Understanding the Brain p. 2 Josephson
2. Basic scheme
We assume in the first place that the nervous system can be characterised as a structured hierarchy whose
elements are of well defined types, each type being associated with specific types of behaviour. To make this
characterisation one that can be precise rather than merely qualitative, we assume also that each type has
associated with it a specific abstract or formal model, in terms of which the behaviour associated with an
element of that type can be understood, these assumptions being merely the translation into the context at hand
of how design can viewed in general, expressed in formal terms. In the case of the nervous system it is
normally not considered that such precise accounts are possible, but our hypothesis that in fact, given sufficient
insight into the question of what models might be appropriate in this context (as will be discussed later), the
kind of scheme proposed does apply. It is relevant to note that a structure of the kind proposed, involving a
hierarchy of systems of specific types, has been found appropriate for designing computer programs required to
function reliably in complicated situations, the typed systems in this case being the objects of object-oriented
programming.
To apply this scheme to the actual brain, we identify the typed abstractions of the scheme with processes that
can be regarded, to a first approximation, as functionally separate and approximately autonomous, such as those
of balance and hand-eye coordination. The physical systems to which the specialised models refer involve the
neural circuits relevant to the specified type of activity together with the environment in which they function.
Intuitively, there is a specific (and specifiable) mechanism for balance, a specific mechanism for hand-eye
coordination, and so on, and this is what the relevant abstract model refers to. As far as the neural-circuitry
component of the model is concerned, in many cases this need not be a model in the form of a neural network,
but instead a signal processing model involving an appropriate mathematical transformation between input and
output signals. In the case of learning, in many cases the model can simply dictate that a consistent
relationship between input and output signals, determined by trial and error to be appropriate for the successful
performance of some task, be learnt , meaning that the system, after learning has taken place, applies the
relevant functional relationship directly rather than having to determine the outcome by trial and error.
Each model, to be complete, needs to contain parameters that may vary, in order to take account of the variation
of the details of the execution of a given task from one context to another. In general, therefore, there will be
many systems for each model. As with the objects of object-oriented programming, the models need to
describe not just a single process but in general a complex of processes, taking into account the way they may
interfere with each other, as well as taking into account the mechanisms of learning.
3. Hierarchical aspects and hyperstructures
The thinking behind what has been proposed is that the nervous system, while very complex in itself is, from a
logical point of view, constructed from comparatively simple components, and works in the way that it does as
a consequence of the laws governing the behaviour of these components, which laws are in principle accessible
to separate investigation and analysis. This somewhat cautious statement of the logical dependences involved
anticipates our use of the hyperstructure concept of Baas (1994), which we now discuss. Normally one infers
directly from behaviour at one level of a hierarchy to the next level up. Baas characterises this as deducible
emergence. He defines in addition a process that he calls observational emergence, where a range of
configurations are explored, with the aid of a specialised observer mechanism, in a search for one exhibiting
some specified target behaviour. If and when the target behaviour is attained, this behaviour is learnt, in the
manner previously discussed, i.e. new connections become established which permit immediate computation
equivalent to the behaviour discovered earlier by trial and error. If the links concerned take time to become fully
established, the computation concerned can be further tested, with the links fully established only if the tests are
successful. In the context of development, a mechanism such as described does not need to know precisely
what to do to obtain a target result, but can take advantage instead of knowing what field to explore in order to