I La Arqueología en la bahía de Algeciras
I. LA ARQUEOLOGÍA EN LA BAHÍA DE ALGECIRAS 1 Una bahía entre dos mares.
I.3. Evolución del conocimiento arqueológico de la bahía de Algeciras 1 De los hallazgos casuales a las primeras excavaciones (1900-1985).
I.3.1.1. Primera mitad del s XX La Conferencia y los arqueólogos extranjeros.
4.2
The novelty model is quite complicated. It is an abstract architecture existing of three different mediums and four novelty anchors. Novelty anchors are abstract constructions, like meta-‐systems they become specific when considering an actual implementation of the novelty model. The essence of the novelty anchors is that they allow agency to arise. The four different anchors can together create the bootstrapping for the novelty model to overcome a conservation constraint. The case of sailing upwind should introduce the mechanism of novelty regulation in a fluid-‐based medium, where the anchors are physical forces.
Consider that in a fluid-‐based medium the artifacts were limited. So the novelty regulation in a fluid-‐based medium is limited as well. The sailing upwind, examined in the first subsection only creates novelty in respect to the movement. Because of the primitive nature, the case can demonstrate the fundamentals of novelty regulation. All other subsections focus on the novelty model that I call "Cohering". The Cohering model can show how knowledge can emerge. In this case, there is no limitation to the agency and so with the Cohering model we build an argument of how the novelty model is essential for an agent.
Sailing upwind overcomes the paradox of momentum conservation. The Cohering model overcomes the bias of knowledge. The bias of knowledge is also a conservation constraint. Knowledge is both a barrier and the source for knowledge creation, similar to how momentum is both a barrier and the source of sailing upwind. This is the essence of the novelty regulation: the ability to overcome a
seemingly unbreakable barrier. In following chapters, other novelty regulations are
examined, while, depending on the implementation of the medium and anchors, some kind of conservation gets tackled. The novelty regulations for these chapters all relate to innovation, they are more complicated to understand and therefore the Cohering model seems most fit to demonstrate the general novelty model.
For the Cohering model, the mediums are memories and the anchors are anticipation processes. The second subsection creates an argument for the three involved memories. The third subsection introduces the anticipation process, as the building block for the novelty regulation process. The following subsections create an extension of a basic process to build the novelty model. The internalizing and externalizing processes create a default flow, which was demonstrated in Section 2.2.4 by the perception of a face. The example is also going to be used to demonstrate a more formal notation. The default flow uses default categories for associations and tags. It demonstrates the default bootstrapping cascade and involves the three mediums. This default flow has to be extended twice to get to an actual novelty model.
The first extension changes the default flow to a directional flow by adding challenges. Challenges change the default categories to context dependent categories, which allows control by focusing on the context. Just like associations
and tags, the challenges have to get anticipated. The anticipation can lead to a cascade of different associations and tags that may lead to reinforcing or falsifying the anticipated challenge. The challenge creates the required tension for novelty to emerge.
The last process is for the novelty to evolve. It changes the directional flows to a constructive flow. With the Cohering model, the evolving process works on experience in the workspace. The evolving process is anticipating what aspects in the workspace are relevant and creates conceptual networks of the novelty. The directing and evolving process are part of the regulation and they can focus on internalizing or externalizing. By focusing on internalizing, the novelty regulation is modeling experiences into knowledge. By focusing on externalizing the novelty regulation is mastering the novelty. It creates, metaphorically, the same oscillating movement as seen with the sailing against the wind. The modeling and mastering are examined in the next section on learning novelty.
Novelty regulation: the case of sailing upwind
4.2.1
The source for sailing is the wind working on the sail. There are actually two forces on the sail: drag and lift. The drag force is the commonly known force. It pushes the boat along in the direction of the wind. Drag is not useful for sailing against the wind. The lift force is more difficult to understand. The basic principle of lift force is to create a curved surface so that a flow (the wind in this case) would pass with different speeds, creating a difference in pressure. The effect of the difference is a force in a straight angle to the sail. The lift force is not only important for sailing against the wind. It is the same force that keeps airplanes in the sky and makes our modern wind turbines work.
