Intelligence

With jaws that bite, and claws that snatch...

Intelligence is understanding

Intelligence is best imagined as a generalized ability to understand problems or situations, sight unseen, without specific preparation.  The intelligent individual (human, animal or artificial) possesses a generic internal template of subjects, behaviors and situations that they are able to deploy with a minimal amount of preparation.  

Intelligence is knowledge about knowledge 
Intelligence is contrasted with knowledge, defined as the (instinctual and/or learned) capacity to invent and govern those behaviors which satisfy the organism's desires.  Knowledge involves the mind's acting in the 'forward' direction, going from
(a) an emotional drive base (desires, appetites, learned ways to satisfy needs), then
(b) an internally representative stage, with action plans as top-level thoughts to
(c) behaviors.
This three part paradigm is sometimes called the Marr Trilayer*.  In this scheme, intelligence is the thematic complement of knowledge, because it involves, in extremum, inferring internal motivation from observations of external behaviors. More generally, intelligence allows us to infer latent causes from manifest effects, irrespective of the type of, or true nature of the agent. Intelligence can therefore be secondarily defined as meta-knowledge, or knowledge about knowledge. If knowledge is the ability to link cause to effects (a function), then knowledge about that linkage/mapping will clearly be needed to link effects back to their causes (to invert the function).

Intelligence is not expertise

With advance knowledge of specific situations, it is possible for agents to appear intelligent, but this is not true intelligence, rather, it is best defined as expertise. Consider for the moment the work of Rodney Brooks, fellow Flinders University Alumnus, and big wheel in the world of industrial robotics. Two of his papers which spring to mind are 'Intelligence without reason' and 'Intelligence without representation'.** 

Brooks intention seems partly that of 'agent provocateur', wishing to goad his readers into re-examination, and perhaps overthrow, of the accepted tenets of the AI field.  In both papers he seeks to suggest that intelligence without a central GOFAI-type of plan or skeleton (eg a reason engine) could perhaps emerge from cooperating sub-systems in an unplanned, bottom-up manner. 

Bottom-up intelligence is an oxymoron

Brooks is, to put it bluntly, wrong. Specific effects have specific causes, and causes have adaptational provenance (past experience of learning). That is (literally) what 'specification' means. Intelligence, however one defines its particulars, must have a generic internal template at its heart, something that looks very much like the TDE. In fact, since I (or any other cognitive scientist) have been unable to conceive of any other governance/ modelling template which exceeds the TDE's ability to properly characterize subjects, behaviours and situations, it suggests that all instances of true intelligence (biological or artificial) must contain a TDE pattern or its functional equivalent somewhere within their 'codebase'.*** 

The TDE is the most competent template for intelligence

In other words, it is hard to imagine**** a pattern which is more generically competent than the TDE, since its origins lie in the most fundamental of cybernetic circuits, the homeostat. 

*covered more thoroughly on another page

**One awaits with bated breath (and tongue in cheek), the final paper in the tryptich, 'Intelligence without intelligence'. 

*** I use this term loosely, with the readers informed permission.

**** Of course, one should never say never. The future hasn't happened yet.

Legg & Hutter (2006) have collected over 70 independent definitions or descriptions of intelligence. They found an overwhelming consensus which focuses on the following mental processes- acquisition and application of relevant knowledge (meta-knowledge) with the goal of understanding the key causal relationships within a situation or domain, employed by an individual or group to understand and optimally exploit their predicament (understanding and implementing causal mechanisms). 

Knowledge, semantics and facts (predicated truth-states) are the same thing
The Cambridge Advanced Learner's Dictionary (2006) suggests that intelligence is- 'The ability to learn, understand and make judgements that are based on reason'. Many, indeed most of the 70 sources quoted included a reference to reason. This is implied by the term 'knowledge' which in this context (TDE theory) is synonymous with semantics or meaning. If knowledge, semantics and predicated factuality (factual hierarchical contexts) are NOT defined as equivalent constructs, then the arguments supporting TDE (or any other similar) theory have no logical closure.  They are equivalent in reality, so they should also be equivalent within any theory that is expected to faithfully reflect reality.

Reason is the use of logical relationships to extrapolate existing knowledge 
If however, the definition of knowledge as per TDE theory is used, namely that it connotes a hierarchical accumulation of inter-related facts (subject - predicate subordinate links or catenae), then the use of reason is not only implied, but necessitated, since reason is defined as the logical combination of factual or circumstantial (primary) data to create new (derived or secondary) facts.   In other words, reason is the use of (propositional or predicate) logic to extrapolate existing knowledge. 

