IC-1400When steerable feature dimension is held constant, increasing the type-l of steerable features does not improve performance of ESCN or EquiformerV2 on IS2RE and S2EF molecular property prediction
Shih-Hsin Wang, Yung-Chang Hsu, Justin Baker, Andrea L. Bertozzi, Jack Xin, Bao Wang
An ablation study on the OC20 IS2RE dataset (10k training molecules) shows that ESCN with l=2, c=206 (feat dim 1854) achieves 842±13 meV energy MAE, while l=4, c=98 (feat dim 1862) gives 929±15 meV and l=6, c=64 (feat dim 1856) gives 872±6 meV. EquiformerV2 shows a similar pattern: l=2, c=34 (feat dim 306) gives 841±21 meV versus l=4, c=16 (feat dim 304) at 832±11 meV. On S2EF, ESCN l=2 c=824 (feat dim 7416) achieves 246 meV MAE, outperforming ESCN l=6 c=256 (feat dim 7424) at 294 meV. The results support the theoretical claim that information-carrying ability of steerable features is characterized by dimension, not by the highest type utilized.
The S2EF comparison uses results from [28] for l=2 c=256 and l=6 c=256 (trained 12 epochs on 16 GPUs) versus this paper's l=2 c=824 (trained 8 epochs on 2 GPUs), so training budget differs. The authors note these tasks do not rule out confounding factors like over-squashing. IS2RE uses only 10k training molecules due to computational constraints.