The activation dynamics is implemented exactly as in the 2016 paper, but this results in activation and deactivation happening at almost the same speed. The function f is intended as a "soft switch" between activation (a-e < 0) and deactivation (a-e > 0). But it's too soft, the value of f stays very close to 0.5 over the whole range of inputs (a-e ranging from -1 to 1).

I suspect that the tanhSteepness parameter should have been 10 instead of 0.1.

See:

I ran some simulations to see the response to a 1 Hz square wave excitation, and this confirms that activation and deactivation speeds are very similar.

I contacted Friedl for input, a few weeks ago, but received no reply yet.