We develop a general approach to distill symbolic representations of a learned deep model by introducing strong inductive biases. We focus on Graph Neural Networks (GNNs). The technique works as follows: we first encourage sparse latent representations when we train a GNN in a supervised setting, then we apply symbolic regression to components of the learned model to extract explicit physical relations. We find the correct known equations, including force laws and Hamiltonians, can be extracted from the neural network. We then apply our method to a non-trivial cosmology example--a detailed dark matter simulation--and discover a new analytic formula which can predict the concentration of dark matter from the mass distribution of nearby cosmic structures. The symbolic expressions extracted from the GNN using our technique also generalized to out-of-distribution-data better than the GNN itself. Our approach offers alternative directions for interpreting neural networks and discovering novel physical principles from the representations they learn.

Paper presents an exciting area of research. All reviewers agree that the paper makes novel contributions. The one weak point of the current submission is that this work is not properly contextualized with prior work. Further, as authors said in their rebuttal -- it would be good to see comparisons with other SR packages and SR only baseline.

Additional Feedback: 0. The notations in the method section especially Section 2 need to be specified, even if it is easy to infer from context,. For example, L_v, v_i, v_j etc. need to be explained. Further, in the case studies sections, the descriptions are not clear, for example, the system should be explained mathematically from a n-body perspective, clearly denoting the particles as nodes at gnn equation level for atleast one case. The authors should discuss the intuitions behind their specific model decisions, for example, as this is a model discovery task, why haven't the authors used generative model frameworks? 2. The input/output dimensionality for eureqa fitting should be explained in Section 3, for example, GNs have multiple layers, how does the proposed method fit equations for the edge/node functions at different layers and put them together? From the simulation dataset, the underlying model does not seem to need multiple layers for GNs. 3. The Hamiltonian Dynamics section is very hard to read, especially to a non-physics person, it would be helpful if the authors add a clear description of the input (like position and momentum) and output for the HGN. 4. What is the intuition behind the sum of pairwise and self for the HGN? Have the authors compared to a model without this assumption? 5. Does the Bottleneck model perform worse simply because its a much smaller model than the other models with a large hidden size? 6. Line 170 states that "models are trained to predict acceleration given current state", do the authors not account for time dependence?

We develop a general approach to distill symbolic representations of a learned deep model by introducing strong inductive biases. We focus on Graph Neural Networks (GNNs). The technique works as follows: we first encourage sparse latent representations when we train a GNN in a supervised setting, then we apply symbolic regression to components of the learned model to extract explicit physical relations. We find the correct known equations, including force laws and Hamiltonians, can be extracted from the neural network. We then apply our method to a non-trivial cosmology example--a detailed dark matter simulation--and discover a new analytic formula which can predict the concentration of dark matter from the mass distribution of nearby cosmic structures.

At age 19, I read an interview of physicist Lee Smolin. The idea that a foreseeable limit exists on our understanding of physics by the end of my life was profoundly unsettling. I felt frustrated that I might never witness solutions to the great mysteries of science, no matter how hard I work. But… perhaps one can find a way to tear down this limit. Artificial intelligence presents a new regime of scientific inquiry, where we can automate the research process itself.