Table 2 Commonly faced challenges in computational biology and potential solution avenues when using DL.
From: Current progress and open challenges for applying deep learning across the biosciences
Challenge | Experimental/non-DL solution | DL solution |
|---|---|---|
Biased results | Improve study design | Identify forms and sources of technical bias |
Fair AI approaches | ||
High infrastructure costs | Optimize code performance | Optimize DL architecture |
Parallelize code | Parallelize to low-cost devices | |
Sub-sample analyzed data | Condense training data (e.g. coresets) | |
Lack of explainability | Statistical analyses | Explainable post-hoc methods |
Limited training data | Generate and label more data | Data augmentation (e.g. GANs) |
Overfitting | Regularization | Dropout |
Early stopping | ||
Smaller models | ||
Additional training data | ||
Poor performance on novel data | Expand databases | Use larger models |
Analyze generalization potential |
