This paper proposes Persistent Evolution Strategies (PES) to estimate unbiased gradients for unrolled computational graphs. PES allows obtaining unbiased gradients even with truncated unrolls, like truncated BPTT for RNNs, which previously only allowed biased gradients. Experiments demonstrate PES can be applied to tasks like RNN-like problems, hyperparameter optimization, reinforcement learning, and meta-learning. PES achieves unbiased gradient estimation for unrolled computational graphs faster than prior methods.
Related topics: