Definitions

Terms such as pattern, linkset and pattern intersection are helpful to limit the number of words needed to describe Pattern-atom (patomic) theory.

At a very high level all that a robotic brain is required to do is receive inputs from senses and then determine and execute appropriate actions. Actions result from controlled muscle contractions (or alternative motion devices like speakers or motors). In fact, the primary purpose of a brain may be to enable movement (Greenfield, 1996) to effectively react to the environment. The challenge of mechanical control appears to be in finding the right sensory pattern elements that identify real-world objects. Following this recognition, the machine then moves itself by selecting the right pattern of muscle elements to operate with the objects. As animals like dogs and cats recognise the function of a door, it is fair to suggest that many brain functions do not rely on the human brain’s unique capabilities, but rather on our brain’s common ancestral capabilities. Here we explore the terms behind the robotic brain concept: snapshots, linksets and Patoms.

Pattern-atoms

The first task is to recognise objects. Matching the indivisible pattern elements that comprise surrounding objects is needed. If there is an indivisible representation, an atom comprising the unique object’s patterns if you will, it must simply be matched. What pattern-atoms identify a square? What atoms comprise your friend, Sally? What atoms recognise your fourth grade school environment? What are the basic pattern elements necessary to recognise a wheelbarrow, archaic fonts or the mechanical workings of a door? This search for indivisible pattern-atoms has been seen in the computational quest to reduce sensory experience to single algorithmic outcomes since the 1950s. That approach seeks to represent visual objects and auditory words as single indivisible sets of time-independent values, stored in databases. The proposed alternative is a layered model in which pattern-atoms exist, but not at the layer that matches the stored instances. In language, for example, the set of elements comprising an object is anchored to a single, indivisible auditory pattern: a word. In fact both the visual object and the auditory word are proposed to represent sets of instances, as will be shown later.

Snapshots

The concept for robotic brain control is based first on recognition (finding and using sensor combinations comprising pattern-atoms) and second on moving using the correct combination of motion atoms. A snapshot pattern (or snapshot for short) is a set of active sensors, motor units or a set of active snapshots at a given point in time. Depending on the type of pattern stored, determined by its position (layer) within the system, a snapshot can be one instance of sensory attributes or the set of multi-sensory instances of objects. Snapshots are stored in sufficient detail to match or repeat the same pattern again. In conjunction with linksets, hierarchical snapshots allow for object identification. As an example, one hundred photographs of your friend Susan represents one hundred snapshot patterns.

Upon recognition of a previously stored snapshot pattern, a unique signal is sent both forwards to the next layer and backwards to the previous layer. Identifying a part of a snapshot is sufficient to identify the entire hierarchy. In other words, snapshots are bidirectional. Note that the snapshots stored in edge Patoms are key, being the only connection to the original experience.

Linksets

Snapshots are found at single points in time, but as objects are not experienced independently of time multiple snapshot patterns are necessary to represent physical objects. Objects seldom appear stationary in space, due to our own changing point of view and consequently, they rarely appear exactly the same way twice. To recognise objects, matching any one of the linked snapshots together is sufficient, removing the need for computation otherwise necessary to identify the atoms of mathematically idealised patterns.

The linked group of snapshots observed over time are called linkset patterns (or just linksets, for short). Like snapshots, elements of a linkset pattern are bi-directional. By automatically linking together snapshot patterns based on continuity (sequence), logical objects are formed therefore eliminating the need to identify reduced object atoms. Successful object recognition relies on correctly adding snapshots to existing linksets. The pattern-atoms that enable object identification are therefore collections of snapshot patterns joined into linksets over time. An example of a linkset is the set of box shapes caused due to a shift in the point of reference. Linksets are sequential patterns, logically represented as pointers to snapshot patterns.

Linkset patterns operate using the concept of intersection. By combining more than one linkset together, the weighted links will reduce the total active patterns to a subset in which the elements are fully consistent. There will typically be more than one consistent pattern, in which case the strongest links will determine the intersection’s best-fit.