The lift, acting on the sail, is not enough to move against the wind. At best, there would be almost no drag force and only the lift force, resulting in an angle close to 90° opposite to the direction of the wind. A price needs to be paid to move in the direction of the wind. This requires a mechanism existing of some extra forces. The mechanism involves the viscosity of the water and the specific position of the keel. Viscosity is a measure of the resistance against a deforming of the fluid. Boats are designed to use viscosity to their advantage. The keel of the boat is a counter weight at the bottom of the boat that often looks like a long blade. It makes sure that the boat can easily cut through water along the length of the boat, while at the same time restricting sideway movements. By positioning the boat in a 45° angle with the wind, a reaction force occurs on the keel that is optimal to move against the wind. The reaction force and the lift force create a tension. The combined force is smaller (the price paid) and goes into the direction of the wind (forces A, B, C in Figure 4.10). One fascinating conclusion is that without tension, no novelty can exist. It gives part of the friction experienced in innovation a very different meaning. Consider that the medium defines what novelty can arise. In this case, the novelty related to the movement. According to the law of conservation of momentum, it is paradoxical to
move against the wind. Thanks to the described mechanism, it does become possible, which requires a saw-‐like path of movement (from A over B to C in Figure 4.10). The same reasoning applies to other novelty regulations showing how bootstrapping can overcome some conservation. The following subsections describe the novelty model around knowledge creation called Cohering. In this case, knowledge creates bias and fundamental new knowledge can only emerge by the saw-‐like path between mastering, meaning actions to refine the regulation and modeling, meaning actions to refine the internal representation.
Figure 4.10 How sailing against the wind can be achieved
After examining the Cohering model in detail, Subsection 4.4.1 returns to the sailing against the wind case. Because fluid-‐based mediums do not store memory, it becomes a fascinating case to generalize novelty regulation as a mechanism to overcome conservations and deal with complex adaptive environments. It creates the proper context for the following chapters where different domains have approached complex adaptive environments and develop at least part of the novelty model for the conservation they get challenged by.
Three types of memory
4.2.2
The mind needs a memory to work. An extended mind makes it complicated, as memory does not restrict itself to the brain but extends to workspaces like note books, blackboards, kitchen tablets, desks, online spaces, etc. In other words, the environment is a memory too. A workspace is needed to create information, which is a third required memory. As argued before, for a cognitive system only space is needed to make an external environment become a workspace. For example, Clark illustrates how space distribution in the kitchen contributes to problem solving, like separating washed from unwashed vegetables. (Clark 2008, p 46) makes an important contribution to our investigation:
It is intuitive that once descriptive complexity is reduced, processes of selective attention, and of action control, can operate on elements of a scene that were previously too “unmarked” to define such operations over. Experience with tags and labels may be a cheap way of achieving a similar result.
Tags are the element regulating the external anticipation process in the default flow. The last remark by Clark "tags are cheap" needs some nuance. It seems to indicate that spatial distribution would be more expensive than labeling. Spatial distribution
is less robust, but a label requires a culture to make sense of the label, the production of paper and markers for the labeling itself. Once the environmental enrichment has built tagging, it can indeed become cheaper, but not by default. Workspaces are mostly overloaded with tags, at least for the trained eye. An untrained eye can show remarkable blindness to such tags, as it is usually not the intention to attract the untrained eye's attention.
The environment is its own best representation. For example, it is hard to memorize spatial aspects exactly, like colors, because it is so easy to perceive the color. Having a general description is often enough. Even for more abstract aspects, the environment can be a proper space to store knowledge. This can include creating maps or writing books. One misconception is that retrieving information from the environment would be slow. (Clark 2008, p 380) mentions Gray and Fu (2004) to prove this is wrong:
Instead, they [Gray and Fu] argue that their results show that “the time retrieving something from memory is weighed the same as time spend in perceptual-‐motor activity” and that it is therefore a mistake to “presume the privileged status of any location or type of operation” (Gray and Fu 2004, p 378, p 380).
In summary, three types of mediums are essential for the novelty model. In the Cohering model it becomes three type of memories:
• System Core = Long-Term Memory: the memory contains knowledge. Knowledge is experience
processed to abstractions and ready to use concepts.
