³⁰I don’t go into any details on how that “division” of the world into categories happens, but for the interested reader, I give some pointers here. Earlier, we considered the case where the neural network recognizes an object and outputs its category. This is a simple starting point; while it can be easily done by supervised learning, it can also be implemented by unsupervised learning methods, in particular methods such as clustering and (Gaussian) mixture modelling. In the case of humans, the connection between neural networks and logic-symbolic processing is related to what is called the symbol grounding problem (Harnad, 1990). It is a topic subject to a lot of debate: some argue no proposed solution is sufficient (Taddeo and Floridi, 2005), while others argue it is essential to consider robots which communicate with each other (Steels, 2008). The operation of neural networks is closely related to one well-known proposal called the prototype theory. It means we define each category by a single point in the space the activities of units in a neural network (preferably in layers close to output); this point is the prototype (Rosch, 1978). Basically, you would find a “prototypical” cat as a point in the very center of all those points that represent cats. A generalization of this idea can be found in Gärdenfors (2004). However, things get much more complicated in the case of abstract categories such as “good” or “beautiful”.