✏

The Far Side of AI:

Neurosymbolic AI

Introduction:

Artiﬁcial Intelligence( AI ) has been witnessing a monumental growth in bridging the

gap between the capabilities of humans and machine.

Building thinking machines have been a human obsession for decades. Despite, the

current advances in AI and ML, concerns about trust, safety, interpretability and

accountability of AI were raised by inﬂuential thinkers. In this article we will discover

the next wave of AI systems by introducing the Neuro-Symbolic AI approach.

Why deep learning is not enough?

In recent years Artiﬁcial Intelligence, speciﬁcally Deep Learning has gained popularity

thanks to the availability of computational power and the accessibility to large amount

of data.

Current advances in DL have achieved unprecedented impact across research

communities and industries. However, deep learning models are black boxes,

The Far Side of AI Neurosymbolic AI1meaning that we don't know what's happening inside which can lead to a lack of

interpretability of the results.

Deep learning algorithms are week when it comes to logical questions and researchers

believe that for AI to advance and think like human brain, it must understand not only

the 'what' but also the 'why' which can gave a clarity about the cause-effect

relationships that can lead to a better explicability of results.

To better understand this limitation let's take the following examples:

Say, an AI program is asked to Look at the picture below and tell "if there are an equal

number of large things and metal spheres?” or "What is the shape of the object closest

to the large cylinder?"

Another Example: "How much blocks are on the right of the three-level tower? "

The Far Side of AI Neurosymbolic AI2it turns out that state-of-the-art AI neural networks systems irrespective of how much

data we give they struggle. This is really interesting for us because these are cases

where no matter how much data we have deep learning architectures doesn't answer

for some reason.

So how can we build machines that can answer these types of questions ??

Back to 1956 researchers were implementing another kind of AI called the symbolic AI.

Symbolic AI (Classic Symbolic AI):

Symbolic Artiﬁcial Intelligence, also known as Good, Old-Fashioned AI (GOFAI), was

the dominant paradigm in the AI community from the post-War era until the late

1980s.

The Symbolic AI uses human-readable symbols that represent real-world

constants (objects/entities) in order to create ‘rules’ (predicates: relations between

constants) for the concrete manipulation of those symbols, leading to a rule-based

system.

Let's take the Simpsons family tree example to understand how symbolic AI works:

The Far Side of AI Neurosymbolic AI3The constants are people for example Abe, Homer and Lisa and the predicates are the

relationship between those people. Using this rule-based paradigm we can infer new

relationships between objects that a simple Deep Learning model is uncapable of

extracting them like Herb is the half-brother of Homer and many more.

This is very similar how our brain works.

Symbolic AI benefits:

Interpretable: when dealing with real-word data we want to have an explanation

of why the algorithm gave us a certain decision (medical areas or law etc)

Generalization at concept level through reasoning

well established solvers/planning algorithms

The Far Side of AI Neurosymbolic AI4Symbolic AI VS Deep Learning

One of the main differences between machine learning and traditional symbolic

reasoning is where the learning happens. In machine- and deep-learning, the

algorithm learns rules as it establishes correlations between inputs and outputs.

In symbolic reasoning, the rules are created through human intervention. That is, to

build a symbolic reasoning system, ﬁrst humans must learn the rules by which two

phenomena relate, and then hard-code those relationships into a static program.

The Far Side of AI Neurosymbolic AI5This is how Deep Neural Networks see a picture of an "Apple" as an Apple, by

extracting the patterns at pixel level.

Source: MIT-IBM Watson AI Lab

but symbolic AI see's the apple as a graph of rules: an apple is a fruit, it could be green

or red, the shape is round etc. this sound pretty cool and this is exactly how our brain

works.

The Far Side of AI Neurosymbolic AI6Symbolic AI downsides and the raise of Neuro-symbolic AI

Symbolic AI sounds much more clever comparing to DL systems and you may wonder

why nowadays we're talking only about Deep Learning applications. A key

disadvantage of Symbolic AI is that for learning process – the rules and knowledge has

to be hand coded which is a hard problem when facing complex real word situation. For

instance, researchers had to explore newer avenues in AI rather than Symbolic AI

because they cannot handcraft the entire word problems and here's where comes the

deep learning mission to automate the handcraft work used to be done manually by

humans and together they form what's called Neuro-symbolic AI.

Neuro-Symbolic AI: Getting AI to reason

The Far Side of AI Neurosymbolic AI7Neurosymbolic AI is a unison of deep neural networks and symbolic AI techniques, it

is a combination of learning and logic. On the one hand, the deep learning is

responsible for capturing complex correlations in massive datasets and on the other

hand Symbolic models have a complementary strength in capturing compositional and

causal structure . The uniﬁcation of the two approaches would address the

shortcomings of each.

Neuro-symbolic [AI] models will allow us to build AI systems that

capture compositionality, causality, and complex correlations,"

Brenden Lake said.

Neural networks help in getting answers from the messiness of the

real-world data to a symbolic representation of the world,

constituting correlations in images. Together, they can do some

pretty magical things in reasoning. – David Cox, Director of MIT-

IBM Watson AI Lab

The Far Side of AI Neurosymbolic AI8Advantages of Neuro-Symbolic AI

Higher Accuracy.

Data Efﬁciency: The biggest problem with deploying neural networks in the real world

is that we rarely have big data. Experiences have proved that working with

Neurosymbolic AI can do better than most of the end-to-end train systems. it requires

less training examples and can perform higher accuracy.

Transparency and Interpretability: we need to understand why the decision is made

other wise we can't trust the AI system. and this is exactly what NeuroSymbolic AI do it

takes a decision that we can understand the rules of it.

Source: MIT-IBM Watson AI Lab

Conclusion

In order to imitate human learning, scientists and researchers working to develop

models of how humans represent the world and frameworks to deﬁne logic and

thought. Neuro-symbolic AI

The Far Side of AI Neurosymbolic AI9contributes towards the development of explainable and accountable AI and

machine learning-based systems and tools.

References

https://searchenterpriseai.techtarget.com/feature/Neuro-symbolic-AI-seen-as-

evolution-of-artiﬁcial-intelligence

https://www.youtube.com/watch?v=4PuuziOgSU4&ab_channel=AlexanderAmini

The Far Side of AI Neurosymbolic AI10