fast.ai: A Fresh take on Learning and Teaching ML
In this video Rachel Thomas provides an interesting take on learning ML: instead of promoting the typical bottom-up approach, fast.ai promotes a top-down approach. From a pedagogical perspective this seems counter intuitive. Surely you need to know the building blocks before you can move on to the theory that builds on the building blocks? Indeed, that is how traditional education proceeds. However, when consultants provide feedback to executives they tend to take a top-down approach. Why is that?
The main reason for taking a top-down approach when writing up/presenting technical findings is that you can provide a roadmap for where you are heading. This means that when you step into the details, the stakeholders can, because they now have a map of where you are heading, know how the details relate to the bigger picture. This is precisely why I think the fast.ai approach to learning ML can be effective. Rachel Thomas provides further motivation for their approach in their video: How to Learn Deep Learning (when you’re not a computer science PhD)
Inheritance
In this post we will look at how different types of inheritance can be translated to OWL. We consider the case where Person
is specialized by Employee
and Client
(Fig. 1). In a UML class diagram if inheritance is not annotated the default annotation {incomplete, disjoint}
is assumed. incomplete
means there are instances of Person
which are neither of type Employee
nor Client
. disjoint
means there is no instance of Person
that is both of type Employee
and of type Client
. The set representation is given in Fig. 2 and the OWL translation in Fig. 3.

Fig. 1

Fig. 2

Fig. 3
The annotation {complete, disjoint}
means every instance of Person
is either a instance of Employee
or an instance of Client
(Fig. 4). The corresponding Venn diagram is given in Fig. 5 and the OWL translation in Fig. 6.

Fig. 4

Fig. 5

Fig. 6
When overlapping
is used rather than disjoint
it means an instance of Person
may be both of type Employee
and of type Client
. Fig. 7 – 9 provides a UML class diagram, Venn diagram and OWL translation as example for the annotation {incomplete, overlapping}
. Fig. 10 – 12 provides a UML class diagram, Venn diagram and OWL translation as example for the annotation {complete, overlapping}
.

Fig. 7

Fig. 8

Fig. 9

Fig. 10

Fig. 11

Fig. 12
A Brief Introduction to Protégé and Reasoners
A question you rightfully may be pondering is: Why translate object oriented classes into OWL? The answer is that it can help you to find logical inconsistencies in your class designs. In this post I will introduce the tools that will eventually enable you to find logical inconsistencies in your class designs.
The tool we will use is called Protégé. Download and installations instructions for Protégé can be found at https://protegewiki.stanford.edu/wiki/Install_Protege5.
In this post I will provide two screencasts:
- In the first screencast I will show you how to enter the OWL representation of the
Person
class introduced in the previous post. - In the second screencast I will show you how to run a reasoner and how an inconsistency can arise.
On to the first screencast:
- Create a
Person
class. - Create the data properties.
name
surname
age
- Through sub-classing state that the
Person
class necessarily have aname
,surname
andage
.
- If we run the reasoner on this ontology, no inconsistencies will be found.
In the second screencast I show how an inconsistency can arise. The steps are as follows:
- Create an individual called
sarah
of typePerson
. - Run the reasoner. You will see the reasoner give no errors (nothing happened). This may come as a surprise to you since we have not set the
name
,surname
orage
data properties for the individual calledsarah
. In OWL this behaviour is expected due to what is called the open world assumption. OWL makes no assumption with regards to knowledge that is not stated explicitly. Since we did not state that thesarah
individual does not have, for example, aname
, the reasoner found no error in our ontology. This is different from typical database behaviour where absence of information is often assumed to indicate that the information does not exist, which is referred to as the closed world assumption. - Now let us change our
sarah
individual to state that it does not have aname
. This is achieved by stating that thesarah
individual is of typename max 0 xsd:string
. This states that thesarah
individual can have a maximum of 0name
data properties of typexsd:string
.
- If we run the reasoner now it shows that we have an inconsistency. We can ask Protégé to explain the inconsistency.
- The explanation states that
sarah
is of typePerson
and of typename max 0 xsd:string
. ButPerson
is a subclass ofname some xsd:string
. This states that individuals of typePerson
must have at least 1name
property of typexsd:string
. Hence, the reason for the inconsistency.
Admittedly this example is contrived: there is not much sense in creating a Person
class which we state must have a name
and then create an individual of type Person
which we then state does not have a name
. But this was done here to show you how to use a reasoner to find inconsistencies in your ontology and to show you what information you can expect when your ontology is inconsistent.