Hi, I recently came across an article about the application of OmniCLIC in multi-omics integration. One of the tables involved a comparison between GLUE and OmniCLIC on dual-omics tasks, where OmniCLIC significantly outperformed GLUE on the human-brain-3k task. I would like to understand what caused this result. While studying the OmniCLIC code, I noticed that its method for modeling embeddings is similar to most self-supervised learning approaches, with the addition of a linear layer used for classification tasks. Therefore, I am wondering whether the reason GLUE performs similarly to OmniCLIC on other tasks except for human-brain-3k might be due to OmniCLIC using 4,000 training epochs, or whether it could be attributed to differences in the way embeddings are modeled, leading to OmniCLIC’s higher accuracy.
Hi, I recently came across an article about the application of OmniCLIC in multi-omics integration. One of the tables involved a comparison between GLUE and OmniCLIC on dual-omics tasks, where OmniCLIC significantly outperformed GLUE on the human-brain-3k task. I would like to understand what caused this result. While studying the OmniCLIC code, I noticed that its method for modeling embeddings is similar to most self-supervised learning approaches, with the addition of a linear layer used for classification tasks. Therefore, I am wondering whether the reason GLUE performs similarly to OmniCLIC on other tasks except for human-brain-3k might be due to OmniCLIC using 4,000 training epochs, or whether it could be attributed to differences in the way embeddings are modeled, leading to OmniCLIC’s higher accuracy.
