Proteins, Particles, and Pseudo-Max-Marginals: A Submodular Approach
Authors: Jason Pacheco, Erik Sudderth
ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach using optical flow benchmarks. We further demonstrate superior side chain prediction accuracy compared to baseline algorithms from the state-of-the-art Rosetta package. |
| Researcher Affiliation | Academia | Jason L. Pacheco PACHECOJ@CS.BROWN.EDU Erik B. Sudderth SUDDERTH@CS.BROWN.EDU Department of Computer Science, Brown University, Providence, RI 02912, USA |
| Pseudocode | Yes | LAZYGREEDY Particle Selection / Initialize: For each node t let M = M T s1t, . . . , M T sdt T be the message foundations of neighbors Γ(t) = {s1, . . . , sd}. Initialize the selection vector z and margins: a=1 M(a, b), z(b) = 0 b {1, αN}. (11) First Iteration: Ensure that the current MAP estimate x is never discarded by setting z(b ) = 1, where b is the index of x t in the augmented particle set Xaug t (see Sec. 3.3). Update the message approximation ˆm(a) = M(a, b ). Iterations 2 to N: Choose the largest margin to update, eb = arg max {b|z(b)=0} (b). If (eb) = 0 then terminate prematurely, the message can be perfectly reconstructed with a subset of particles. If (eb) has already been updated on the current iteration then set z(eb) = 1 and update the message approximation, ˆm(a) = max( ˆm(a), Mt(a,eb)). Otherwise, update the margin and repeat, h max( ˆm(a), M(a,eb)) ˆm(a) i . Selections are performed in parallel and updates at one node do not affect the selection at neighboring nodes. Figure 3 graphically demonstrates LAZYGREEDY selection on the small toy protein of Fig. 2. |
| Open Source Code | Yes | A MATLAB library, built on UGM (Schmidt, 2007), implementing these methods is available3. 3http://www.cs.brown.edu/ pachecoj |
| Open Datasets | Yes | We evaluate on the Middlebury optical flow benchmark (Baker et al., 2011) using 11 random initializations. The Middlebury training set contains 8 images with ground truth flow, and we report log-probability quantiles over this set (Fig. 7 (left)). |
| Dataset Splits | No | The Middlebury training set contains 8 images with ground truth flow, and we report log-probability quantiles over this set (Fig. 7 (left)). |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | A MATLAB library, built on UGM (Schmidt, 2007), implementing these methods is available3. |
| Experiment Setup | Yes | We use the Charbonnier widths σ = 0.001 recommended for this model (Sun et al., 2014), but learn different scaling parameters (λs = 16, λd = 1) to compensate for our superpixel representation. We compute SLIC superpixels (Achanta et al., 2012) with region size 5 and regularizer 0.1; We run PMP with 50 particles for 50 iterations. D-PMP and T-PMP proposals are 50% random walks from Gaussians wrapped to account for angular discontinuities, and 50% samples from the rotamer marginals. |