Reading What Machines “Think” - Brain Informatics - page 168

Table 1. Classification accuracy over the different experimental settings

Task Smoothing DecTree NaiveBayes
algo no 80.37 64.44

lang no 98.89 99.25

algo yes 81.48 62.96

lang

yes 98.89 99.62

The per-algorithm task is our final task, i.e., understanding whether or not it
is possible to determine the “ cognitive task ” performed by the machine. The
per-language task is instead a control test. We want to see if it is possible to
determine the “ cognitive substrate ”wherethe“ cognitive task ” is performed. As
this task seems to be easier and it is similar to the per-algorithm task ,wewant
also to see if it is solvable with the feature space for images we are using.
For the two classification tasks, we prepared two different settings according
to the smoothing applied to the final images:

- no-smoothing , images are kept with the highest quality available;

- smoothed , images are smoothed according to the equation (4).

We need these two settings to determine if the performance of the two classifi-
cation tasks is affected by a degradation of the images. As we already discussed,
the degradation approximates the real operational conditions where the captured
image cannot have the quality of the single byte of memory.
Finally, as machine learning algorithms we used a decision tree learner [8]

( ). We

selected these two types of algorithms because they behave completely differ-
ently. The decision tree learning algorithm recursively selects features. At each
step, the one selected is the best discriminatory one. The pruning done in the J48
algorithm also performs a feature selection. Yet, the naive bayes algorithm uses
and weights all the features in the probabilistic model. These two algorithms
then give the possibility of analyzing performances in two completely different
setting. We used the implementation given in [18].

DecT ree ) and a probabilistic model, i.e., the naive bayes ( NaiveBayes

5.2 Results

The results of the experiments are reported in Tab. 1. The first column describes
the tasks that have been analyzed. The second column reports if the smoothing
has been applied. The third column reports the accuracy obtained using the
decision tree learning algorithm. Finally, the third column reports the accuracy
obtained using a naive bayes probabilistic learning algorithm.

5.3 Discussion

As expected, the task of deciding the language of the process is simple and
solvable. Accuracies are extremely high. There is no difference in performance