DeepCausality still has some more work to do because of its early stage. In its current state, a handful of limitations exist:
-
Counterfactual reasoning is missing. DeepCausality cannot reason counter to the fact; for example, if the drone had not accelerated more than 50mp/h, it would not have crashed into the tree.
-
Causal structural learning is missing. Right now, causal models have to be designed and built by hand. Possible areas of exploration for causal structural learning are:
- Causal Reinforcement Learning (Elias Bareinboim)
- Deep Neuro Evolution (Kenneth O. Stanley)
-
Tooling is absent. Specifically, tools such as model server, dashboard, and model visualization are missing.
None of these limitations is definitive, meaning with some creative work, counterfactual reasoning might become solvable, and, likewise, causal structural learning might become feasible when adapting, for example, deep neuroevolution to the hyper-graph structure central to DeepCausality. Tooling may require a deeper work commitment since good tooling usually requires solid design and implementation.