Matthew Muller, our key computer scientist here at CADDS Group, recently published a ground-breaking paper on artificial intelligence (AI), discovering a new measurement for AI learning.
While AI is a fascinating field, it can often seem abstract and difficult to grasp for those not directly involved. Matthew’s professional background as a Machine Learning Engineer and Data-to-Application focused researcher centres on his curiosity about how things come together.
In a conversation with Matthew, I learned that AI has more in common with the human brain than I initially thought. ”We have only previously seen this in the brain” speaking on functional structures during learning, much like neurons forming pathways in the brain.
These structures are visualised using topological methods, for those interested, you can read more on Topology Here.
Numbers below indicate the presence of intricate learning dynamics within the AI model. The visualisations (In the example below) use colour to differentiate various features, “blue” represents disconnected neurons, while “green,” “orange” and “red” signify increasingly complex layers, building upon each other, ”the right environment conditions are essential for formation”. The presence of higher-order Betti numbers, such as those represented in red, indicates more sophisticated learnt behaviours, Matthew describes this as the ”Oracle of the network”. ”We didn’t know these would show up” the existence of these higher-order Betti numbers suggests complex learning within the AI. The chart below (a snapshot in time) illustrate how these topological features correlate with the learning process.
What does this all mean? Teaching AI to Run in Testing Environments
In practical terms, AI can be taught almost any scenario, from humanoid figures learning to stand to simulations like the “Half Cheetah” where the goal is to move a Cheetah-shaped agent forward as quickly as possible. Galloping isn’t the goal for the Half Cheetah, but it turns out to be the optimal way to move quickly.
In the video below Matthew applies his experiment to the Half Cheetah environment, the diagrams at the top shows the progression of these Betti numbers in this simulated environment, with an outline of the environment state itself in the box on the right.
As you can see, the Cheetah runs quite successfully through the experiment, in which higher order (or in the case of above ”red” Bettis) are prevalent.
Tracking these topological changes enables us to investigate optimal conditions for AI development, leading to more reliable AI applications.
A Note on Human and AI Learning
While I’ve used the analogy of the brain to explain AI structures, it’s important to recognise that human and artificial intelligence learn in fundamentally different ways. Therefore, AI should not be approached with the same methods used for human learning. Misunderstandings about AI learning and it’s core principles, and applying it without understanding how it processes information, are already causing cascading effects in our society. This is why research on how and why AI learns is so critical.
This research contributes to developing integrity around the unpredictable nature of AI, leading to more established outcomes when applied.
Matthew plans to continue his experiments and further his research in collaboration with Stellenbosch University in South Africa.
CADDS Group integrates only tested, measured, and verified artificial intelligence into our solutions while remaining committed to ongoing research and development of new applications.