Fig. 3: Performance of fine-tuned embedding models.

a The convergence of fine-tuned models. The top eight plots fix OR₀ as the target dataset (we plot eight out of the total fifteen signals). Dark green lines show the convergence of a randomly initialized LSTM trained in OR₀ and light green show the convergence of an LSTM trained in OR₀ initialized using weights from the best model in OR₁ (fine-tuning). The bottom two rows show the analogous plots with OR₁ as the target dataset. Because deep models are typically trained iteratively using some variant of stochastic gradient descent, convergence plots are used to assess the convergence of deep models as a function of the number of iterations (epochs) based on the performance on a held out validation set (validation loss). b The performance of GBT models trained on embeddings from standard embedding models (next), transferred embedding models (next'), and fine-tuned embedding models (next^ft) (best models from light green in (a)). We report the average precision value of the raw model in parenthesis on the x-axis.