How long a forget gate remembers
limit▲▲△Exact values where possible; six decimal places for non-exact numerical results.
STATEMENT
Turn a constant forget value into a number of steps, bound that number, and find the forget value and bias a given horizon requires.
GIVEN
One coordinate’s forget value is at every step. The carry-line factor after steps is . The memory horizon is .
FIND
(a) and for , with its horizon and half-life. (b) Bounds on in terms of . (c) The smallest with , and the bias that produces it when the gate’s other inputs are zero. (d) The behaviour of as and as the bias grows.
STRATEGY
Write and use twice. For the bias, invert the sigmoid.
SOLUTION
(a) Directly, and . The horizon is steps. The half-life solves , so steps.
(b) The inequality is strict for . At it gives . At it gives , that is, . Taking reciprocals,
The two bounds differ by exactly one, so gives the horizon to within a step. For they are and .
(c) The condition is , so . A gate with only its bias has , so .
(d) As , and ; by (b), grows like . With , (I.10.5) gives . Each extra unit of bias multiplies the horizon by about , but no finite bias reaches .
Answer
(a) , ; horizon ; half-life . (b) . (c) , bias at least . (d) like , never infinite at finite .
Check — sanity
At the bounds give , and indeed . A zero forget bias forgets most of a value within two steps.
Where this breaks
A trained gate’s value changes with the input, so the horizon of a real cell is a distribution over steps, not a constant. The bound also concerns the carry line only; I.10.B02 shows the full derivative can be several times larger or smaller.
Variation
For a horizon of exactly in the sense of , set . For that is , slightly below the answer to (c), because exceeds .