Deep Learning has achieved tremendous results by pushing the frontier of automation in diverse domains. Unfortunately, current neural network architectures are not explainable by design. In this paper, we propose a novel method that trains deep hypernetworks to generate explainable linear models. Our models retain the accuracy of black-box deep networks while offering free lunch explainability by design. Specifically, our explainable approach requires the same runtime and memory resources as black-box deep models, ensuring practical feasibility. Through extensive experiments, we demonstrate that our explainable deep networks are as accurate as state-of-the-art classifiers on tabular data. On the other hand, we showcase the interpretability of our method on a recent benchmark for empirically comparing prediction explainers. The experimental results reveal that our models are not only as accurate as their black-box deep-learning counterparts but also as interpretable as state-of-the-art explanation techniques.

Authors: Arlind Kadra, Sebastian Pineda Arango, Josif Grabocka

# The following commands assume the user is in the cloned directory
conda create -n inn python=3.9
conda activate inn
cat requirements.txt | xargs -n 1 -L 1 pip install

The entry script to run INN and TabResNet is main_experiment.py. The entry script to run the baseline methods (CatBoost, Random Forest, Logistic Regression, Decision Tree and TabNet) is baseline_experiment.py.

The main arguments for main_experiment.py are:

• --nr_blocks: Number of residual blocks in the hypernetwork.
• --hidden_size: The number of hidden units per-layer.
• --nr_epochs: The number of epochs to train the hypernetwork.
• --batch_size: The number of examples in a batch.
• --learning_rate: The learning rate used during optimization.
• --augmentation_probability: The probability with which data augmentation will be applied.
• --weight_decay: The weight decay value.
• --weight_norm: The L1 coefficient that controls the sparsity induced in the final importances per-feature.
• --scheduler_t_mult: Number of restarts for the learning rate scheduler.
• --seed: The random seed to generate reproducible results.
• --dataset_id: The OpenML dataset id.
• --test_split_size: The fraction of total data that will correspond to the test set.
• --nr_restarts: Number of restarts for the learning rate scheduler.
• --output_dir: Directory where to store results.
• --interpretable: If interpretable results should be generated, basically if INN should be used or the TabResNet architecture.
• --mode: Takes two arguments, classification and regression.

A minimal example of running INN:

python main_experiment.py --output_dir "." --dataset_id 1590 --nr_restarts 3 --weight_norm 0.1 --weight_decay 0.01 --seed 0 --interpretable

The plots that are included in our paper were generated from the functions in the module plots/comparison.py. The plots expect the following result folder structure:

├── results_folder
│   ├── method_name
│   │   ├── dataset_id
│   │   │   ├── seed
│   │   │   │   ├── output_info.json

@misc{kadra2023breaking,
      title={Breaking the Paradox of Explainable Deep Learning}, 
      author={Arlind Kadra and Sebastian Pineda Arango and Josif Grabocka},
      year={2023},
      eprint={2305.13072},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}