Gradient descent optimization of smoothed information retrieval metrics

By Olivier Chapelle, Mingrui Wu. 2010

214 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition, Information Systems

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Information Systems

Citations in the corpus

Recommendations as Treatments: Debiasing Learning and Evaluation (2016, background): “Solving this training objective for other δ u,i ( Y, ˆ Y ) that are non-differentiable is more challenging, but possible avenues exist (Joachims, 2005; Chapelle & Wu, 2010).”
Fast Differentiable Sorting and Ranking (2020, background): “While several works opt to approximate ranking metrics directly (Chapelle & Wu, 2010; Adams & Zemel, 2011; Lapin et al., 2016), others introduce “soft” ranks, which can then be plugged into any differentiable loss function.”
A Vector-Contraction Inequality for Rademacher Complexities (2016, background): “These examples are not exhaustive, in fact it seems that many applications of Slepian’s inequality in the machine learning literature can be circumvented by Theorem 2 (see also [6], [7], [17] for applications to information retrieval and generalization of autoregressive models).”
Hand Pose Estimation via Latent 2.5D Heatmap Regression (2018, background): “(14) is known as soft-argmax in the literature [61].”
H+O: Unified Egocentric Recognition of 3D Hand-Object Poses and Interactions (2019, unknown): “In principle, it is straightforward to merge and train the single-image and temporal models jointly with a softargmax operation [5, 71].”
On Graduated Optimization for Stochastic Non-Convex Problems (2015, methodology): “For a set A ⊂ R d , u ∼ A denotes a random variable distributed uniformly over A .”
Bagging gradient-boosted trees for high precision, low variance ranking models (2011, unknown): “…MAP SVMmap [34] 0.2933 0.3035 0.3007 0.2049 RankSVM-Struct [20] 0.3467 0.3371 0.3192 0.2196 ListNet [9] 0.3600 0.3573 0.3325 0.2231 SmoothRank [11] 0.4000 0.3832 0.3555 0.2326 RankSVM [17] 0.4133 0.3467 0.3240 0.2237 AdaRank-MAP [33] 0.4133 0.3757 0.3602 0.2189 AdaRank-NDCG [33] 0.4267 0.3688…”
TFMAP: optimizing MAP for top-n context-aware recommendation (2012, background): “Based on insights in [8], we approximate I(rmjk ≤ rmik) by the following logistic function:”
LIFT: Learned Invariant Feature Transform (2016, methodology): “We also replace the traditional approaches to non-local maximum suppression (NMS) by the soft argmax function [12].”
LF-Net: Learning Local Features from Images (2018, methodology): “From this scale-invariant map we choose the top K pixels as keypoints, and further apply a local softargmax [7] for sub-pixel accuracy.”
CLiMF: learning to maximize reciprocal rank with collaborative less-is-more filtering (2012, methodology): “Inspired by recent developments in the area of learning to rank [6], we derive an approximation of I(Rik < Rij) by using a logistic function:”
Understanding the impact of entropy on policy optimization (2018, background): “One approach to visualize an objective function is to interpolate θ in the 1D subspace between two points θ 0 and θ 1 (Chapelle & Wu, 2010; Goodfellow et al., 2015) by evaluating the objective at O ((1 − α ) θ 0 + αθ 1 ) for 0 ≤ α ≤ 1.”
On the Link between Gaussian Homotopy Continuation and Convex Envelopes (2015, methodology): “Chapelle has utilized continuation in various applications such as semi-supervised learning [12,13,59], semi-supervised structured output [21], multiple instance learning [28], and ranking [14].”
StereoNet: Guided Hierarchical Refinement for Real-Time Edge-Aware Depth Prediction (2018, background): “The first of which is soft argmin, which was originally proposed in [6] and was used in [29].”