Snapshot and linkset interaction

Snapshot patterns include not only the sensor patterns, but also links to other sensor patterns stored in higher layers. The difference between a snapshot and a linkset is that while snapshots are instantaneous patterns, linksets are the time-based, weighted collections of snapshot patterns. Snapshots rarely have much value without the time-based components. In language, for example, it is only at a specific point in time that a physical object is named. Linksets enable language by allowing a single association (a word) to be linked to numerous instances (snapshot patterns), effectively naming multiple instances in time with a single snapshot. While both the physical object and its name appear to be independent of time, the name may only be associated verbally for part of a second, despite hours of experience with the object. Yet our recall of the name for any of the object snapshots must be almost instantaneous. To stress the point, regardless of the shortness of object naming, immediate recall of this association is available through the linkset.

To illustrate the interaction, let’s walk through the sequence of events necessary to name a person you meet. Snapshots are established based on concurrency and linksets are established based on continuity. When introduced to Robert and hearing the sound “robert”, components of the images of Robert are stored as visual snapshots. These are grouped together in a linkset due to continuity. Elements of the sounds comprising the word “robert” are stored as snapshots, grouped together in an auditory linkset. Based on the concurrency of the two active linksets, at the layer above them, a snapshot connects the two sensory patterns. This connection effectively joins the two linksets. In other words, by association, the sounds comprising “robert” are joined with the visual images of Robert.

Linkset intersection

Linksets enable the concept of decision making through intersection. Intersection is the result of combining two or more linksets to establish the common ground between them. At a low level, intersection maintains consistency with the stored patterns involved. At a high level, intersection results in the most favourable or least unfavourable emotional outcome. The following example illustrates these concepts.

When viewing a known box, one of the previously stored snapshot patterns is matched. The snapshot signals the match to the layer above, which activates the linkset. The linkset signals forward and backward to its constituents including all the other snapshots comprising the box. These other snapshots are activated, in turn. One of the forward links may result in a link to the auditory linkset pattern for the word “box”, effectively naming the object. This forward and backward linking is bidirectional pattern matching since either pattern identifies the other, albeit indirectly.

Our brain needs to do more than simply name objects, of course. More typically the brain’s requirement is to direct motion based on numerous pattern inputs. Leaving aside the question of learning at this stage, let’s consider two examples in which a brain needs to take action based on previously established patterns.

If smell A, movement B, sight C and touch D happen simultaneously, take action E. In isolation to all other stored patterns, this allows for A and B only to activate action E. As linksets are bidirectional, the pattern E must identify A-D when receiving only one of the inputs. This bidirectional nature of linksets is invaluable to function in the world’s complex sensory environment.

If the sensory pattern group above results in E, but a similar one with A, B, C plus F leads to G, how does the brain select which action to take when presented with just A? The concept of pattern intersection comes into play. Linksets connect using activating and inhibiting weighted links, not binary switches. Provided the A link is slightly stronger than E, in isolation A leads to E (Fig. 1). In more complex intersections, the strongest links identified over a number of patterns dictate the action. If intersection fails to find a consistent pattern, no action is taken.

Fig. 1. Examples of bi-directional “best-fit” pattern matching: (a) Sense A, B, C and D link to action E (strength 5). (b) Sense A, B, C and F link to action G (strength 4). (c) A-only active leads to E due to strength 5 (5>4). If A and F were active, G would be the action (4+4>5). In this latter case, G still activates all of A, B, C and F in reverse.

In human examples when a decision cannot be reached, actions can be compelled with the addition of a time limit as the best fit can be selected by lowering the accuracy threshold. If the human brain uses this approach, full knowledge of all previous experiences in the relevant domain, including emotional patterns from all elements in the brain’s linksets, is necessary to predict a human decision. Even for simpler animals, like dogs and cats, predicting behaviour is theoretically complex.

Patoms

To control these complex theoretical constructs, the concept of the Patom is introduced. It is the basic building block for the robotic brain – a standard device enabling complex hierarchical and bidirectional pattern storage. Patoms are implemented in layers forming a hierarchy in which the lowest layers store sensor attribute patterns and the higher layers store complex, multi-sensory patterns. Despite the Patom itself being fairly simple in nature, the approach described rapidly results in complex capabilities.

