Collaborative Filtering Yun-Ta Tsai 1 , Markus Steinberger 2 , Dawid - PowerPoint PPT Presentation

FastANN for High Quality Collaborative Filtering Yun-Ta Tsai 1 , Markus Steinberger 2 , Dawid Pająk 3 , Kari Pulli 4 1 Google, Inc. 2 TU Graz 3 NVIDIA 4 Light Research

Collaborative Filtering - Aggregation 

Collaborative Filtering - Result

Related Work

Distance Table Distance Table (sorted) 0.13 1.43 0.98 1.33 1.21 2.33 0.13 0.98 1.21 1.33 1.43 2.33 1.22 0.31 0.45 2.01 1.75 0.48 0.31 0.45 0.48 1.22 1.75 2.01 2.11 1.12 0.92 3.16 0.33 0.21 0.21 0.33 0.92 1.12 2.11 3.16 Garcia et al. ICIP 2010

Cayton et al. ADMS 2010

Adams et al. SIGGRAPH 2009

Limitation

Our solution

Our solution Efficient implementation on GPU General solution for different filters High image quality Applicable to different applications

Our solution

Our solution Clustering kNN Query Filtering

Design challenges Register pressure Memory access pattern Thread divergence Kernel launch overhead Memory footprint

Our solution Clustering kNN Query Filtering

Tiling  

Clustering

Warp-wide operation 3 8 2 6 3 9 1 4

Warp-wide operation 3 8 2 6 3 9 1 4 3 8 2 6 3 9 1

Warp-wide operation 3 11 10 8 9 12 10 5 3 11 10 8 9 12

Warp-wide operation 3 11 13 19 19 20 19 17 3 11 13 19

Warp-wide operation 3 11 13 19 22 31 32 36 Reduce register usage Better parallelism Minimize thread and warp divergence

Clustering kNN kNN kNN kNN

Our solution Clustering

Our solution kNN Query

kNN Query kNN kNN kNN kNN

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 p 1 kNN p 2 p 3 p 0 p 1 p 2 p 3 p 4 p 4

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 0.0 0.4 1.3 0.9 0.7 kNN p 0 p 1 p 2 p 3 p 4

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 0.0 0.4 1.3 0.9 0.7 1 1 0 0 1 ≤ 0.8

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 0.0 0.4 1.3 0.9 0.7 1 0 0 0 0 ≤ 0.0

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 0.0 0.4 1.3 0.9 0.7 1 0 0 0 0 ≤ 0.0 1 1 0 0 0 ≤ 0.4

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 0.0 0.4 1.3 0.9 0.7 1 0 0 0 0 ≤ 0.0 1 1 0 0 0 ≤ 0.4 1 1 1 1 1 ≤ 1.3

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 0.0 0.4 1.3 0.9 0.7 1 0 0 0 0 ≤ 0.0 1 1 0 0 0 ≤ 0.4 1 1 1 1 1 ≤ 1.3 ≤ 0.9 1 1 0 1 1

kNN Query p 0 p 1 p 2 p 3 p 4 p 0 0.0 0.4 1.3 0.9 0.7 1 0 0 0 0 ≤ 0.0 1 1 0 0 0 ≤ 0.4 1 1 1 1 1 ≤ 1.3 ≤ 0.9 1 1 0 1 1 ≤ 0.7 1 1 0 0 1

kNN Query p 0 p 1 p 2 p 3 p 4 1 1 1 1 0 1 1 1 0 1 kNN 0 1 1 1 1 1 1 1 0 1 1 1 0 1 1

Filtering

Our solution Hierarchical 2-mean clustering Clustering Warp-wide operators kNN search Distance table kNN Query Voting Binary coding Filtering Filtering and aggregation

Results

NN quality % of patches matching the kNN result, k=16 (the higher the better) 97.88 39.01 35.21 34.86 24.87 7.18 0.22 Randomized K-means Composite Hierarchical Generalized Random Ball Ours KD-trees Clustering PatchMatch Cover

NN quality D ann /D knn , k=16 (the lower the better) 23.91 7.38 3.01 2.00 1.99 1.32 1.01 Randomized K-means Composite Hierarchical Generalized Random Ball Ours KD-trees Clustering PatchMatch Cover

