The notion of "Causes and Clauses" is one of the most appealing theories in Minsky's Society of Mind. In this paper, we seek to reconcile this abstract idea with some of the more structured theories found in The Emotion Machine. In particular, we have asked ourselves, how can the notion of "Causes and Clauses" be made to fit with his notion of the six-level model of mind?
 Let us remind ourselves briefly what these two theories say. "Causes and Clauses", described prominently in the Society of Mind, but also in an essay entitled "Alien Intelligence", is a description of tendencies we use to represent the world. The theory has four parts: Things, Differences, Causes, and Clauses. Things are components of a scene, roughly corresponding to nouns. Differences are a comparison of different things, and can themselves be Things. Causes are the person, process, or things that we hold responsible for differences. Clauses are linguistic structures that can be built out of Things and Differences, and that can describe Causes. Clauses can also be Things. One of the features of this theory is the way in which all four parts can be considered special kinds of Things. The ability for Clauses to be considered Things allows Clauses to be parts of Clauses, and for chains of thought to be composed recursively.
 The six-level model of mind (model six) is detailed in the fifth chapter of The Emotion Machine. At the bottom of the model are instinctive reactions such as reflexes and reactions that we have at birth. Learned reactions are the next level up, and are made up of those reactions that we acquire over time. Deliberative thinking, on the next level, describes forward-looking kinds of thought that can test and act out hypothetical plans. Further up, reflective thinking allows us to look back on the results of deliberative thought, recognize patterns, and improve our future deliberations. Further up yet, self-reflective thinking adds a model of self to the actions of the world, and considers that self's actions from an outside perspective. Finally, self-conscious thinking incorporates the perceived opinions of others into evaluations about what the self should be doing.
 
[Intelligence as decomposition]
 
 Now that we have established what the two theories say, let us take a stab at integrating the two. Our strategy will be to first form a naive marriage between the two, examine where it falls short, and then attempt to combine them in a more profitable way in light of those shortcomings.
 
A Naive Combination
 
 The first combination we will examine is motivated by the view that the Causes and Clauses idea describes a hierarchy. Since model six is clearly a hierarchy, a first attempt at unifying the two might try to match the hierarchical levels together. Let us explain why this might seem like a good idea at first.
 The hierarchical structure of Causes and Clauses is not too difficult to see. Clauses are non-existent without being composed of nouns and verbs, to which Things and Differences relate. One could certainly describe a Cause in a Clause. It seems like Clause is in several ways the "highest-level" of this theory. Causes cannot be identified without recognizing Differences. Differences are only meaningful as a comparison between Things. Thus there appears to be a hierarchy of compositionality as one progresses from Things to Clauses.
 How can we map this hierarchy onto the hierarchy in model six? It appears as though one can simply associate the Causes and Clauses idea with the first four levels of model six. It is easy to think about instinctive reactions as involving sensory perception. Minsky describes instinctive reactions as stimulus-response activities. For the purposes of matching these hierarchies, let us therefore think about Things as sensory stimuli. The first level of understanding of the world that a baby perceives might be described as a world full of stimuli that it does not yet understand, but responds to in ways that are genetically pre-programmed. In this way, as we are trying to combine these two hierarchies from the bottom up as we might zip up a zipper, we notice that this combined hierarchy seems more like a hierarchy of child cognitive development than a hierarchy of compositionality. 
 As we continue to zip up these two hierarchies into one, we find ourselves comparing Differences with Learned Reactions. As the developing child begins to make sense of the world, she starts categorizing stimuli by their similarity or difference. Only by a process of dissociating stimuli and identifying the differences can any learning occur. Thus Differences appear to be crucial to this level of model six.
 Further upwards we compare Causes with deliberative thinking. Keeping the child development example in mind, it seems that only once a child has the ability to begin assigning agents of action to the differences she is observing can she begin to mentally experiment with deliberation. Because deliberation involves thinking about action in the future--action which is set in motion by an agent--the notion of Cause appears to be an important precondition. Children certainly cannot begin to identify causes for things before they are able to group stimuli based on their similarity or learn differences. It seems that Causes are appropriately matched with this level of model six.
 Finally, we compare reflective thinking with Clauses. The fit between these two levels is rather straightforward. Because a clause allows the composition of the other elements, including itself, it is a structure well suited to the process of reflection. The clause can be thought of as the Petri dish or test tube of thinking, a container for fragments of ideas which that allow them to be manipulated and examined as separate units. And since they Clauses are themselves things, they further lend themselves to reflection as they can be analyzed and composed in a recursive manner.
 

