Amortized Monte Carlo Integration Adam Goli ski*, Frank Wood, Tom - PowerPoint PPT Presentation

Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* 11/06/19

Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* 2

CPXicbVBNS+tAFJ34/epX9S3dDBah3ZREBQURxMeDt1SwWmhCmExv7OBkEmZupCHmj7nxP7ydOzcufMjbunVau/DrwMCZc85l5p4ok8Kg6947U9Mzs3PzCz9qi0vLK6v1tfVzk+a Q4enMtXdiBmQ kEHBUroZhpYEkm4iK5+jfyLa9BGpOoMiwyChF0qEQvO0Eph/cxPGA6iqPxdhWXWHNIbWrSqXtwcHlAfB4CsFdBD6guF4TjKmSy7FX0foHbupmjZKwyx7FfDsN5w2+4Y9CvxJqRBJjgJ63/9fsrzB Ry YzpeW6GQck0Ci6hqvm5gYzxK3YJPUsVS8AE5Xj7im5ZpU/jVNujkI7V9xMlS4wpksgmRwuYz95I/M7r5RjvB6VQWY6g+NtDcS4p nRUJe0LDRxlYQnjWti/Uj5gmnG0hd sCd7nlb+S8+2 t9PePt1tHB1P6lg G2STNIlH9sgR+UNOSIdwckseyBP5 9w5j86z8/8tOuVMZn6SD3BeXgEhva3/</latexit> <latexit sha1_base64="5pFQvrg9hoF5zxocjJwc/T KI9Q=">A Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* E p ( x | y ) [ f ( x ; θ )] 3

<latexit sha1_base64="5pFQvrg9hoF5zxocjJwc/T KI9Q=">A CPXicbVBNS+tAFJ34/epX9S3dDBah3ZREBQURxMeDt1SwWmhCmExv7OBkEmZupCHmj7nxP7ydOzcufMjbunVau/DrwMCZc85l5p4ok8Kg6947U9Mzs3PzCz9qi0vLK6v1tfVzk+a Q4enMtXdiBmQ kEHBUroZhpYEkm4iK5+jfyLa9BGpOoMiwyChF0qEQvO0Eph/cxPGA6iqPxdhWXWHNIbWrSqXtwcHlAfB4CsFdBD6guF4TjKmSy7FX0foHbupmjZKwyx7FfDsN5w2+4Y9CvxJqRBJjgJ63/9fsrzB Ry YzpeW6GQck0Ci6hqvm5gYzxK3YJPUsVS8AE5Xj7im5ZpU/jVNujkI7V9xMlS4wpksgmRwuYz95I/M7r5RjvB6VQWY6g+NtDcS4p nRUJe0LDRxlYQnjWti/Uj5gmnG0hd sCd7nlb+S8+2 t9PePt1tHB1P6lg G2STNIlH9sgR+UNOSIdwckseyBP5 9w5j86z8/8tOuVMZn6SD3BeXgEhva3/</latexit> Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* E p ( x | y ) [ f ( x ; θ )] AMCI = novel estimator + amortization objectives 4

<latexit sha1_base64="5pFQvrg9hoF5zxocjJwc/T KI9Q=">A CPXicbVBNS+tAFJ34/epX9S3dDBah3ZREBQURxMeDt1SwWmhCmExv7OBkEmZupCHmj7nxP7ydOzcufMjbunVau/DrwMCZc85l5p4ok8Kg6947U9Mzs3PzCz9qi0vLK6v1tfVzk+a Q4enMtXdiBmQ kEHBUroZhpYEkm4iK5+jfyLa9BGpOoMiwyChF0qEQvO0Eph/cxPGA6iqPxdhWXWHNIbWrSqXtwcHlAfB4CsFdBD6guF4TjKmSy7FX0foHbupmjZKwyx7FfDsN5w2+4Y9CvxJqRBJjgJ63/9fsrzB Ry YzpeW6GQck0Ci6hqvm5gYzxK3YJPUsVS8AE5Xj7im5ZpU/jVNujkI7V9xMlS4wpksgmRwuYz95I/M7r5RjvB6VQWY6g+NtDcS4p nRUJe0LDRxlYQnjWti/Uj5gmnG0hd sCd7nlb+S8+2 t9PePt1tHB1P6lg G2STNIlH9sgR+UNOSIdwckseyBP5 9w5j86z8/8tOuVMZn6SD3BeXgEhva3/</latexit> Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* E p ( x | y ) [ f ( x ; θ )] AMCI = novel estimator + amortization objectives 5

