Reading What Machines “Think” - Brain Informatics - page 162

The applicative perspective is also an extremely interesting and unexplored
area of research. Using the ideas developed on the brain side of the parallelism
(Fig. 1), we can try to apply them to the electronic computer side. Can we
develop technologies that “ read the computer mind ”. This predictive model can
have a wide variety of applications, e.g., detecting malicious software, detecting
the intentions of hostile computers by looking at their activation patterns. We
need specific devices that can capture activation images of computers. We can
then study the application of machine learning to induce models that can predict
what a computer is doing by analyzing its activation patterns.
The complete realization of the electronic computer side of the vision is a long
term goal. It requires physical devices to capture the activation state of electronic
computers. Yet, as electronic computers are easily and directly observable, it is
possible to set up a scenario where we can test the idea. This scenario can
help in preparing the ground of the complete research program. This first phase
of analysis is the virtual observation of electronic computers . We exploit the
fact that we can directly observe the memory state of machines and, then, we
can draw their activation state. As the observation of the activation state is
done through a software system instead of a physical device, we call it virtual
observation . We will describe this scenario in Sec. 4.

3 Background

Categorization is the cognitive ability of classifying objects into concepts. This
ability is extremely important for this study as, on both the brain side and the
electronic computer side, we want to study the correlation between activation
images (i.e., objects) and cognitive processes (i.e., concepts). The final aim is to
develop models that determine the performed cognitive process by observing an
activation image. For example, we want to have a model that determines that
the brain in Fig. 1 is performing the act of looking at a chair. This should be
done only by observing the brain image.
One of the objectives of machine learning is to define models and algorithms
that can learn categorization functions from existing training data. Observing
some brain images grouped into classes, i.e., grouped according to the cognitive
process, machine learning algorithms induce classifiers that can predict the class
for a new and unseen brain image. A classification function

C is defined as:

C : I → T (1)

where is the set of possible categories. This clas-

sification function will observe objects

I is an instance space and T

.Thecate-
gorization is possible if some regularities appear in the space of the instances

i ∈ I assigning a class t ∈ T

.
To discover these regularities, we need to observe instances using some feature.
These instances are then represented as points in feature spaces

I

F 1 × ... × F n

where each F i is an observable feature. We can then define a function F that

maps instances i in I to points in the feature space, i.e.