Refining Connectivism

Do all networks learn?

The notion that all networks learn, just by virtue of being networks seems off. It does not correspond to what most of us would call learning.

Is there some way to look at this, which makes more intuitive sense? Admittedly, there will be grey areas, but my view is that it is possible to conceptualise learning more clearly.

Different networks vary on:

• The breath of signals they are responsive to,
• The intensity of signals they are responsive to,
• How varied their response is,
• How much directionality there is to the process,
• Possibly others (still thinking on it).

The hypothesis is that, what we ordinarily call a learning system maps to:

High-complexity systems that respond to a wide variety of input signals in a wide variety of ways, and change in a way that makes future reactions appropriate (which is left to intuition for the moment).

Usually, networks of complex systems or of hybrid systems also change complex ways and will also be learning systems.

On the other hand, low-complexity systems change in narrow and even predictable ways whichare not considered learning systems.

In a sense learning is not the mechanism, but it's just the label. Complex systems that change directionally and in various ways learn; simple systems that change --by the same mechanism-- but do not behave in complex ways are not learning systems.

Just reinforce it: the proposal is that learning mainly points to high-complexity, but the actual kind of change could be the same in both systems.

We could say that there is a threshold of a binary function $f(Complexity,Degre{e}_{A}dapt,Store{s}_{C}hanges)$ were systems are considered at learners. There will also be, certainly, grey areas.

With that, which one learns?

We can apply the previous criteria to a large number of cases, and see whether we are satisfied:

Do biological networks learn?

It is generally agreed that learning involves changes in the wiring pattern, or degree of wiring, and this change can happen in different ways, and it is built into the system (brain) e.g. by Hebbian learning. We can take this case as our paradigmatic learning system.

Do artificial neural networks (ANNs) learn?

During training phase, connections (weights) are updated towards a goal through backpropagation, and the network behaviour (predictions) change. General-purpose process a wide variety of signals are are closer to learning systems than narrow AIs. Both only learn during training phase.

Does a colony of insects learn?

Complex organisms communicate and change their behaviour: Ants propagate pheromones which adapt their behaviour; humans similarly can do it by imitation of some kind. This "node update" propagates out, and may return to update the node further. Finally the network is seen as reacting appropriately.

But is this "reaction", or emergent behaviour, the same as learning? Only if the change is permanent. The variety of signals they process and the behaviours is quite varied and certainly are learning systems.

Does the world wide web, or a library learn?

The system will learn only if it is designed to do so. For example, after a certain event (like a query) a program may change the system to respond more appropriately in the future.

When humans have designed them to adapt (say with software) these have limited learning ability and change in response to a narrow variety of events, and are not quite learning systems. Considered without the programs, these wouldn't have any learning at all.

Hybrid systems such as humans-web, or humans-library systems certainly are learning systems.

The structure, interface, reliability that they have though, matters a lot in terms of how usable they are to humans, and so forth. These properties are still important across most network systems, be it learning networks or not. In other words, all non-learning networks can become learning networks when associated with other systems that can change them

Does the electric grid learns? Same as above.

Does a molecule or a material, which can be represented as a graph, learn?

Signals can change them. But even if we define appropriate as increasing the likelihood to persist / survive, the ways in which it reacts and the number of stimuli it responds to are narrow, and won't be learning systems.

Do mixes of nodes made of proved-learning-systems (say humans) and non-learning systems (say a book), learn?

Yes, especially if we consider a company, a community and so forth.

So all the different kinds may be considered to learning, if they permanently change in response to a stimuli and this adequately changes future responses.