DaXicdVJdixMxFM20fqz1a6svoi/BYmlBSmcVFKGwKAs+6Qq2u9CMQyaTacNOMrNJxraE/C3/i+Crv gnzHQGut3WCzPcnHtP7rn3JspTpvRw+N rNG/cvHX74E7r7r37Dx4eth9NVFZIQsckSzN5HmF UyboWDOd0vNcUsyjlJ5F x/K+Nl3KhXLxFe9ymnA8UywhBGsHRS2vc/dLuJYz6PInNjQoJz1ln07Tdw/gF2E81xmS4gWLKZzrA3ihYXvRhAlEhPjW/PJQqQKHhox8u238piglCa6t3SQRZLN5roPF+UBItTqdl16jgVEmi61Wcyp NDCZRVXjMNLV/l TRhVZ yo3TutudwDrktsGp uCXcNOeX/63ZDldxMsNxLrvRcTXLc0O8vYOgH5T Cw85wMFwb3HX82umA2k7dCjwUZ6TgVGiSYqWm/jDXgcFSM5JS20KFojkmF3hGp84VmFMVmPXqLXzhkBgm XSf0HCNXmUYzJVa8chl qLV9VgJ7otNC528DQwTeaGpIFWhpEihzmD5jmDMJCU6XTkHE8mcVkjm2E1Hu9e2VSVWpT Xh6ALknGORWzQiTWbRZRD86+PaNeZHA38V4OjL687x+/r8R2AZ+A56AEfvAH 4CM4BWNAvB/eL+ 396fxt9luPmk+rVIbXs15DLas2fkHiDwaEg= </latexit> <latexit sha1_base64="OGZnm3y0pBK8vdPyYZ3BhoGjD/M=">A Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* Importance Sampling (IS) N X µ := 1 E π ( x ) [ f ( x )] ≈ b f ( x n ) w n N n =1 where x n ∼ q ( x ) , w n = π ( x n ) q ( x n ) E [ b µ ] = E π ( x ) [ f ( x )] µ ] = Var[ f ( x 1 ) w 1 ] Var[ b N 6

<latexit sha1_base64="MBAE4aFP6XlFwELjzqAbmxnjrcQ=">A B/3icbVDLSgMxFM3UV62vUcGNm2AR6qbMVEGXRTcuK9gHdIaS TNtaCaJSUYstQt/xY0LRdz6G+78G9N2Ftp64MLJOfeSe08kGdXG876d3NLy upafr2wsbm1vePu7jW0SBUmdSyYUK0IacIoJ3VD SMtqQhKIka 0eBq4jfvidJU8FszlCRMUI/TmGJkrNRxD+5KDyeBVEIaAQNJ7Su21XGLXtmbAi4SPyNFkKHWcb+CrsBpQrjBDGnd9j1pwhFShmJGxoUg1UQiPEA90raUo4TocDTdfwyPrdKFsVC2uIFT9f EC VaD5PIdibI9PW8NxH/89qpiS/CEeUyNYTj2UdxyqC9dRIG7FJFsGFDSxBW1O4KcR8phI2NrGBD8OdPXiSNStk/LVduzorVy OPDgER6AEfHAOquAa1EAdYPAInsEreHOenBfn3fmYteacbGYf/IHz+QNu 5US</latexit> <latexit sha1_base64="hWekqsPQwtoW6yfEV/LWKl+0y2c=">A B8XicbVA9TwJBEJ3DL8Qv1NJmIzHBhtyhiZ EG0tMBIxwIXvLHGzY27vs7hkJ4V/YWGiMrf/Gzn/jAlco+J Xt6bycy8IBFcG9f9dnIrq2vrG/nNwtb2zu5ecf+gqeNUMWywWMTqPqAaBZfYMNwIvE8U0igQ2AqG1 O/9YhK81jemVGCfkT7koecUWOlh7D8dEo6fSRut1hyK+4MZJl4GSlBhnq3+NXpxSyNUBomqNZtz02MP6bKcCZwUuikGhPKhrSPbUsljVD749nFE3JilR4JY2VLGjJTf0+Ma T1KApsZ0TNQC96U/E/r52a8NIfc5mkBiWbLwpTQUxMpu+THlfIjBhZQpni9lbCBlR ZmxIBRuCt/jyMmlWK95ZpXp7XqpdZXHk4QiOoQweXEANbqAODWAg4Rle4c3Rzovz7nzMW3NONnMIf+B8/gDBHY+p</latexit> Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* Importance Sampling (IS) When and q ( x ) ∝ π ( x ) f ( x ) f ( x ) ≥ 0 yields an exact estimate using a single sample 7

