Platt Perspective on Business and Technology

Commoditizing the standardized, commoditizing the individually customized 12: neural nets and self-assembling systems

Posted in strategy and planning by Timothy Platt on July 7, 2013

This is my twelfth installment in a series on the changing nature of production and commoditization (see Business Strategy and Operations – 2, postings 363 and loosely following for Parts 1-11.) And my goal here is to discuss neural networks, and both biological systems as a starting point and artificial neural networks as a goal. And I plan on discussing them in the larger context of complex self-assembling and self-organizing systems in general, here as manufacturing and production systems goals. As a foretaste of a related posting to come, I add that I am also going to address this general topic area in a second series: Some Thought Concerning a Rapidly Emerging Internet of Things, where I will discuss the functionality of these systems. (See Ubiquitous Computing and Communications – everywhere all the time 2, postings 211 and loosely following for that series, where my goal as of this writing is to address this topic area there in its Part 10 installment.) But I focus here on the production – from the outside and through self-assembly processes of these complex systems here, as production of neural net-organized devices.

Artificial neural networks are often thought of in the abstract and as mathematical and mathematical logic-based conceptualizations. And in keeping with the biological neural networks that they simulate and build from, they are often formulated conceptually as being constructed from arrays of artificial neurons with all network connections and nodes formed from them. I would at least start this discussion by taking a more abstracted and open-ended perspective as to what functional nodes and connectors would look like and be built from. My goal in that is to at least attempt to limit obligatory features of biological neurons per se that might or might not always make sense when:

• Specifying,
• Designing or
• Building an artificial neural network, or
• Setting up a starter core system that could be grown from, and with outside assembly, self-assembly processes or both managing that process.

And with that, I begin with what might be considered to be a conclusion.

• Envisioning and building artificial neural and other complex self-assembly systems should be seen as an evolutionary process in which specific biological systems features and details are identified and abstracted so their functional qualities can be achieved through other means as necessary.
• So my goal here is to at least begin a discussion of what the overall structural and functional goals are for what these systems can and will do, and even in contexts quite different than found in biological systems per se.
• And with that in mind, I begin with observable biological neural systems and with a goal of more abstractly outlining something of what their individual components do with their requirements and constraints, and about how they functionally connect together in forming at least low-level simple systems subassemblies.

Biological neurons are highly specialized cells that produce and convey electrochemical signals down their length and through specialized intercellular junctions called synapses. Simplifying this discussion for purposes of this posting, and leaving out a great deal of terminology in that process, neurons connect with sensors and convey and process input from them. They connect with other neurons and in a variety of ways, forming information processing systems. And they connect through specialized intercellular junctions to muscles where they cause contractions and where they control muscle activity, and to glands and other effectors. Combinations of sensory input receiving, intermediate information processing and effector activating neurons come together to create feedback driven systems.

To put the systems complexity that can arise in that into scale, and citing an average healthy adult human brain as a benchmark, it is usually estimated that that brain contains in it some 100 billion neurons, with each on average carrying some 7000 synaptic connections to other neurons. Some of these synaptic connections convey excitatory signals that provoke signal transmission and positive activity to adjacent, functionally connected neurons and some synapses provide inhibitory signals that serve to damp down that activity, and together, patterns of excitatory and inhibitory signals serve to produce controlled, organized functional activity that can be turned on and then dampened down and turned off as needed.

It is sometimes stated as if a fact, that we only use a small percentage of our brain cells. Neurons that do not functionally connect and actually function are not maintained. They die off and are lost. If there is any truth to that particular “fact” it is that all of our neurons are not all wildly firing signals at once and in all possible directions; we do in fact use all of them, but only when and as they, and the functionalities that they control are needed. So at any given time, some areas of our brains show greater, and others show lesser levels of activity, and at the individual cellular level and also when observed by technologies such as positron emission tomography (PET) where overall energy utilization can be determined for specific areas of the brain, as a measure of overall levels of activity in the cells of those brain regions.

But as I stated above, my goal here is not to focus on individual biological neurons or on small-scale assemblies of them. It is to focus on the properties that they have in general and even as abstractions, and on the abstracted level discussion of how they functionally interact as connectors and nodes in general. So I turn here to consider some general properties, coming out of what I have just written.