fit with M6

naïve combinations

prose 6x4 matrix

what’s missing from the matrix – explain why new columns are necessary

mammoth grid/prose – explanatory/generative cross-products

lessons learned

processing vs representation

new hierarchy

things + differences + top level

clause as a frame – then viewed as things at the level above

reward – goals, causes, consequences

adjectives
 

New matrix

	things	differences	clauses	reward	causes	consequences	goals	adjectives	example	processing
instinctive	basic low-level percepts should this be features too percept data	change in percepts	episodic memories of action-reaction pair		stimulus-response based on some feature of the total input vector				burning and recoiling from stove
learning	low-to medium- level perceptual data percept data – represented as a vector, ranging from the pixel to the geon level somewhere above the edge-detection level prototype features in your state-space where ‘state’ = low/mid level perceptual data	action comparisons, clustering statistical stuff, SOMs etc.	episodic memory state-action-state triple	reward for given state-action pair (introduce)	agent some part of the state that’s always present what’s common to all the episodic memories that elicit a given action principle components (not needed)	the action associated with a given stimulus??? no real sense of consequence the state following your action – i.e. a difference the last state a special type of thing that occurs after a state+action (not needed)	maximising reward (not needed)	adjective - distance from prototype vector to current state vector could also have adverbs too – distance from prototype to current action vector ignore		2 types of learning: unsupervised learning predictive feedforward model
	things	differences	clauses	reward	causes	consequences	goals	adjectives	example	processing
deliberative	transitions state or input situation feature vector chains	between transitions difference between input + output vector	plan – sequence of transitions plus learned consequence	whether you like the output vector that you’ve calculated	whichever features affect the output most for a given input vector (introduce)	difference between input vector and output vector (introduce)	maximising reward (not needed)	variations on your transitions, i.e. the differences between them that describe new test transitions (not needed)	“when I go to bake a cake today, its okay to touch the handle of the stove, but not the stove, because its hot”	sequences of transitions are formed by establishing test chains of transitions, evaluating those test chains on the basis of reward, discarding parts that don’t work and adding parts that do.
	things	differences	clauses	reward	causes	consequences	goals	adjectives	example	processing
reflective	sequences of transitions, i.e. ‘possibility-chains’	comparisons between them how your plans differ from each other	record of that process what it is about a plan that makes it better than another a plan (sequences of transitions) incorporating causes, consequences and goals? sequences of transitions + “expected consequence” + affective response + cause (reason why it worked)		reasons for coming up with those plans reason why a plan worked	what the plan did that you liked	how you judge the plans – based indirectly on reward – reward has taught you what kinds of states you like to be in, which you’ve clustered as goals (introduce)	different permutations on plans ways to describe the output of a “matching process” which tries to find similarities between sequences of transitions highlights a feature that’s different – useful in an intentional way (introduce)	“When I planned to only touch the handle of the stove, and then burned myself, I didn’t consider that it was too was hot, because it looked insulated.”	has fairly transparent access to the internal workings of the deliberative, so that it can draw inferences from the deliberative’s dead ends --notion of curiosity comes out of an affective response to plans and consequences (maybe just reward)
	things	differences	clauses	reward	causes	consequences	goals	adjectives	example	processing
self-reflective	different versions of self instantiation of self as thing … see reflective clause maybe not so much a model of self as a model of someone	comparisons between reflections comparison between comparisons between the possibility chains comparing itself with itself at other times	records of process		self-rationalisations			comparing me at the moment to the prototype description of myself – I was in a much bigger hurry, or more distracted etc. than usual	“How could I not have realized that the handle was going to burn me before touching it? I must have been distracted.”	process centers around a model of self. The model is a testbed of experimentation. Different “reflection clauses” or reflections are placed into the testbed. Now, however, the outer workings can also be examined. The combination between the outer workings and the inner workings can be used to improve future reflections. Phrases that involved referring to the “I” involve self-reflection. what determines what I infer from looking at the possibility-chains? i.e. how does the state of the system (i.e. me) affect its own processing and self-understanding
self-conscious	self-reflections	highlighted between self- reflections of different people imagining they are you - you imagining you are them, no???	record of process		Causes are the cheap emotional reasons people attribute to your actions				“If anyone else had seen me, they would have made fun of me, because they would have thought I was stupid.”	Your model of other people’s model of yourself.