Reasoning is applying transitivity to set inclusion

Reason is a built-in feature of contextual factual (possibility) hierarchies. What turns them from 'flat' semantic networks into hierarchical structures is the superset-subset relationships we see in reason. {Socrates is a human + All humans will die = Socrates will die} is the famous 'syllogism' expressed as three sentences, but it is also a Venn diagram consisting of three nested sets. The outermost set describes the things that die. Inside this set is the set of all humans, and inside both sets is the singleton set of Socrates. Putting Socrates (or indeed any other person) inside the human set automatically places him inside the mortal set. 

linguistic description is reasoning about semantic sets

The use of this construction reveals another important feature - nouns (names of classes of similar things) are recursively decomposable into adjective-subset pairs. Humans are mortal things. Politicians are a necessary evil. Bicycles are two-wheeled vehicles. Indeed, all sentences are logical linkages, acts of extrapolative reasoning which 'grow' the logical remit of their enclosing paragraph. Or, put most simply of all, linguistic description is reasoning about set inclusion (feature set attribution). A is in B, B is in C, thus A is in C. This we immediately recognise as the mathematical property of transitivity.  The Venn diagram is one form of depiction, but another, entirely equivalent, method is a hierarchical tree.

In set notation.......... (R⊆S)∧(S⊆T)⟹R⊆T. . . . . . . . . . . Eq(3) ie all subset relations are transitive. Another type of subset relation is inequality. To say that one value is greater than another implies that, if expressed numerically, the first value contains the range of numbers of the second value. Why is this important? Because it provides a point of consistency between measured (analog) and enumerated (discrete) quantities and variables, suggesting that they can both be represented by common semantic mechanisms.

Each sentence or clause (verb phrase) processed adds a point to an internal semantic hierarchy
Each natural language description is a way of placing a semantic value (a point, a singleton) in a superveilant hierarchy (tree) of other semantic values. Each description automatically joins several subtrees, allowing new logical relationships to be created (entailed). This is what reasoning is, it is what we do when we deduce or infer or adduce or entail use any of the other types of logical process.

The mechanism which underlies belief belongs to this class of semantic operations. With belief, however, the mental operations are subconscious. We collect data by means of daily observations, but most of these are not attended to. That is, the observation event occurs without leaving a trace on our episodic/ autobiographical memory (the one in the LCH, according to Tulving). Each observation consists EITHER placing a new percept within an existing percept class, (paradigmatic - a subtree operation) OR linking percepts at the same semantic level (syntagmatic). If these combinations 'pass muster', ie they are logically consistent, they do not 'break' the relational database (eg do not fire 'tautology-redundant' or 'oxymoron-inconsistent' warning signs),  then these novel semantic values will be candidates for inclusion in general knowledge (permanent world facts, stored in the RCH according to Tulving's model). General knowledge is INTERsubjective, as opposed to episodic knowledge, which is INTRAsubjective.   The way that belief and knowledge are linked is discussed in another page. 

Intelligence is abstract, top-down use of reason

Brooks' idea of 'subsumption' architectures describes computational systems which use bottom-up methods to perform intelligent behaviors. But from the above analysis, reasoning must involve top-down processing, since it involves placing sets and singletons into hierarchies. Hence subsumption as a way of implementing reasoning seems, prima facie, to be fundamentally flawed, indeed, oxymoronic. In other words, intelligence cannot be defined without reference to abstract (= high level = top down = hierarchical = global-local) concepts. To suggest otherwise is disingenuous at best.

This is not to say that Brooks' research is without merit. Far from it. From his earliest work on insectoids such as the ultra-cute Ghengis, through the equally charming Kismet and Cog, to his most recent foray into selling industrial robots that are better value for money*,  Brooks, together with Boston Dynamics**, has become the public face of the global (= American, mostly) robotics industry. Without publicity there are no funds, and without funds, there are no research projects.

*because they are designed to be generic and flexible (intelligent) instead of specialized and dedicated (bespoke), as well as needing no dedicated programmer on call.

**who could forget the headless and terrifyingly life-like Big Dog! 

AI vs A-squared-I-squared

We have established that in order that we preserve commonplace semantics, we must also preserve a consistent relationship between the concepts of reasoning and intelligence - they must therefore both be top-down. Therefore we need (i) another name for the sort of intelligence-like behavior that Brooks has drawn our attention to, as well as (ii) an explanatory mechanism for it. 

It is possible for a bottom-up system to Act As If Intelligent (pronounced 'A-squared I-squared')* This is certainly what Brooks suggests with Ghenghis and other insectoids. These 'arthrobots' embody an apocryphal tale which also happens to be true.  Attempts by Brooks to write a supervisory executive (a mini-operating system) linking the drivers (firmware, or ROM) for the six articulated leg actuators continually met with problems, until, in frustration, Brooks disabled the central executive module, effectively allowing each of the six legs to 'do their own thing'.  Surprisingly, the insectoid was able to walk better without it. Furthermore, the rougher the terrain, the greater the improvement in overall locomotive efficiency.