A Patom (Pattern-matching ATOM) is defined as the smallest unit that stores, matches and uses patterns in a brain (Ball, 2000). The word Patom was selected since it sounds a bit like “pattern” and is a combination of the words pattern-matching and atom. The term Patom recognises the value to science from postulating the indivisible physical element (atom) upon which chemistry is built, an approach reignited from antiquity by John Dalton in the early 1800s (Gribbin, 2002). In cognitive science also, there is great value in finding reliable ways to convert complex experiences into groups of pattern-atoms.

In humans, there is evidence supporting the existence of such a construct. The observable effects of brain damage variations resulting in cell loss suggests that localised brain areas are collectively capable of capturing specific patterns. Damasio (1994), for example, describes effects resulting from losing the six layers of cortical cells. Greenfield (1997), McCrone (1999) and others popularise this understanding through the visual presentation of graphic brain scan techniques. The question of whether the brain is modular has moved to the question of how many modules there are in the brain (Jacobs, Jordan and Barto, 1991).

A Patom converts patterns (input signals) into links (output signals). Patoms are hierarchically structured with each Patom signalling uniquely (including a feedback signal) to correspond with the received patterns. By storing and linking snapshot patterns, the volume of data sent through the system is minimised. This effectively distributes pattern storage at the received location and not centrally in some form of processed abstraction. While higher-level Patoms are essential to multi-sensory patterns over long timeframes, only sensory patterns in edge Patoms interact directly with our body. Snapshots stored in edge Patoms contain sufficient detail to recognise the transduced sensory pattern, while the rest of the system simply supplies links. The contents of the links from the edge Patoms are incomprehensible without referencing back. Our brain’s architecture is structured with multiple layers of bidirectionally connected neural networks. While it may appear confusing when compared with a model in which senses are connected directly to motor control areas (Damasio, 1994), it aligns with the needs of a reasonably slow brain driven by evolutionary necessity to act rapidly and the proposed layered approach to control complex multi-sensory patterns.

Although the alignment between our brain’s topology and the theoretical Patom construct requires more research, current evidence is sufficiently detailed to propose the following functional specification: (1) A Patom automatically stores and matches patterns based on received inputs. The location, dictated by its topology of connected links, determines whether the patterns are single-sense or multiple-sense. (2) Matched patterns send a unique output, weighted by experience. This output directly links to other Patoms. (3) Patoms are hierarchically connected and operate bi-directionally based on concurrency and continuity. See Fig. 2 for a high-level diagram of interconnected Patoms.

Fig. 2. Simplified Patoms for hearing, vision and grammar. Apparent complexity stems from the great number of forward and backward connections to multi-use Patom elements. Patom 5’s learned patterns link back to sensory patterns as the system is bidirectional. If Patom 1 receives input X, it sends output Y. If Patom 2 receives input Y, it sends output Q, which includes feedback Z. If Patom 1 receives input Z, it sends output Y. In other words, a linked pattern works both forward and backward and compels experience to align with “what you expect” since near matches direct to previous experience, if consistent.

References

Ball, J. S. (2000). ABC Radio National, Ockham’s Razor, Our Brain, the Patom-Matcher http://www.abc.net.au/rn/science/ockham/stories/s73842.htm.

Damasio, A.R. (1994). Descartes’ Error New York: Picador, 26-27, 92-93.

Greenfield, S. (1997). The Human Brain: A Guided Tour, London: Phoenix, 18-28.

Greenfield, S. A. (1996). The Human Mind Explained, Sydney: Reader’s Digest, 12-13.

Gribbin, J. (2002). Science: A History 1543-2001, London: Allen Lane, 359-360.

Jacobs, R.A., Jordan, M.I. & Barto A.G. (1991). Task Decomposition Through Competition in a Modular Connectionist Architecture: The What and Where Vision Tasks. Cognitive Science 15, 219-220.

McCrone, J. (1999). Going Inside, London: Faber and Faber, 186 and Plate 10.