 
 
Where this combination falls short
 
 This thought experiment, however, has its limits. As we will now show, there are many reasons to think that this naive combination is an abuse of these two concepts, and does not in fact clarify either one. We feel that several issues are important to examine. Firstly, we feel that the notion of a Clause is too undetermined. Secondly we feel that because of the recursive nature of the Causes and Clauses idea, that it does not neatly fit into a 4 part hierarchy, but in fact is a continuum which extends both above and below the 4 basic components. Thirdly, we feel that a Cause does not necessarily have to be a part of a Clause, thus breaking the hierarchical requirement. One might argue that a Cause and a Clause could fit alongside one another in a hierarchy. Fourthly, we feel that the bidirectional nature of model six is at odds with the unidirectional explanation of Causes and Clauses as a hierarchy.
 
[Better explanation of these concerns here]
 
Better Combinations
 
 What we believe is a more profitable way to think about these two ideas in combination is as a 6x4 grid, which allows all parts of the Causes and Clauses framework to exist within all levels of model six.
 
START HERE
 Let us imagine how the Causes and Clauses framework fits into each of the levels of model six. In each case, we will define Things, Differences, Causes, and Clauses when viewed through the lens of the particular level. Let us begin with the Instinctive level and take the example of a person who, by reflex, pulls their hand away after touching a hot stove. We can think of two different types ofa Things that apply to this level. Type 1 Things are as low-level perceptual data, roughly represented as a feature vector. and Type 2 Things are low-level motor responses.  A a Difference can be framed as the informational “distance” between perceptual data or motor responses that ariseing from different stimuli. In this case, the type 1 Things  isare the percepts of a regular hand and a burning hand, and the Difference is a measure of the way that those two percepts are dissimilar. The type 2 Things are the reflex motions that pull the hand away from the pain. Our notion of dissimilarity is inspired by the dot product between two vectors, which yields a numerical measure of similarity by quantifying how much the vectors point in the same direction. Our notion of dissimilarity can be thought of as the opposite of this measure.
 When we turn to discovering a place for a Cause or a Clause at this level, we turn up empty-handed.  It doesn’t make sense to have a Cause for an instinctive reaction because the association between stimulus and response  (type 1 and type 2 things) cannot be attributed to anything other than the genetically predetermined association between pain and reflex.  While we might say that there was a cause for the reflex when looking from outside the system, looking from inside the system there is little reason to think that a Cause representation exists at this level.  
 Clauses also seem misplaced on this level. The presence of Clauses suggests the ability to compose Things and Differences.  However, at such a low level, where it is not clear that learning occurs, it seems meaningless to talk about combinations of Things and Differences.
 The Cause corresponds to the agent responsible for the Difference. In our example, the Difference between the burning hand and the regular hand is that one is in pain (though this is not represented explicitly). Under the influence of reflexes that are genetically determined, the body pulls away from the pain in the direction opposite from its source. While there may not be a detailed concept of what that cause was, the fact that the hand moves correctly to avoid the pain suggests that the Cause is known in some form. 
 In this framework, the Clause can be thought of as a structure for composing Things and Differences, but not Causes. This is because we feel that upper levels of thinking are not aware of what reflexes the body has until a person develops a model of itself. Therefore, while it makes sense that Things and Differences, percepts and the differences between them, may be aggregated and categorized and passed up to a higher level for consideration, it does not make sense that Causes could be aggregated in this way.
 