• Workspace = Working Memory: the temporal memory that creates tension and focus, allowing
novelty to emerge. The workspace is the centerpiece in the novelty model.
• Environment = external memory: tags, signs and meaning in the environment. Some are created by
the agent, other are serendipitous.
Anticipation as a building block
4.2.3
The anticipation process is considered the cognitive building block of the novelty regulation process. Anticipation works on an information flow with feedback and feedforward. To provide additional detail of what is happening during an event (t), I consider a time moment just in front of the event (t-‐) and a moment just after the event (t+). At t-‐ the feedforward suggests what may happen, at t+ the feedback contains what has happened. The moment that things are happening is when output is created. The t-‐ and t+ are relevant to understand the anticipation process. Figure 4.11 illustrates two ways to present an anticipation process. The left side of the figure illustrates the information flow from input to output. On t-‐ input to the feedforward process predicts an outcome. After the output, t+ feedback updates the long-‐term memory used by the anticipation process. The right side of Figure 4.9 shows an alternative presentation by combining input and output to the same workspace. The benefit of this figure is that multiple anticipation processes can work in parallel on the working memory, which is required for the novelty model. Figure 4.11 Two presentations of anticipation: left side focuses on the flow, the right side on the
An event progresses in time, which can be approximated by a discrete number of steps (t1, t2, t3, …). Each step has a feedforward t-‐ and feedback t+. In real situations,
the t1+ feedback and t2-‐ feedforward would overlap. So t2-‐ may still be working on
the incorrect memory. To simplify the process for the theory, consider small enough steps, so no overlap would exist : t1-‐ < t1+ < t2-‐ < t2+. The pattern used in the
anticipation process will be presented as a set of variables X = {x1, x2, x3,…, xn}. The
variables can be presented as numbers. It is, however, essential that we consider sets, as networks of associations, tags, challenges and experience are in some way connected and the connections allow for the novelty regulation to manifest.
The anticipation function ' ϒ ' predicts what X would be. We can use t-‐ and t+ to describe when an anticipation is incorrect: ϒ(Xt-‐) ≠ Xt+. Delta-‐Learning in neural
networks (Jacobs 1988) uses the difference between target output and the actual output to calculate a delta: Δw = (target – actual) * neural_activation. For the novelty regulation, a similar delta is expressed as the symmetric difference:
X ΔY = (X∪ Y) \ (X ∩ Y)
The symmetric difference contains the novelty. It is the set of what was expected and did not manifest (wrong knowledge) and what manifests that was not expected (unknown observation). In case what was expected and what manifest are the same we get an empty set. This delta can be used with the anticipation function:
ΔN (X) = Xt+ Δ ϒ(Xt-‐) = (Xt+∪ ϒ(Xt-‐)) \ (Xt+ ∩ ϒ(Xt-‐))
When an anticipation is correct we get ΔN (X) = Ø, if not we get ΔN (X) = Y, where Y is
the set of differences that need to be understood. To express this, consider two experiences at a different time around the same set of events: . For example, pouring out some liquid in a glass can be an event. By pouring out different liquids and by using different glasses a more general conceptualization of the event is reached by the different experiences.
A correction has occurred if ΔN(exp1) ≠ Ø and ΔN(exp2) = Ø. The correction does not
directly mean novelty, maybe exp1 was a miscalculation, maybe the anomaly did not
occur any more at exp2. To deal with such perturbations, it is essential that many
experiences exist of a particular event. The event would be incomprehensible if:
ΔN(expi) ≠ Ø, .
Default flow
4.2.4
The default flow is the bootstrapping cascade between associations and tags as illustrated with the perception of a face (see Section 2.2.4). The associations and tags are specific to the Cohering model. In general, the associations become internal matches, while the tags become external matches. The basic anticipation design (right side of Figure 4.11) only has a working memory(the workspace). With the default flow, the basic anticipation design is extended, relating to the System Core (the Long-‐Term Memory) via an internalizing anticipation process and the environment via an externalizing anticipation process. Notice that in contrast to the basic anticipation design, this design does not contain a correction, which is reinstated with the constructive flow. The default flow does use all three mediums of the novelty model and allows regulation, if knowledge exists.
