More Research
Hierarchical Conditioning of Diffusion Models Using Tree-of-Life for Studying Species Evolution
ECCV 2024
1Mridul Khurana, 2Arka Daw, 1M. Maruf, 1Josef Uyeda, 3Wasila Dahdul, 1Caleb Charpentier, 4Yasin Bakış, 4Henry L. Bart Jr., 5Paula Mabee, 6Hilmar Lapp, 7James Balhoff, 8Wei-Lun (Harry) Chao, 9Charles Stewart, 8Tanya Berger-Wolf, 1Anuj Karpatne
1Virginia Tech, 2Oak Ridge National Lab, 3University of California, Irvine, 4Tulane University, 5Battelle, 6Duke University, 7University of North Carolina at Chapel Hill, 8The Ohio State University, 9Rensselaer Polytechnic Institute
mridul@vt.edu, karpatne@vt.edu
Phylo-Diffusion
We introduce Graph as a modality to conditional Diffusion Models, Phylo-Diffusion, for understanding in evolutionary traits in biological specimens. We leverage this new modality and introduce two experiments for understanding evolutionary traits inspired by gene-knockout experiments:
- Trait Masking: where we mask out one or more level embedding by substituting it with gaussian noise to verify that the embeddings learned capture hierarchical information.
- Trait Swapping: which involves substitution of level-l embedding of a source species with the level-l embedding of a sibling subtree at an equivalent level to visualize the trait differences in the generated images before and after trait swapping that can help us understand the evolutionary traits that branched at a certain level.
Demo
Please note that the demo is currently running on free CPU resources provided by Hugging Face, so it may take up to 10 minutes to generate an image. We're working on securing additional resources to speed up the process. Thank you for your patience!
Experiments
We evaluate Phylo-Diffusion and show that it achieves high-fidelity at par with the text-to-image and class-conditional diffusion models.
Phylo-Diffusion performs at par with state-of-the-art generative models
Scroll to see all results.
Model Type | Method | FID ↓ | IS ↑ | Prec. ↑ | Recall ↑ |
---|---|---|---|---|---|
GAN | Phylo-NN | 28.08 | 2.35 | 0.625 | 0.084 |
Diffusion | Class Conditional | 11.46 | 2.47 | 0.679 | 0.359 |
Diffusion | Scientific Name | 11.76 | 2.43 | 0.683 | 0.332 |
Diffusion | Phylo-Diffusion (ours) | 11.38 | 2.53 | 0.654 | 0.367 |
Trait Masking
Trait Swapping
Figure 4 illustrates trait swapping for the source species Noturus exilis (left), where the information at Level-2 is swapped with that of a sibling subtree at Node B (right). The image in the center is generated using the trait swapped embedding. This visualization of the perturbed species helps us study the trait changes that would have branched out at level-2 between Node A and Node B.In the generated image (center), we observe the absence of barbels(whiskers), and the caudal fin (tail) is getting forked (or split) highlighted in pink, which are traits adopted from species in the subtree at B (Notropis). Whereas other fins like the dorsal, pelvic, and anal fin still resemble the source species Noturus exilis highlighted in green. The same is also reflected in the change of probability distribution after perturbations; the probability distribution of source species Noturus exilis decreases and the probability of it being a Notropis increases slightly.Comparision to PhyloNN
References
Please cite our paper if you use our code, data, model or results.
@article{khurana2024hierarchical, title={Hierarchical Conditioning of Diffusion Models Using Tree-of-Life for Studying Species Evolution}, author={Khurana, Mridul and Daw, Arka and Maruf, M and Uyeda, Josef C and Dahdul, Wasila and Charpentier, Caleb and Bak{\i}{\c{s}}, Yasin and Bart Jr, Henry L and Mabee, Paula M and Lapp, Hilmar and others}, journal={arXiv preprint arXiv:2408.00160}, year={2024} }
Also consider citing LDM:
@inproceedings{rombach2022high, title={High-resolution image synthesis with latent diffusion models}, author={Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj{\"o}rn}, booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, pages={10684--10695}, year={2022} }
Acknowledgements
This work was supported by the Imageomics Institute, which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under Award #2118240 (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.