• Individual neurons connect out very widely to other neurons in assembling incredibly complex circuits that in effect organize on the fly and then functionally shut down.
• Neuron to neuron connections can either facilitate and positively contribute to larger scale neural processing activity, through contribution of excitatory signals or they can contribute to dampening down or blocking specific activity through contribution of inhibitory signals – and the overall emergent result is functional balance and control where specific functional activity ramps up and is maintained, and then ramps down and it shut down as needed.
• Bringing in self-assembly here, new connections can and do form, in building new preferred neural pathways that would support recurring brain activity. Repeatedly activated and used synapses are reinforced through accretion of specific proteins at them, that make them more functionally robust. And unused connections – unused synapses can and do die off, as do functionally disconnected neurons. Even in older mature brains, neural stem cells are still present and new neurons are still formed, if at a much slower rate than would be found in a still actively growing and developing younger brain.
• The brain overall can readily be seen to follow a large scale functional organization with a seeming myriad of specifically functional areas, and with heavily parallel processing within them. But at the same time, this same brain is capable of exhibiting a great deal of functional and organizational plasticity in developing work-arounds for example in recovery from traumatic injury such as stroke. And when someone learns to play a musical instrument to a professional level of competency, for example, and particularly if they learn that from childhood, PET scans and other mapping techniques show certain functional areas of these musicians’ brains expand to facilitate this skill. Similar growth and expansion to facilitate recurring functional use can be observed in highly skilled chess players too, as a second example of this type of resource allocation and reallocation process. So neurons and the larger neuronal systems of entire brains show higher level pre-established functional organization but they also show capacity to flexibly rewire and reorganize in a self-organizing way so as to meet current and recurring, and new and emergent need.

One of the ongoing trends in production and manufacturing has been the organized assembly of products on a progressively larger and more complex scale.

• This means producing more and more copies of the same product items at a faster and faster rate, and with progressively lower and lower production failure rates – with fewer and fewer defective copies made.
• This also means production of progressively more complex and functionally flexible individual products and that point is crucial here.

Consider the central processing unit (CPU) chip of a standard and even basic computer and how the number of transistors, as functional elements in it have doubled roughly every 18 months in ongoing accordance with Moore’s law, and how this trend has persisted since the invention of the integrated circuit in 1958.

• The cost of design and production of these computer chips and of the development, construction and maintenance of facilities required to produce them rises as fast as or faster than does the complexity of the chips that these manufacturing facilities produce.
• And increased chip complexity and fineness of structural/functional detail in them from inclusion of progressively smaller-scale circuit elements means greater costs from chip defect rates, and both from production of flawed chips that have to be discarded, and from more costly processes for limiting product production failure rates in the first place.

One possible approach to breaking out of the cost spiral for producing progressively more complex integrated circuit chips such as CPU’s would be to build in capacity to reroute and reestablish functionality around point defects in the circuit patterns produced, as is achieved in neural networks. Another would be to find ways to enable self-organization and functional self-assembly as neural systems display when specific patterns of neurons fire to meet specific time-sensitive needs, to disorganize from that coordinated activity as its need falls away. Such a circuit might show overall fixed functional compartmentalization as is found in a brain, in in fact within its individual functional partitions, but it would also show functional plasticity too, and capacity to adapt.

As a first and still primitive step in achieving that type of goal I would cite gate array and field-programmable gate array integrated circuits. But these are and can only be seen as first steps that would support development as is, of only simple systems. Among other things, these technologies are not readily reprogrammable and certainly not on-the-fly during operation and with a brief refractory period before that can be done between setups. My point here is that “artificial neuron based systems” might not be constructed out of elements that look in any way like neurons. Functionally homology (with this link pointing to biological homology) and homologous structures are more important here.

I leave off from this discussion here, and with the final thought that there is a great deal of market-driven and competition-driven pressure to develop new approaches to expanding both the scope and the functional flexibility of these systems and these circuit development options for creating them. And this drive to progressively greater complexity in both structure and function, and at limited and competitive cost per unit is all but pervasive and across most if not all industries.

I am going to turn back to the topic of 3-D printers in my next series installment where I will discuss its emerging use in medicine and surgery – an application where customized 3-D printed constructs that serve as scaffolding and a patient’s own stem cells will open up whole new avenues for personalized treatment. Meanwhile, you can find this and related postings at Business Strategy and Operations and its Part 2 continuation page.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: