Keynote@FOIS'24 - Ontologies for Machine Learning - Frank van Harmelen

Intro
AI seems to be a tale of two towers: the Statistical AI tower vs. the Symbolic AI tower
It seems more like a religion, "you don't want to read the other's books because they are wrong anyway".

But he wants to put these two towers together.

It is important to note that ML work is extremely unaware of ontology work. They are now learning about simple things we already know (e.g., about types)

In this talk, he is going to give examples from the literature where ontologies are solving problems in ML contexts.

Hard problems in ML
ML is not:
- Explainable
- Generalisable: Generalisation is good only for the data similar to the one for which they train
And ML is:
-Data Hungry: Data sparcity - data acquisition bottleneck
- Black box
- (they have a) Performance Cliff

He uses a graphical notation to talk about ML systems

Synmbol or data

Conditions to be a synbol according to What the -0.70, 1.17, 0.99, 1.07- is a Symbol?, Istvan Berkeley, Minds and Machines, 2008
1. A symbol must designate an object, a class or a relation in the world (= the interpretation of a symbol)
2. symbol can either be atomic or complex (=composed of other symbols according to compositional rules)
3. there must be a system of operations that, when applied to a symbol, generates another sybol that again could have a designation.

Examples

Example 1:
He gives an example to explain how ontology can help solve the problem of semantic loss-function. How can a ML understand the difference between a flower and a cushion with flowers in it. The system should see that if this is attached to a chair and that cushions may have flowers in it, the probability that the object is a cushion is higher than a flower. And if the ontology shows the relation between cushion and chair and the ML recognizes the chair, then they know that the object is a cushion.

Example 2:
Informed ML using KR. An ontology can give the background knowledge you need to guide the final neural network (NN) layer. Example of the photo of Queen Elizabeth in which the system gave a 99% answer saying the object in her head was a showercap. But if they know what queens usually wear in public, then the NN would know that the object is a crown. The ontology helps to provide justification/explanation for the response.
In this case, the ontology is not involved in the prediction, but afterwards in the justification.
There is a mistake now on how the ML communicty looks for explanations (by undesrstanding how the NN arrived at the result), because people will not be interested in understanding how the NN arrived at the result, but they require a plausible explanation for why that result is correct.

Example 3:
Learning intermediate abstract for learning. He shows an example in wich the authors proposes an architecture composed of two agents: a Neural Back End and a Symbolic Backend. Comparing the performance of the two, the symbolic agent did much better because by the time the neural agent learned about the world, then the world had already changed.

Example 4: he gave an example of early cancer diagnosis. In this example, ontology provides intermediate abstraction. Try to learn early symptoms of CRC from GP data (symbol, life style, diagnosis, drugs)

1. Raw data: no meaningful signal
2. Raw data + snomed = abstracted data: significant results.
Predictive modeling of colorectal cancer using a dedicated pre-processing pipeline on routine electronic medical records, R. Kop et al.

About LLMs...
LMMs Modulo
LLMs can't calculate, but they can learn when to call a calculator.
LLMs can't do formal ontological derivations, but they can learn when to call an reasoner.

The Vienna Study: looked in 300+ semantic web systems to understand the role of the ontology in these studies
- Clean the data
- Abstract the data
- Explain the results
- Guide the search
- etc.

Open Questions
We know that domain ontology can help ML, how about foundational ontologies?
- invariance to time, space, roles
- improve out-of-distribution generalization

What ontological representation languages are suitable to interfact with ML
- differentiate vs. discrete?
- classes vs. prototypes

Can ontologies help to resolve some of the inherent weakness of LLMs?
- halluncinamtion
- lack of planning, inference.

We have early empirical evidence that ontology knowledge can help ML, but no theory which types of ML will/won't benefit, for which functionalities?

In the discussion, Nicola makes a point that theory may be started by looking for principles to construct such theory. For example, perhaps understanding the difference between justification and explanation.
Someone in the audience made a point that even wrong ontology may help, but then I said that little ontology may help, but the wrong ontology may hurt. Frank said these are claims that we must experiment.
Data will always underrepresent the world. We deal by this by common sense reaoning. However, we have made little progress in that, because we've been obsessing about the symbol when we should actually be worried about the concepts. Analogical reasoning can play a role here to apply structures in similar contexts. Frank says that ML can do a lot for Common sense reasoning, because they are trained in huge corpora and they can lead to the extraction of common sense knowledge. He says that he would love to see analogical reasoning come back to being fashionable.