Hardware architecture and processing units for exact Bayesian inference with on-line learning and methods for same
OwnerAndreas G. Andreou, Tomas Figlioliainventors
Life of this patent
Estimating the term…
Abstract
A reconfigurable computing architecture for Bayesian Online ChangePoint Detection (BOCPD) is provided. In an exemplary embodiment, the architecture may be employed for use in video processing, and more specifically to computing whether a pixel in a video sequence belongs to the background or to an object (foreground). Each pixel may be processed with only information from its intensity and its time history. The computing architecture employs unary fixed point representation for numbers in time, using pulse density or random pulse density modulation—i.e., a stream of zeros and ones, where the mean of that stream represents the encoded value.
Claims5
A change point detector, comprising:
- a sample input, configured to receive a sample xt at a sample frequency Fsamp and output n stochastic streams representing xt;
- a set of Nwin Ro registers representing a probability distribution value of a prior runlength, P(rt-1|X1:t-1=x1:t-1), wherein each Nwin Ro register of the set of Nwin Ro registers is configured to store a count of prior Bernoulli samples;
- a Wo register configured to store a sum of all values in the set of Nwin Ro registers as a normalizing constant for the prior runlength probability distribution value;
- an encoder for probabilistically encoding each value stored by the set of Nwin Ro registers and normalized by the value stored in the Wo register;
- a mu_update circuit having an input for receiving a stochastic stream representing sample xt and variance σ20 for calculating mean parameters μok for k∈[1:
- Nwin—1];
- a set of Nwin Bernstein circuits, each Nwin Bernstein circuit of the set of Nwin Bernstein circuits configured to receive a set of n Bernstein coefficient signals and stochastic streams representing sample xt and mean μok for a particular value k wherein each Nwin Berstein circuit operates on a value k different from the other Bernstein circuits, and to provide an output of P(Xt|{right arrow over (θ)}k);
- a set of Nwin combinatorial circuits, each Nwin combinatorial circuit of the set of Nwin combinatorial circuits configured to receive the output from a corresponding Nwin Bernstein circuit and a corresponding encoded Nwin Ro register from the encoder, and wherein a set of combinatorial circuits provide outputs of Bernoulli samples;
- a set of Nwin Ri registers for storing a probability distribution values of a current runlength P(rt|X1:t-1=xi:t-1), wherein each Nwin Ri register of the set of Nwin Ri registers is configured to count Bernoulli samples;
- received from a corresponding combinatorial circuit of the Nwin combinatorial circuits, and wherein each Nwin Ri register is configured to output its stored value to a corresponding Nwin Ro register each time a new sample xt is received by the change point detector;
- and
- a Wi register configured to store a sum of all values of the set of Nwin Ri registers as a normalizing constant for a current runlength probability distribution.
Dependencies are inferred from the claim text.