HCNN - Init by minimising the correlation between simulated states

[ulf1]

Corr Loss

1. Init network
2. Simulate (without teacher forcing)
3. Compute sum of corr between all features as loss
4. Backprop loss

Correlated Surrogate

1. init network with 0 hidden states
2. generate random numbers as sequences and 1-leg inputs (simulate 1st order changes; and 2nd order => is there a pkg)
3. train with teacher forcing
4. export weight matrix without other params

Why?

• we assume that the RNN is dynamic system to produce chaotic sequences
• chaotic data and random data is not really diatinguishable