the question to ask is not whether a program can “understand or not” but rather: what type of processes would a “very special kind of machine” need to have in order to achieve what the human mind achieves? I think that the Emotion Machine addresses that question and I hope the model of understanding contributes by tackling one of those processes.

References [1] Minsky, M (in preparation). The Emotion Machine. [2] Searle, John. R. (1980) Minds, brains, and programs. Behavioral and Brain Sciences 3 (3): 417-457 [3] Merriam-Webster OnLine. http://www.m-w.com/ [4] Wikipedia. http://en.wikipedia.org [5] Leake, J (2006) Dolphins ‘know each other’s names’. The Sunday Times. May 6, 2006. http://www.timesonline.co.uk/article/0,,2087-2168604,00.html [6] Janik, V.M.; Sayigh, L. S.; Wells, R. S. Signature whistle shape conveys identity information to bottlenose dolphins. Proceedings of the National Academy of Sciences USA (in press) [7] Truth Journal. http://www.leaderu.com/truth/2truth03.html [8] Perkins, D. (1992) Smart Schools: Better Thinking and Learning for Every Child. Chapter 4: Towards a Pedagogy of Understanding. pp73-95. [9] Searle, John R (1998). The Philosophy of Mind (Course Guide and Audio). Lectures three and four. The Teaching Company. [10] Turing Test. Wikipedia http://en.wikipedia.org/Turing_test