In "Neural architecture search with reinforcement learning" (https://arxiv.org/pdf/1611.01578.pdf), the authors use a recurrent neural network as the controller to generate hyperparameters, but do not describe the architecture of the controller in detail.
- How are the predictions for each hyperparameter made? The paper says that softmax is used and hyperparameters are generated sequentially (first filter width/height, then stride width/height, etc., then repeating for the next layer). Thus, what architecture allows predictions to be made sequentially in such a way?
- How are the skip connection sigmoids incorporated into this architecture?