Patom Theory Explains Language Acquisition: Part 1

Pattern Atoms (Patoms) solve many NLU puzzles today

Last time, we explored how meaning matching removes ambiguity by eliminating parts of speech. Today, we turn our focus from what doesn’t work, to what is working today, exploiting my Patom brain theory.

As usual, you can view the companion video with animations and description on our YouTube page at https://youtu.be/9eEsnRa0IAQ.

Patom theory postulates that all a brain can do is store, match and use patterns. The assumption is that a pattern cannot be divided beyond a certain point, and as it is indivisible, I call it an atom. That’s why we have the name pattern atom, or patom.

We will look at what the theory is claiming and then we introduce how it applies to language learning.

This is the core design of our natural language understanding engine, and so far, we have found no limitations in converting words to meaning and back again.

The 1990 Patom Model

Patom theory is the model the meaning matcher adheres to. A brain in this theory comprises the entire nervous system, from the senses themselves (like vision, auditory, olfactory, gustatory, touch, balance) to the output mechanism in muscles.

In the 1990s, the model was described[i] with an example as seen in figure 1. The main features are the edge patoms that connect transducers, or senses, to the brain. These connect to single-sense patoms that refine the recognition of that particular sense, such as in human brains the specialized recognition of motion and color. These single-sense patoms then connect to multisensory patoms that enable the multisensory experience to be consolidated.

Figure 1. A 2-sense brain model from the 1990s.

A key feature of the system is its bidirectional nature. A pattern that connects to another can find the patterns that comprise it. There is no tv screen in a brain, so instead the sensory experience that initiated a higher-level pattern can be accessed in reverse, using the bidirectional links. This certainly makes for a confusing model to reverse engineer.

A Dog’s Brain

My dog recognizes me. She is only 7 months old, but it is obvious that when she sees me she gets excited. The interesting observation is that when it is dark and she can’t tell who I am visually, she gets scared and barks, but when I call out she instantly recognizes me and stops.

As we know that a brain has specialised areas, even a dog has a multisensory capability to recognize objects, such as specific people and their voices in this example.

The system exploits the well-known expression, “neurons that fire together wire together[ii].” While Patom theory considers what a brain does, and not how it does it, the concept that 2 patterns can connect when matched at the same time is compelling to use as a tool to automate and control learning.

In the case of my dog, each sense must recognize me separately and jointly to allow each to direct her reaction to me. This capacity, shared by many animals, is modelled by the theory, but we don’t attempt to mimic it yet, due to the lack of sensory equipment and motor/muscle control equipment available today. Instead, we take advantage of the pattern atoms that represent objects (referents) and their relations (predicates) to connect it to human-only capabilities, such as language.

Introducing the Set (Snapshot)

Figure 2. A snapshot pattern is a set of matching elements.

A set is the first kind of patom. Here, many versions of the same thing are connected based on their associated meaning. While all of the symbols are of the letter ‘c’, each uses a combination that includes font, size, and case.

The set is like a template, except instead of needing a single, ideal, mathematical description, the input patterns to the set can be of any type. Any input to the set determines if it matches.

And introducing the sequence

The sequence takes elements of a set and orders them. These same elements are used to learn and use meaning in language, but at this point we are just looking at the basic concept while describing a model like that of formal linguistics — meaningless lists of letters that look like words.

Figure 3. A sequential pattern orders other patterns.

As each element in the list is a reference to another pattern (snapshot or list), we show them as ‘x’es in the diagram.

The next diagram shows the challenge in looking at a system implemented with Patom theory. As the patterns that relate to sense are found in the edge patoms, the intermediate patoms only refer to higher and lower levels.

Figure 4. How do we reverse-engineer this? Nothing is labelled in a running system!

Once the labels are removed as in the right diagram, the patom elements require reverse engineering to see the connections of the larger pattern.

AI Foundations

AI was launched in 1956 at the Dartmouth conference by John McCarthy. The late professor Marvin Minsky, who I corresponded with in the early days of my research, was a key contributor and influencer in AI until recently. He demonstrated his colleague, Oliver Selfridge’s illustration[iii]. It shows the need for AI to operate at more than “one level” (something that computer scientists can struggle with as computers ‘bake-in’ encoding of letters).

Figure 5. Selfridge shows us that recognition should be done over layers, not in isolation.

Here, the meaning of the funny H, or is it a cut-off A needs to be determined. By simply checking which sequences of letters are stored, only ‘the’ and ‘cat’ will be matched and used. ‘tae’ and ‘cht’ aren’t English words.

The solution to this puzzle by using known letter sequences is the same approach used to solve ambiguity in words, phrases and meanings in our NLU’s meaning matcher. As some like to say, a brain only uses one algorithm so, as I say, NLU only needs one memory strategy — store, match and use.

Applying the theory to Language Learning

The introduction to language learning can now start. Here we are given 2 types of meanings: one refers to something (a referent) and the other relates things (a predicate). Contiguity is sufficient to connect the two very different kinds of elements, auditory input that represents the sound of a word or visual input that represents the spelling of the word with the relevant meaning representation.

Figure 6. Referents are related by predicates, as arguments.

In the referent diagram, the sound and the written word both create bidirectional links with the active element, a referent. A referent needs at least a category, and it can have other relations like possessions and modifiers like qualities.

Figure 7. Predicates relate referents. Referent and predicate are semantic (meaning-based) terms.

In the predicate diagram, the relations for the predicate must connect to their arguments, however the predicate itself also needs the bidirectional links connected.

In both the cases, language learning of words simply connects 2 active elements together. That’s the first step in language learning: knowing what a word means. But of course we don’t talk with single words, but with phrases that can use multiple embeddings. That’s the next step in the language learning process which relies on the same mechanisms.

So there you have it, the introduction to Patom theory and I hope some food for thought.

The brain doesn’t need to look anything like a digital computer and our experience shows that for human language emulation, our model removes the complexity a programmer would need to handle.

(Next time we will flip our view from this brain-based model, to one starting with the science of language, linguistics, and the amazing theory based on the study of the world’s diverse languages, Role and Reference Grammar.)

[i] John Ball, Machine Intelligence: The Death of Artificial Intelligence, Hired Pen Publishing, 2016.

[ii] Donald Hebb, The Organization of Behavior: A Neuropsychological Theory. Wiley, New York, 1949.

[iii] Marvin Minsky, The Society of Mind, Simon and Schuster, New York, 1985, P209.