Method and system for determining an optimal or near optimal set of contexts by constructing a multi-directional context tree
View Patent ↗In various embodiments of the present invention, optimal or near-optimal multidirectional context sets for a particular data-and/or-signal analysis or processing task are determined by selecting a maximum context size, generating a set of leaf nodes corresponding to those maximally sized contexts that occur in the data or signal to be processed or analyzed, and then building up and concurrently pruning, level by level, a multidirectional optimal context tree constructing one of potentially many optimal or near-optimal context trees in which leaf nodes represent the context of a near-optimal or optimal context set that may contain contexts of different sizes and geometries. Pruning is carried out using a problem-domain-related weighting function applicable to nodes and subtrees within the context tree. In one described embodiment, a bi-directional context tree suitable for a signal denoising application is constructed using, as the weighting function, an estimated loss function.
1. A method for estimating the effectiveness of a denoiser for denoising a noisy symbol sequence z received following transmission of a clean symbol sequence x through a noisy channel, the method comprising:
computing, for each symbol z i , in the noisy symbol sequence, an estimated loss resulting from replacing the symbol z i , with a replacement symbol computed by the denoiser; and
adding together the estimated losses computed for each symbol z i to produce an estimated loss resulting from denoising the noisy sequence z by the denoiser.
2. The method of claim 1 wherein computing, for each symbol z i , in the noisy symbol sequence, an estimated loss for replacing the symbol z i , with a replacement symbol computed by the denoiser further includes:
for each symbol x in an alphabet X from which symbols of the symbol sequences are chosen,
computing an unbiased estimate of a selection function δ(x i =x) where x is a symbol in the alphabet X and x i , is the clean symbol sequence symbol corresponding to the symbol z i ,
computing an average loss term equal to the computed loss for replacing a random symbol z with a replacement symbol selected by the denoiser, given that the clean signal symbol x i , is equal to x, and
multiplying the estimate of a selection function δ(x i =x) by the computed average loss term to compute an estimated loss for the symbol x; and
summing together the estimated losses computed for each symbol x.
3. The method of claim 1 used to determine a context set for a context-based denoiser to apply to the noisy sequence z by selecting a context set that produces a minimal estimated loss for denoising the noisy sequence z.
4. Computer instructions encoded and stored in a non-transitory computer readable medium that implement the method of claim 1 for estimating the effectiveness of a denoiser for denoising a noisy symbol sequence z received following transmission of a clean symbol sequence x through a noisy channel, the method comprising:
computing, for each symbol z i in the noisy symbol sequence, an estimated loss resulting from replacing the symbol z i with a replacement symbol computed by the denoiser; and
adding together the estimated losses computed for each symbol z i to produce an estimated loss resulting from denoising the noisy sequence z by the denoiser.
5. A method for selecting a denoiser to apply to a noisy symbol sequence z received following transmission of a clean symbol sequence x through a noisy channel, the method comprising:
for each denoiser,
estimating the effectiveness of the denoiser for denoising a noisy symbol sequence z received following transmission of a clean symbol sequence x through a noisy channel by
computing, for each symbol z i in the noisy symbol sequence, an estimated loss resulting from replacing the symbol z i with a replacement symbol computed by the denoiser, and
adding together the estimated losses computed for each symbol z; to compute an estimated loss resulting from denoising the noisy sequence z by the denoiser; and
selecting the denoiser with a smallest computed estimated loss.
6. Computer instructions encoded and stored in a non-transitory computer readable medium that implement the method of claim 5 for selecting a denoiser to apply to a noisy symbol sequence z received following transmission of a clean symbol sequence x through a noisy channel, the method comprising:
for each denoiser,
estimating the effectiveness of the denoiser for denoising a noisy symbol sequence z received following transmission of a clean symbol sequence x through a noisy channel by
computing, for each symbol z i , in the noisy symbol sequence, an estimated loss resulting from replacing the symbol z i , with a replacement symbol computed by the denoiser, and
adding together the estimated losses computed for each symbol z i , to compute an estimated loss resulting from denoising the noisy sequence z by the denoiser; and
selecting the denoiser with a smallest computed estimated loss.
7. A system that selects a denoiser to apply to a noisy symbol sequence z received following transmission of a clean symbol sequence x through a noisy channel, the system comprising:
logic that computes, for each symbol z i , in the noisy symbol sequence, an estimated loss resulting from replacing the symbol z i , with a replacement symbol computed by a particular denoiser;
logic that adds together the estimated losses computed for each symbol z i , for a particular denoiser to compute an estimated loss resulting from denoising the noisy sequence z by the particular denoiser; and
logic that selects as the denoiser to apply to the noisy symbol sequence the denoiser with the lowest computed estimated loss.
8. The system of claim 7 wherein the logic that computes, for each symbol z i , in the noisy symbol sequence, an estimated loss for replacing the symbol z i , with a replacement symbol computed by the denoiser computes the estimated loss by:
for each symbol x in an alphabet X from which symbols of the symbol sequences are chosen,
computing an unbiased estimate of a selection function δ(x i =x) where x is a symbol in the alphabet X and x i , is the clean symbol sequence symbol corresponding to the symbol z i ,
computing an average loss term equal to the computed loss for replacing a random symbol z with a replacement symbol selected by the denoiser, given that the clean-signal symbol x i , is equal to x, and
multiplying the estimate of a selection function δ(x i =x) by the computed average loss term to compute an estimated loss for the symbol x; and
summing together the estimated losses computed for each symbol x.