Fig. 3: Benchmarking transfer learning and mixture of experts performance versus dataset size. | npj Computational Materials

Fig. 3: Benchmarking transfer learning and mixture of experts performance versus dataset size.

From: Towards overcoming data scarcity in materials science: unifying models and datasets with a mixture of experts framework

Fig. 3

Percent improvement on MAE across 19 materials property regression tasks of transfer learning from Materials Project formation energies (TL) and our mixture of experts approach (MoE) over single-task learning with random initialization (STL). For clarity, the y axis above and below the break have different scales. Examples of positive and negative transfer are indicated by points above and below the gray dotted line, respectively.

Back to article page