Possible extensions

[dgcnz]

• Using L1 regularization instead L2 to induce more sparseness/interpretability
• Measure causality of final linear layer
• Explicitly limit the residual component in PCBM-h
• Test finetuned CLIP in medical domain with custom concepts: https://huggingface.co/datasets/alkzar90/NIH-Chest-X-ray-dataset
• try multimodal audio models/test for audio classification
• random projection
• complexity measures

[Nesta-gitU]

• Using L1 regularization instead L2 to induce more sparseness/interpretability
• Measure causality of final linear layer
• Test for audio classification

l1 already used because we have l1_ratio of 0.99