Md. Asif Uddin
    I.10.X01

    A saturated gate still leaks

    limit▲△△

    A forget gate has bias 1010 and no other input, so f=σ(10)f=\sigma(10). Is the carry line lossless? Compute ff, 1−f1-f, the carry-line factor after 10510^5 steps and the memory horizon.

    Hint

    Use 1−σ(b)=σ(−b)1-\sigma(b)=\sigma(-b), and (I.10.5) for the horizon.

    Solution

    f=σ(10)=0.9999546f=\sigma(10)=0.9999546 and 1−f=σ(−10)=4.539787×10−51-f=\sigma(-10)=4.539787\times10^{-5}. After 10510^5 steps the factor is f100000=0.010675f^{100000}=0.010675: about one per cent of the signal is left.

    The horizon is −1/ln⁡f=22026.97-1/\ln f=22026.97 steps. By (I.10.5) it lies just below 1/(1−f)=1+e10=22027.471/(1-f)=1+e^{10}=22027.47.

    The carry line is not lossless. A sigmoid gate can be pushed as close to one as the bias allows, which sets a long but finite horizon. Only the 1997 design, with no forget gate at all, had a factor of exactly one, and I.10.X03 shows what that cost.

    Draws on