Reading What Machines “Think” - Brain Informatics - page 163

f 1 , ..., f n ) (2)
This model is generally called feature-value vector and underlies many algo-
rithms.
The field of machine learning has delivered a wide range of algorithms to
analyze huge amounts of data in order to find regularities. Supervised (e.g.,
[8,9]), semi-supervised (e.g., [10]), and weakly supervised (e.g., [11]) algorithms
and models are available to automatically learn classifiers or decision making
systems. These models are widely and successfully applied in many important
fields, e.g., homeland security (e.g. [12]), data mining for business intelligence
(e.g. [13]), and computational linguistics [14].
Machine learning algorithms have been used to discover regularities in images
of brains performing cognitive and semantic tasks. The work in [7] follows the
idea that it is possible to discover regularities in brain images of individuals
observing or thinking of objects in the same semantic class such as chairs, houses,
etc. (e.g., [6]). Machine learning has been applied to induce brain activation
patterns for words where the activation image is not observed. Words with similar
meaning should have similar activation patterns. Using corpus linguistics, word
similarity is determined comparing their distributional meaning , i.e., their vectors
of co-occurring words. This is the distributional way of determining the meaning
of a word [15]. The induced activation patterns have high predictive performance.

F ( i )=(

4 Virtual Observation of Computational Machines

Electronic computers have a very nice property with respect to our research
program. The activity of these machines is observable using software programs.
Then, we can simulate the electronic computer side of our vision without actually
having a physical device to observe the activation state of machines. We can write
a software program that snapshots the memory of the machine. These snapshots
can then be used to produce activation images as if these were taken from an
external device.
Using these virtual observations of the activation states, we can test the over-
all process of the electronic computer side . Then, we can study if it is possible
to derive a correlation between the images of the activation states and the per-
formed “ cognitive processes ”. For this purpose, we will extract features from
activation images to feed machine learning algorithms. Given a set of train-
ing examples, i.e., training activation states, associated with different types of
“ cognitive activities ”, the machine learning algorithm can extract prototypical
models of activation for these types of cognitive activities. These latter models
can be used to classify novel activation states, i.e., recognize the type of cognitive
process that the activation state suggests. If classifiers have good performances
with respect to a set of testing activation states, we can conclude that the task
of reading “ machines' thoughts ” is reachable using the proposed features. Fi-
nally, we can repeat the process using smoothed images of the activation states.
Smoothed images better approximate the images produced with physical obser-
vation devices. Then, we can determine if the final vision is viable.