3.4.4 Perceptual vs Gestalt Semantics

One of the questions raised by the experiment described in 3.4.3.2 is whether or not the semantic input representation used for CLASPnet might not -- accidentally -- contain a few units which give away too much information. Given the kind of loose representation used, with its division in perceptual, Gestalt and instinctive units, it is not inconceivable that, for example, the Repetition input unit had effectively been coded in such a way as to provide sufficient information to the network about, for example, the Infinity output unit.

In order to investigate this conjecture, I first looked at the connection weights linking the 25 semantic input units (X-axis in Figure 3.27) to the 15 units of the semantic hidden layer (Y-axis). Figure 3.27 provides a graphical representation of the weights: white squares indicate inhibitory connections and black squares excitatory ones, with the size of the squares indicating their strength. Although it is obvious that some units have a higher number of strong connections than others, most of them carry at least one significant weight (Note 33). If we compare the first 14 columns with weights leading from the perceptual units, to the other 11 columns (7 Gestalt + 4 instinctive), then it appears that the latter group has a stronger overall influence. On the basis of this insight, I generated new training and test corpora of 3,000 sentences each (maximum length of 20 words), and then zeroed out the first 14 semantic units in one copy, and the last 11 in another. The two nets were then trained to see which one would learn the task best (Note 34). The results on the test corpora are shown in Figure 3.28.

As in the previous figures, the leftmost bar provides a reference marker. But in this case, the 'Normal' bars were derived from a test corpus other than the one used to generate the results for the other two bars, so care should be taken with direct comparisons. Of more interest, however, is the relation between the two partial semantic networks. Even a cursory glance suffices to show that the network which could rely on information from the Gestalt and the instinctive units performed better than the one which only had access to the perceptual units. For most bars, however, the difference is small, so one can also learn from Figure 3.28 that the perceptual units do contain useful information themselves. (The large difference for the missed clausal type units is mainly the result of one aberrant score: the network scored 60% missed errors for the Yes/No-Question unit versus a much more usual 32% for its Gestalt counterpart.)

Note 33
This type of superficial analysis is of course severely limited: a combination of many weak connections can, if connected appropriately, achieve a much larger effect than a single strong weight.

Note 34
Comparing the relevance of the semantic units by switching them off one by one is not an option unless one has access to very fast computers indeed.