Neural Symbollic Regression Using Deep Learning and Sparse Modelling
The paper proposes using a neural network first, then LASSO, to recover sparse symbolic equations from data.
The authors frame the neural model as a noise-robust preconditioner for symbolic regression, rather than the final black-box predictor. Their pipeline learns a smooth nonlinear feature space, then applies sparse regression to extract closed-form expressions. On the Nguyen benchmark suite, they report better RMSE, noise robustness, and out-of-distribution generalization than SINDy and untuned neural baselines. Ablations attribute the gains to feature interactions, network depth, and hyperparameter tuning with Ray Tune and ASHA. ArXiv · AI/CL/LG's note
The authors frame the neural model as a noise-robust preconditioner for symbolic regression, rather than the final black-box predictor. Their pipeline learns a smooth nonlinear feature space, then applies sparse regression to extract closed-form expressions. On the Nguyen benchmark suite, they report better RMSE, noise robustness, and out-of-distribution generalization than SINDy and untuned neural baselines. Ablations attribute the gains to feature interactions, network depth, and hyperparameter tuning with Ray Tune and ASHA. ArXiv · AI/CL/LG's note
score 3