These are my own personal philosophical musings on AI research, geometric deep learning, and life in general. Everything I build in my academic career rests on one premise. All data, whether images, point clouds, molecular graphs, sensor signals, or language, is a representation of phenomena taking place in a shared reality, and that reality is geometrically organized. The structure present in data is the trace of symmetry and physical law. Data does not just have geometric structure the way a house has a coat of paint. Data is geometry, because the world that cast it is.
From this I draw an engineering conclusion: a model whose internal organization ignores the structure of the world it represents discards what is most principled about its data. That conclusion, and the machinery it leads to, is what my course on Group Equivariant Deep Learning is about, and maybe one day I will write the technical posts too. But first I'd like to focus on the whole premise itself.
So: data inherits the structure of reality. But this kind of assumes we already know what reality is. Does anyone? I don't think we get to leave that to philosophy, because representation learning is at bottom the science of drawing valid inferences about reality from representations of it. We can at least try to reach some consensus on what we are talking about.
These posts ask the questions underneath my research program, in order. How should we think about reality in the first place? What structure do mind and matter share, and why do I keep answering geometry? What is a mind, and what is intelligence? And when a machine reasons well enough to unsettle us, does it become something more than a technology, or does it stay one? The posts below are my answers.