Generalized P{\'o}lya Urn for Time-Varying Pitman-Yor Processes
Authors: François Caron, Willie Neiswanger, Frank Wood, Arnaud Doucet, Manuel Davy
JMLR 2017 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate our models and algorithms on epidemiological and video tracking data. We illustrate experimentally that this model succeeds at filling this role. The remainder of the paper is organized as follows: ... Section 5 on various applications. ... We demonstrate the models and algorithms on simulated data, modeling of the spread of a disease, and multi-object tracking. |
| Researcher Affiliation | Collaboration | Fran cois Caron EMAIL Department of Statistics University of Oxford Oxford, UK; Willie Neiswanger EMAIL Machine Learning Department Carnegie Mellon University Pittsburgh, USA; Frank Wood EMAIL Department of Engineering Science University of Oxford Oxford, UK; Arnaud Doucet EMAIL Department of Statistics University of Oxford Oxford, UK; Manuel Davy EMAIL VEKIA 165 Avenue de Bretagne 59044 Lille, France |
| Pseudocode | Yes | Algorithm 1 Generalized P olya Urn; Algorithm 2 Sequential Monte Carlo for Time-Varying Pitman-Yor mixture processes; Algorithm 3 N best algorithm for Deterministic Deletion |
| Open Source Code | Yes | A preliminary version of this work has been presented as a conference paper (Neiswanger et al., 2014). Matlab code is available at https://github.com/willieneis/Dirichlet Process Tracking. |
| Open Datasets | Yes | We use a benchmark humantracking video from the International Workshop on Performance Evaluation of Tracking and Surveillance (PETS) 2009-2013 conferences (Ellis and Ferryman, 2010), due to its prominence in a number of studies (listed in Figure 10(f)). |
| Dataset Splits | No | The paper discusses data generation and collection in a time-series context, but does not specify explicit training/test/validation splits. For example, in Section 5.1: "The observations zt are generated for t = 1, . . . , 1000 from a sequence of mixtures of normal distributions". In Section 5.3: "At each frame t, we assume we are given a set of Nt foreground pixels..." |
| Hardware Specification | No | No specific hardware details such as GPU/CPU models, processor types, or memory amounts are mentioned. The paper describes the computational methods (SMC, PMCMC) and particle counts (e.g., "A SMC algorithm is implemented with 1000 particles"), but not the underlying hardware. |
| Software Dependencies | No | The paper mentions "Matlab code is available at https://github.com/willieneis/Dirichlet Process Tracking." in footnote 2 of Section 5.3. However, no specific version of Matlab or any other key software component is provided. |
| Experiment Setup | Yes | We assume here that the hyperparameters of the base distribution are assumed fixed and known µ0 = 0, κ0 = 0.1, ν0 = 2 and Λ0 = 1. ... where τ = 0.5 and r = 4000 are fixed parameters, ... The DP scale parameter is θ = 3. ... where aρ = 1000. ... We select a mixture of uniform and cluster deletions with ξ = 0.98. A SMC algorithm is implemented with 1000 particles ... If Neff NT , multiply the particles with large weights and remove the particles with small weights, resulting in a new set of particles denoted (i) t with weights w(i) t = 1/N. ... We consider a deterministic deletion model where r = 7; ... Inference is performed using the N best algorithm 3 with N = 1000 particles. ... where we choose τ = 1 for all experiments, ... In the following experiments we perform inference using both the SMC and PMCMC inference algorithms with N = 100 particles, and compare performance of both algorithms. ... Note that we run PMCMC as described in Section 4 for 100 iterations, where the first pass is equivalent to the SMC algorithm. |