Or "[diffusion-based] generative inpainting of incomplete Euclidean distance matrices of trajectories generated by a fractional Brownian motion". The context:
> One of the most widely used experimental techniques for examining how DNA molecules are folded in space is fluorescence microscopy [...] such data is inevitably fragmentary. [...] "Once you know the distances between a sufficient number of genes, determining the remaining distances for which there is no experimental data takes the form of a mathematical problem with a specific solution", the principal investigator of the study, Assistant Professor Kirill Polovnikov from Skoltech Neuro, commented. "We have shown for the first time that generative models are capable of solving such problems. This is an unconventional application of the kind of AI usually employed for more "creative" tasks - generating images and text based on a user prompt. At the same time, this is a new approach to the study of chromatin structure, where polymer physics has historically reigned supreme"
What is not clear is, "how can that be reliable".
They write «Our dataset reveals that, in the regime of low missing ratios, data imputation is unique, as the remaining partial graph is rigid, thus providing a reliable ground truth for inpainting». And «our findings indicate that diffusion behaves qualitatively differently from simple database searches, allowing for generalization rather than mere memorization of the training data».
But in this context, the space for failure, the possibility of sneaky errors, is what counts...
mdp2021•8mo ago
> One of the most widely used experimental techniques for examining how DNA molecules are folded in space is fluorescence microscopy [...] such data is inevitably fragmentary. [...] "Once you know the distances between a sufficient number of genes, determining the remaining distances for which there is no experimental data takes the form of a mathematical problem with a specific solution", the principal investigator of the study, Assistant Professor Kirill Polovnikov from Skoltech Neuro, commented. "We have shown for the first time that generative models are capable of solving such problems. This is an unconventional application of the kind of AI usually employed for more "creative" tasks - generating images and text based on a user prompt. At the same time, this is a new approach to the study of chromatin structure, where polymer physics has historically reigned supreme"
What is not clear is, "how can that be reliable".
They write «Our dataset reveals that, in the regime of low missing ratios, data imputation is unique, as the remaining partial graph is rigid, thus providing a reliable ground truth for inpainting». And «our findings indicate that diffusion behaves qualitatively differently from simple database searches, allowing for generalization rather than mere memorization of the training data».
But in this context, the space for failure, the possibility of sneaky errors, is what counts...