Single Frame Noise Reduction Nonlocal Means – PSNR [dB], k=16 (the higher the better) 28.55 27.83 27.79 27.13 27.02 26.88 25.65 21.24 Randomized K-means Composite Hierarchical Generalized Random Ball Ours Exhaustive KD-trees Clustering PatchMatch Cover Search

Single Frame Noise Reduction BM3D – PSNR [dB], k=16 (the higher the better) 31.10 31.05 30.72 30.71 30.68 30.57 29.87 28.92 Randomized K-means Composite Hierarchical Generalized Random Ball Ours Exhaustive KD-trees Clustering PatchMatch Cover Search

Single Frame Noise Reduction Run-time [ms], k=16, 0.25MPix (the lower the better) 8930 Query Clustering 1024 1017 982 788 8.19 Randomized KD- K-means Composite Hierarchical Generalized Ours (GPU) trees Clustering PatchMatch

Single Frame Noise Reduction Run-time [ms], k=16, 0.25MPix (the lower the better) 26359 Query 11328 Clustering 594.99 48.3 8.19 kNN-Garcia (GPU) Random Ball Cover Window Search Window Search Opt Ours (GPU) (GPU) (GPU) (GPU)

Architectures MPix/s/Watt, k=16, 0.25MPix (the higher the better) 0.29 0.16 0.02 Core i7-950 Geforce GTX 680 Tegra K1

Burst Denoising – Single Frame First frame of stack First frame of stack 26.45dB 26.45dB

Burst Denoising – Single Frame GK GK D- D- T T r r e e e e s s / / ฀r ฀r st st f f r r a a me me Ou Ou r r s s NL NL M M / / ฀r ฀r st st f f r r a a me me First frame of stack First frame of stack 3 3 1 1 . . 0 0 1 1 d d B B / / 1 1 1 1 . . 3 3 s s 3 3 1 1 . . 9 9 0 0 d d B B / / 0 0 . . 0 0 2 2 s s 26.45dB 26.45dB

Burst Denoising – All Frames GKD-Trees / stack GKD-Trees / stack Ours NLM / stack Ours NLM / stack First frame of stack First frame of stack 31.53dB / 1080s 31.53dB / 1080s 34.10dB / 0.52s 34.10dB / 0.52s 26.45dB 26.45dB

Global Illumination

Geometry Noise Reduction Noisy Input

Geometry Noise Reduction Ours

Geometry Noise Reduction Exhaustive Search Ours

Conclusions Efficient implementation on GPU High image quality Applicable to different applications

Thank you Paper and Binary: http://bit.ly/fast-ann

Collaborative Filtering Yun-Ta Tsai 1 , Markus Steinberger 2 , Dawid - PowerPoint PPT Presentation

FastANN for High Quality Collaborative Filtering Yun-Ta Tsai 1 , Markus Steinberger 2 , Dawid Pajk 3 , Kari Pulli 4 1 Google, Inc. 2 TU Graz 3 NVIDIA 4 Light Research Story Collaborative Filtering - Aggregation Collaborative Filtering -

CS490W: What is Collaborative Filtering? Collaborative Filtering (CF): Making recommendation

Collaborative Filtering Presentation by Alex Hugger Filtering Documents Mittwoch, 28. April 2010

Training deep Autoencoders for collaborative filtering Oleksii Kuchaiev & Boris Ginsburg

Collaborative Filtering at Scale Recommender engines with Mahout and Hadoop Berlin Buzzwords Sean

Collaborative Filtering Practical Machine Learning, CS 294-34 Lester Mackey Based on slides by

Distributed Itembased Collaborative Filtering with Apache Mahout Sebastian Schelter

1 Collaborative Filtering Collaborative Filtering Maintain a database of many users

1 2 3 4 5 Second Project Implement collaborative filtering algorithm Apply to

Music recommenda tion System - Spotify Collaborative Filtering and Feedback System 1 Mithun

A Privacy-Protecting Architecture for Collaborative Filtering via Forgery and Suppression of

RLT: Residual-Loop Training in Collaborative Filtering for Combining Factorization and

CoFiGAN: Collaborative Filtering by Generative and Discriminative Training for One-Class

k -Reciprocal Nearest Neighbors Algorithm for One-Class Collaborative Filtering Wei Cai a , b , c

