Séminaire PhilMath

Nous aurons le plaisir d'accueillir Denis Bonnay Université Paris Nanterre

Mathematics and the Creativity of Neural Networks

Résumé :

As opposed to purely predictive machine learning, generative AIs make it possible to "create" texts or images. How creative are those AIs, and which role does mathematics play in the process?

In this talk, I will focus on image generation machines, such as Midjourney, Dall.e or Stable Diffusion.Whether those AIs are really creative is a matter of controversy, some assuming that they would only mix and match what they have been trained on, some insisting that there is no reasonable sense in which the generated images already 'exist' in the training data. It might seem difficult to answer the question without presupposing too much about creativity or machines, but taking one step back and looking at the underlying maths and network structures actually helps.

More precisely, my aim in the talk is threefold:

1/ explaining the role mathematics plays in the generation process. Mathematics is instrumental in producing a shared space representations where image contents and text contents are suitably aligned

2/ showing that a proper understanding of this role makes it possible to adjudicate the creativity controversy. AIs are creative in as much as they are imaginative, their key component being the ability to travel along that space and “translate” representations from one medium to the other.

3/ suggesting that this role makes the case for a specific take on the toolbox view of the applicability of mathematics. Gen AIs build on pre-existing tools of representational learning and make those tools work together.

Zoom