How Hard is Inference for Structured Prediction?

Authors: Amir Globerson, Tim Roughgarden, David Sontag, Cafer Yildirim

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, our analysis is validated via experimental results on 2D grid graphs. and We begin with an empirical study of the structure of real-world probabilistic inference problems. and Fig. 4 shows the expected error for the different algorithms, as a function of edge noise.
Researcher Affiliation Academia Amir Globerson GAMIR@CS.HUJI.AC.IL Tim Roughgarden TIM@CS.STANFORD.EDU David Sontag DSONTAG@CS.NYU.EDU Cafer Yildirim CAFERTYILDIRIM@GMAIL.COM
Pseudocode Yes Algorithm 1 A(X) for inference in grids. input Edge and node observations X 1: b Y arg max Y P uv E Xuv Yu Yv 2: if P v V Xv b Yv < 0 then 3: b Y b Y 4: end if output b Y
Open Source Code No No explicit statement or link providing access to the authors' own source code was found. The paper mentions using code from Sontag et al. (2012) for relaxations but does not provide their own.
Open Datasets Yes We use the Weizmann horse dataset (Borenstein & Ullman, 2002) and consider the inference problems arising from using a conditional random field (CRF) to perform foreground-background segmentation.
Dataset Splits No The parameters of the model were learned by Domke (2013, Sec. 8.3) using 200 training images. and In addition to the two-step algorithm we consider the following: Marginal inference predicting according to p(Yi|X). ... We consider a 20 20 grid, with high node noise of 0.4 and variable edge noise levels. ... Results are averaged over 100 repetitions.
Hardware Specification No No specific hardware details (like GPU/CPU models, memory, or cloud instances) were mentioned for running experiments.
Software Dependencies No The MPLP algorithm (Sontag et al., 2008) was used for inference and For both the cycle and local relaxations we use the code from Sontag et al. (2012).
Experiment Setup No We consider a 20 20 grid, with high node noise of 0.4 and variable edge noise levels. In addition to the two-step algorithm we consider the following: Marginal inference predicting according to p(Yi|X). ... Results are averaged over 100 repetitions.