Figure 4.12 Internalizing and externalizing creating a default flow
The memory element X now gets a dimension related to the four processes involved: internalizing (in), externalizing (ex), directing (di) and evolving (ev). With the default flow, the memory element becomes: X = {xin1, xin2, xin3,…, xinn, xex1, xex2, xex3, …, xexn'}, where there are n associations and n' tags. The information flow from the design that is defined as a specific anticipation becomes ϒα and works on Xα = {xα1, xα2, xα3, …, xαn} ∀𝛼 ∈ 𝑁𝑜𝑣𝐷𝑖𝑚. 𝑁𝑜𝑣𝐷𝑖𝑚 is short for the four dimensions that the novelty can have (in, ex, di, ev). The anticipation processes are parallel processes that have no overlap of values: Xα ∩ Xβ = Ø for α ≠ β and ∀𝛼, 𝛽 ∈ 𝑁𝑜𝑣𝐷𝑖𝑚
To illustrate the default flow, I reconsider the Gestalt bootstrapping example for perceiving a face (see Figure 3.8). This example has some simplification that I undo in the later formalization, but it seems best to introduce the notion in a simplified way. The example demonstrates how anticipation happens with intermediate stages before a stable pattern would emerge: ϒ(X) = X', meaning that the X was get refined
to X', in that case X was a more simplified description than X'. Many steps may be required before a useful description is reached: ϒ(X') = X'', …, ϒ(Xn-‐1) = Xn In our face recognition example, only a few steps existed: ϒe first finds Xround, became more
concrete with Xface to finally fix on XAlice. To indicate that intermediate steps exist, I
use the ϒ(X) → Xn as a simplified notation.
The face recognition demonstrated how two anticipation processes can reach what each cannot do separately, starting from nothing it recognizes the face: ϒ(Ø) → Xface while ϒα (Ø) Xface∀𝛼 ∈ 𝑁𝑜𝑣𝐷𝑖𝑚. In other words, coordination between the
anticipation processes is needed. The initial step in the figure is ϒex(Ø) = Xshape and
ϒin(Ø) = Ø. The externalizing gets light as input and recognizes a shade, resulting in Xshape = { 𝑥!"#$%$"&!" !, 𝑥!"#$%$"&!" !, …}. Xshape only contains concrete tags and no
associations. Internalizing only works if there is some concept to also start associating. Shapes are recognized in milliseconds resulting into ϒe(Xshape) = Xround. Round is a concept that triggers associations: ϒin(Xround) = { xinball, xinapple, xinface}. In the meantime, the external is too complicated and no information is gained:
ϒex(Xround) = Xround. This local maximum is only temporal and leveraged by the internal process. To visualize this, think of the Gestalt emerging where no clear pattern was recognized (Figure 2.15 D of a dog sniffing a tree). In the case of the phase, the working memory contains Xround = {xinball, xinapple, xinface, …, 𝑥!"#$%!" !"#$!!, 𝑥
!"#$%!"#$!!
!" , …}, where part-‐1 and part-‐2 are concrete external tags and xiface refers to Xface.
By retrieving the Xface from the Long-‐Term Memory, a new set of relations is added
to the workspace: Xface = {Xnose, Xupper-‐lip, Xleft-‐eye, …, xinfriend-‐or-‐foo, …}. If an object is more complex, as with the face, several steps are needed to make the object concrete. In the novelty model, an operation queue exists. It contains the first associations and push new associations to the back: Q = {Xball, Xapple, Xnose, Xupper-‐lip , …}. In the next steps ϒex(Xround) ≠ Xball and ϒex(Xround) ≠ Xapple, meaning that external feedback rejects the conceptualization of a ball or apple. While ϒex(Xface) is falsifying
the associations, ϒin would get the next associations related to round, maybe ϒin(Xround) = {Xwheel, Xplate}. Of course, it becomes harder to find more associations if
no new verification comes from the external. Lucky ϒex(Xround) = Xface is verified, so
the next steps in the bootstrapping cascade can be explored, which leads to XAlice.
Notice how both ϒin(Xround) = Xface and ϒex(Xround) = Xface are needed to become