 Looking through the lens of Learned reactions, we find that a different notion of the Causes and Clauses framework can be applied. We see as context for this level the combination of unsupervised and supervised learning, as will be explained below. We will again talk about Things as perceptual data, but this time as residing on a medium-level. Such a Thing might well be thought of as a prototype or the principle components that represent “objects” whether they be physical or abstract. A Thing would be well described using a noun. Things could be thought of as composed of lower-level perceptual data into a more coherent and general symbol, such as the concept of a chair can be the combination of many views of chairs plus a notion of how a chair is used. As before, Differences are medium-level sensory descriptions of dissimilarity.
 The preceding description of Things and Differences captures the unsupervised learning aspect of the Learning level. This is demonstrated upon consideration of the way that these representations would be acquired. The most reasonable way that this could work is through a self-organizing process that categorizes or clusters similar percepts. Thinking about Things as feature vectors and Differences as dissimilarity lends itself to self-organization because algorithms that accomplish self-organization use those data types.
 A different kind of learning, supervised learning, suggests a different way of viewing Things and Differences in this level. But while unsupervised learning is content to occur without much reference to the temporal realm, supervised learning cannot be work without it. In particular, we choose to add a temporal dimension, as the sequence of presenting a stimulus and observing the response is meaningless without it. Once we can stamp Things and Differences with times, we can begin to look at Things as the state of an object, and Differences as trajectories that evolve over time. Clauses in this level can be described as state-trajectory-state transitions. [here we cite Jackendoff]
 This description covers some ground towards describing supervised learning, but leaves out a significant notion, that of reward. [Some initial musings on why reward is generally a good thing to have in an explanation of intelligence] In particular, for our explanation, it is clear that supervised learning is meaningless without some way of deciding between the things you want to remember and the things you don’t mind forgetting. . We feel that the notion of reward fits that purpose, and thus we will carry it through the following sections alongside the Causes and Clauses framework and analyze how further levels can be thought to involve reward.
 
<new>
 
 The last thing that needs description in the Learned reactions level are Causes. A starting point for thinking about Causes would be to consider them as Before, we had thought about Causes as the explanatory agents responsible for Differences, but . In this case, this description falls short for a few reasons. For unsupervised learning, it does not make much sense to attribute an agent to the difference between objects. Asking why a prototypical apple is different from a prototypical orange isn’t likely to add much useful information. However,But in supervised learning, it also does makedoesn’t make sense to ask why one state is different from another state. In any kind of supervised learning, the “agent” responsible for the state of the system being different is always the same; it is the sheer fact that a Difference between the two states exist.  In the example of the backpropagation algorithm, the weights of the synapses of the network will be caused to change simply because What might have caused a particular apple to be transformed into a particular orange would certainly add useful information. Because of this, it seems like Causes are far more useful when considering supervised learning. In order to define Causes in these terms, let us turn back to the state-trajectory-state transitions. It makes sense to think of the Cause as the initial state plus the trajectory from that state. It seems somewhat meaningless, however, to consider this cause in the absence of a consequence. [general discussion about causes and consequences as a good way to slice up the world]. Therefore, we also feel it important to carry the notion of consequence alongside the Causes and Clauses framework, and examine how it may be incorporated into higher levels.
 
 
STOP HEREthe network returned a value which didn’t match a target. In general, we don’t think of a feedforward network as representing why the synapses change; it is only important by how much they should change.  We feel that this same philosophy generalizes to all of the supervised learning methods
 
 
 used at this level.
 