Programming language compilers are formally similar to (artificial and biological) intelligences 

There are both superficial and deep similarities between programming language compilation and both sorts of intelligence (ie biological and artificial). Every compiler has at least the following two stages, with the 'program'** as the input data (see figure 7(a), reprinted below for your convenience)-
Step 1 - Paradigmatic (local symbols)- First the compiler uses a lexical (symbolic) analyzer to identify the tokens (symbol copies, which are the reserved words and user-defined names), and put them into semantic classes (super-symbols).
Step 2 - Syntagmatic (global expressions)- Then, having established the semantics (meaning) of the symbolic 'atoms', the compiler uses a 'parse tree' to first check for correct expression (ie check that the sentences make sense), then implement the rules of the grammar using the annotated words (typed tokens) output from the lexical analyzer.

When the compiler (which is itself a program) is 'run', the output is equivalent to the behavior we call 'intelligent' - in other words, we have found a way to differentiate AI (the compiler, the generic knowledge base) from A-squared-I-squared (the running program, the intelligent behavior).

Brains can be regarded as hardware and minds are software 
The interesting thing about this formal definition of intelligence is that it focusses attention on something that is easily overlooked, namely the program.  This formal definition reminds us that intelligence is a tool that does nothing without its user's intentions, a set of experience-led drives to satisfy (=drive reduction of) inner goals. Computer software is directly comparable to the contents of the programmer's mind. This is theoretical affirmation that brains can be regarded as hardware and minds are software.

The end goal of compilation is semantically correct machine behavior 
This directly refutes the view of GOFAI as incorrect, and indirectly refutes previous attacks on GOFAI, based on semantic arguments, such as Searle's Chinese Room. Clearly, viewing intelligence as compilation is GOFAI, since it uses one of the most common computer science paradigms, namely software compilation. Equally as clearly, this vision of AI semantics, it positively embraces it, since the end goal of compilation is semantically correct machine behavior.

*Almost any sort of mathematical function or measurement applied to data is partially intelligent, because it does part  of what a compiler does - eg run-length encoding of any file (equivalent to a form of lexical analysis), measuring the area of a room (equivalent to parsing the overall structure).

**FYI programs can be viewed as consisting of fixed and variable parts- commands ('must-do') and conditional requests ('wish-list'). A useful mnemonic is cooking food- the fixed part is the recipe, while the flexible part is the shopping list. Variability in the fixed part occurs with problems that are handled by exceptions, while variability in the flexible part occurs with conditional tests such as if-then constructs.

Language use is semantic (knowledge base) management

Reasoning and belief systems have been more than adequately explained as semantic hierarchies, in which considerable effort is made, via subconscious processes, to manage our (episodic and general) knowledge repositories, keeping them both efficient (by trimming tautologies and flagging redundancies) and consistent (by pruning oxymorons and eliminating contradictions). Language can now be viewed as that knowledge-base management process. External language (eg heard speech or read text) introduces the possible inclusion of new facts, as long as they pass efficiency and consistent tests.  

Internal use of language is thought

Internal language (thought*) serves a similar process. Figure 16 depicts the three TDE levels of biocomputational embodiment, but it also suggests something more profound - the association between modes of neural representation (internal) and modes of external behavior (external). Specifically, it associates TDE level 1 constructs - eg joint angle rotation values- with low-level ('motor') memory, TDE level 2 constructs -eg posture array values- with symbolic combination of level 1 constructs into compound (eg visual) representations, and TDE level 3 constructs - language expressions- with symbolic combination of level 2 constructs into compound (eg acoustic) sequences for processing as speech.

*according to TDE theory

Visuospatial (eg faces) and Phonological (eg voices) channels have similar information capacities

Figure 16 summarises a specific set of associations, from which we may reasonably infer some important predictions. The most obvious one is that TDE level 3 development, that is, the formation of human language from animal communication was caused by a troika* of co-evolutionary positive-feedback loops between (i) increasing vocal range (acoustic- anatomic) (ii) phonological repertoire (mnemonic- neurological) and (iii) additional memory size and data structures.  Each human individuals voice (formant array treated as a perceptual representation) contains just about as many identifying features as does the image of their face (feature array). In other words, vision (visuospatial scene understanding) and speech (phonological loop/sequence understanding) are perceptual abilities of roughly equal individual capacity**. But the most significant improvements in fitness are not just operational and external and operational, they are internal, neurocognitive and organisational. According to TDE theory, these co-evolutionary processes eventually lead to the increase of the number of fractal TDE levels from two ( in most animals) to three (in all humans). This represented not just a stepwise increase, but a whole order of magnitude enhancement in evolutionary fitness. Indirectly, it gave rise to language itself***. 

*originally a light horse carriage towed by three horses (Russian)

**This is not news to the millions of people worldwide who love radio plays, avidly follow cricket commentaries, or who listen to audiobooks while driving long distances. It is also a godsend to the blind. A blind animal which is not a human pet has a poor prognosis indeed, while a blind human is merely disabled, and can live a relatively normal life after minimal training. 

***We must be careful to distinguish between the linguistic nature of all biocomputation (Chomsky's i-language), and the rise of full-featured human language.