Problem Domain Collaborative filtering (CF)-based recommender systems (RS). Issue:

Applying Subsequence Matching to Collaborative Filtering Alejandro Bellog n Pablo S

Collaborative Filtering Alejandro Bellogn 1 , Jun Wang 2 , and Pablo Castells 1 1 Escuela

Algorithm Recommendation as Collaborative Filtering Mich` ele Sebag & Mustafa Misir &

Transfer Learning for Heterogeneous One-Class Collaborative Filtering Weike Pan , Mengsi Liu and

Neighborhood-Enhanced Transfer Learning for One-Class Collaborative Filtering Wanling Cai 1 , 2 ,

Collaborative Filtering Radek Pel anek Notes on Lecture the most technical lecture of the

Collaborative Filtering: basic ideas (slides based on chapter 2 of Programming Collective

Nearest-Biclusters Collaborative Filtering Philadelphia, 20 August 2006 Speaker : Panagiotis

Boolean Matrix Factorisation for Collaborative Filtering: An FCA based approach Dmitry Ignatov

On the Effectiveness of Linear Models for One-Class Collaborative Filtering Suvash Sedhain 1,2 ,

Collaborative Filtering Yun-Ta Tsai 1 , Markus Steinberger 2 , Dawid - PowerPoint PPT Presentation

FastANN for High Quality Collaborative Filtering Yun-Ta Tsai 1 , Markus Steinberger 2 , Dawid Pajk 3 , Kari Pulli 4 1 Google, Inc. 2 TU Graz 3 NVIDIA 4 Light Research Story Collaborative Filtering - Aggregation Collaborative Filtering -

CS490W: What is Collaborative Filtering? Collaborative Filtering (CF): Making recommendation

Collaborative Filtering Presentation by Alex Hugger Filtering Documents Mittwoch, 28. April 2010

Training deep Autoencoders for collaborative filtering Oleksii Kuchaiev &amp; Boris Ginsburg

Collaborative Filtering at Scale Recommender engines with Mahout and Hadoop Berlin Buzzwords Sean

Collaborative Filtering Practical Machine Learning, CS 294-34 Lester Mackey Based on slides by

Distributed Itembased Collaborative Filtering with Apache Mahout Sebastian Schelter

1 Collaborative Filtering Collaborative Filtering Maintain a database of many users

1 2 3 4 5 Second Project Implement collaborative filtering algorithm Apply to

Music recommenda tion System - Spotify Collaborative Filtering and Feedback System 1 Mithun

A Privacy-Protecting Architecture for Collaborative Filtering via Forgery and Suppression of

RLT: Residual-Loop Training in Collaborative Filtering for Combining Factorization and

CoFiGAN: Collaborative Filtering by Generative and Discriminative Training for One-Class

k -Reciprocal Nearest Neighbors Algorithm for One-Class Collaborative Filtering Wei Cai a , b , c

Problem Domain Collaborative filtering (CF)-based recommender systems (RS). Issue:

Applying Subsequence Matching to Collaborative Filtering Alejandro Bellog n Pablo S

Collaborative Filtering Alejandro Bellogn 1 , Jun Wang 2 , and Pablo Castells 1 1 Escuela

Algorithm Recommendation as Collaborative Filtering Mich` ele Sebag &amp; Mustafa Misir &amp;

Transfer Learning for Heterogeneous One-Class Collaborative Filtering Weike Pan , Mengsi Liu and

Neighborhood-Enhanced Transfer Learning for One-Class Collaborative Filtering Wanling Cai 1 , 2 ,

Collaborative Filtering Radek Pel anek Notes on Lecture the most technical lecture of the

Collaborative Filtering: basic ideas (slides based on chapter 2 of Programming Collective

Nearest-Biclusters Collaborative Filtering Philadelphia, 20 August 2006 Speaker : Panagiotis

Boolean Matrix Factorisation for Collaborative Filtering: An FCA based approach Dmitry Ignatov

On the Effectiveness of Linear Models for One-Class Collaborative Filtering Suvash Sedhain 1,2 ,

Training deep Autoencoders for collaborative filtering Oleksii Kuchaiev & Boris Ginsburg

Algorithm Recommendation as Collaborative Filtering Mich` ele Sebag & Mustafa Misir &