 To understand the notion of thinking on the  On the level of DDeliberative thinkinglevel, let itus begin by defining what sort of Clause we see resulting from this level, and work backwards to explain what parts are required to construct it. We view this level of thinking as planning, where “sub-plans” are chained together in a sequence for some purpose. In order to investigate further, let us consider first what these “sub-plans” might be, and then consider what we can say about the purpose that planning serves. We believe that a profitable way to think about “sub-plans” is as state-trajectory-state transitions, as we established were the Clauses from the Reflective level.  Generally one thinks of a plans as: “first I do X1, then I do X2, finally I do X3”. We envision this as a sequence of three transitions, where a single transition involves going from a state where Xn isn’t done to a state where Xn has been done. The trajectory in this case could be described as “doing”. For our purposes, it makes sense to call these transitions the  makes more sense to think about Things for this level. Additionally, we can imagine Differences as dissimilarities between transitionsas transitions and Differences between transitions.
 What about the purpose of a plan? It is now important to consider the role of Reward on this level. One way to think about how a plan is motivated is by the fulfillment of some kind of need. Thus, we can define Reward as the degree to which the current plan is determined to fulfill the present need. With this in mind, we can be more specific about the process of Deliberative thinking.  The process can now be described as generating test chains of transition s, evaluating those test chains on the basis of Reward, discarding the parts that don’t work and adding parts that do to, resulting in a final plan.
 To aid in the process of Deliberative thinking, it now seems important to introduce the concept of Cause.  [In particular, we will say that a Cause is a transition which produces a Consequence, which is some transition that happened as the result of a Cause???]As before, Causes are the explanation for there being a Difference between Things. We might think of the following Clause that illustrates these roles: "When I bake a cake today, it is okay to touch the handle of the stove, but not the stove, because it is hot". In this case, the transitions concern such transitions as baking a cake (from not baked to baked), touching the handle (from not touched to touched), and touching the stove (from not touched to touched). As these transitions are compared, their Differences become obvious (touching the handle versus touching the stove). Here, the Cause implies a further transition that is not stated in the clause--that your hand will go from not painful to painful if you touch the stove. This transition serves as an explanation for the difference between touching the stove and the handle. [Talk about Causes-Consequences here] The Cause-Consequence relationship can be formed on top of the sequence of transitions transparently. It is not necessary for a transition to be either. A plan may include a series of transitions that are Causes followed by a single Consequence, or a plan may have a set of nested Causes and Consequences, or a plan may have neither Causes or Consequences, but just a serious of transitions which have not been assigned any particular causality.
 
</new>
 
 
 As we move above the level of Deliberative thinking, we can begin to see a pattern where the Things one level up are the Clauses from the level below. For Reflective thinking, a Thing can be thought of as a Clause from the Deliberative level, or a 'deliberation'. Differences are formed between these deliberations. Causes are the rationale you have for coming up with those plans, and Clauses are the records of that process. Here's an example of a Clause in reflective thinking using this framework: "When I planned to touch the handle of the stove rather than touch the stove itself, I still burned myself. I didn't consider that the handle would be too hot, because it looked insulated". The deliberations are the plan to only touch the handle of the stove, as well as the consideration that the handle would be too hot. The Cause in this case is the rationale that the handle looked insulated.
 Self-reflective thinking follows the same pattern started below in Reflective thinking. Its Things are 'reflections', its Differences are comparisons between those reflections, and its Causes are the rationalizations for those Differences. An example Clause for this level would be: "How could I have failed to realize that the handle was going to burn me before touching it? I must have been distracted."
 Finally, Self-conscious emotions are the crowning layer in this recursive process. Here, Things are 'self-reflections', and differences are found by comparing the 'self-reflections' that different people you know might place upon your actions. An example clause might be: "If anyone else had seen me burn myself on the stove handle, they would have laughed at my ignorance."
 
 Now that we have outlined how these ideas might be more profitably connected, let us examine a few trends that this explanation appears to signal. First of all, one can see the chaining process suggested by the Causes and Clauses framework operates as one goes up the levels. Because higher levels use the structures created on lower levels, an important feature of the framework is conserved in this combination. Secondly, we notice that Clauses become more complicated sentences as one goes up. Lastly we notice that this treatment demonstrates that Causes become less necessary as we move upwards through the levels.

conclusions

new glossary

1. we like C&C – the idea that decomposition, and certain kinds of decmposition are integral to the notion of ingelligence generically

though we see clause as just a kind of way of composing new compound things out of other things

2. that you need reward as a fundamental C&C component

3. that if you take the sparse cross-product of the M6 and a version of C&C, then you get a richly generative matrix

4. we try generating one level down from things, differences + reward

causes, consequences and goals

other possible shit that might pop out

adjectives – might be a useful notion too – reflexive difference on a thing, comparing it to its prototype

extra paragraph

What might have caused a particular apple to be transformed into a particular orange would certainly add useful information. Because of this, it seems like Causes are far more useful when considering supervised learning. In order to define Causes in these terms, let us turn back to the state-trajectory-state transitions. It makes sense to think of the Cause as the initial state plus the trajectory from that state. It seems somewhat meaningless, however, to consider this cause in the absence of a consequence. [general discussion about causes and consequences as a good way to slice up the world]. Therefore, we also feel it important to carry the notion of consequence alongside the Causes and Clauses framework, and examine how it may be incorporated into higher levels.