Generating Random Parameters in Feedforward Neural Networks with Random Hidden Nodes: Drawbacks of the Standard Method and How to Improve It
This work addresses a specific technical issue in neural network initialization, likely incremental for researchers in random neural networks.
The paper identifies drawbacks in the standard method of generating random weights and biases from a uniform distribution in feedforward neural networks with random hidden nodes, and proposes a new method that ensures the most nonlinear parts of sigmoids are retained in the input hypercube and generates activation functions with uniformly distributed slope angles.
The standard method of generating random weights and biases in feedforward neural networks with random hidden nodes, selects them both from the uniform distribution over the same fixed interval. In this work, we show the drawbacks of this approach and propose a new method of generating random parameters. This method ensures the most nonlinear fragments of sigmoids, which are most useful in modeling target function nonlinearity, are kept in the input hypercube. In addition, we show how to generate activation functions with uniformly distributed slope angles.