tvzs2xs7oUxkbR9yC8cvXa9RsrNwe3bt+5u7q2fm/fVI1mfMyqstKHGTW8FIqPrbAlP6w1pzIr+UF28rqtH3zk2ohKvbOzmieSHilRCEYtUOl6sPHh/ZPN6RZ+TGpd1b CBWSkFu2bDIiQsAU3+G/XS0wKTZn73+UdEcpi8nBhEhPLp9blfurxpROS2uMsc7s+dT03aeuJB0 Aabx1msat2jzpP9G1Qupdt06L8NKWf6qQLbBLlbtNl1XTtWE0ir AF0HcgyHqYw+MDUhesUZyZVlJjZnEUW0TR7UVrOR+QBrDa8pO6BGfAFRUcpO47kA93gAmx0Wl4QFvO3ZxwlFpzExm0NlubM7XWvKy2qSx YvECVU3lis2FyqaEsO5t7cD50JzZs ZAMq0gF0xO6ZgpIU7dEYlN+1q8B+Kn7JKSqpyR3YXDfZgWnzeo tgf3sUPx1tv3023HnV27eCHqBHaBPF6DnaQW/QHhojFnwKPgdfgq/ht/BH+DP8NW8Ng37mPjoT4e8/d4YE3g= </latexit> B/3icbVDLSgMxFM3UV62vUcGNm2AR6qbMVEGXRTcuK9gHdIaS TNtaCaJSUYstQt/xY0LRdz6G+78G9N2Ftp64MLJOfeSe08kGdXG876d3NLy upafr2wsbm1vePu7jW0SBUmdSyYUK0IacIoJ3VD SMtqQhKIka 0eBq4jfvidJU8FszlCRMUI/TmGJkrNRxD+5KDyeBVEIaAQNJ7Su21XGLXtmbAi4SPyNFkKHWcb+CrsBpQrjBDGnd9j1pwhFShmJGxoUg1UQiPEA90raUo4TocDTdfwyPrdKFsVC2uIFT9f EC VaD5PIdibI9PW8NxH/89qpiS/CEeUyNYTj2UdxyqC9dRIG7FJFsGFDSxBW1O4KcR8phI2NrGBD8OdPXiSNStk/LVduzorVy DP3icdVL btQwFHXCqwyPtrBkYxgVtSxGSUGCDVIFqsSySExbaRwix3Faq7ETbIfOyPIf8DVsYcNn8AXsEFt23GQCTB9zpUTn <latexit sha1_base64="A/7/9jd65UHiJ4a3Da8It+sCfpY=">A Xt6bycy8IBFcG9f9dnIrq2vrG/nNwtb2zu5ecf+gqeNUMWywWMTqPqAaBZfYMNwIvE8U0igQ2AqG1 O/9YhK81jemVGCfkT7koecUWOlh7D8dEo6fSRut1hyK+4MZJl4GSlBhnq3+NXpxSyNUBomqNZtz02MP6bKcCZwUuikGhPKhrSPbUsljVD749nFE3JilR4JY2VLGjJTf0+Ma T1KApsZ0TNQC96U/E/r52a8NIfc5mkBiWbLwpTQUxMpu+THlfIjBhZQpni9lbCBlR ZmxIBRuCt/jyMmlWK95ZpXp7XqpdZXHk4QiOoQweXEANbqAODWAg4Rle4c3Rzovz7nzMW3NONnMIf+B8/gDBHY+p</latexit> B8XicbVA9TwJBEJ3DL8Qv1NJmIzHBhtyhiZ EG0tMBIxwIXvLHGzY27vs7hkJ4V/YWGiMrf/Gzn/jAlco+J <latexit sha1_base64="hWekqsPQwtoW6yfEV/LWKl+0y2c=">A OPDgER6AEfHAOquAa1EAdYPAInsEreHOenBfn3fmYteacbGYf/IHz+QNu 5US</latexit> <latexit sha1_base64="MBAE4aFP6XlFwELjzqAbmxnjrcQ=">A Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* Importance Sampling (IS) When and q ( x ) ∝ π ( x ) f ( x ) f ( x ) ≥ 0 yields an exact estimate using a single sample q ∗ ( x ) ∝ f ( x ) π ( x ) f ( x ) π ( x ) f ( x ) π ( x ) = ⇒ q ∗ ( x ) = f ( x ) π ( x )d x = R E π ( x ) [ f ( x )] f ( x 1 ) w 1 = f ( x 1 ) π ( x 1 ) π ( x 1 ) q ∗ ( x 1 ) = f ( x 1 ) = E π ( x ) [ f ( x )] f ( x 1 ) π ( x 1 ) E π ( x ) [ f ( x )] 8

<latexit sha1_base64="PUc2 1srs+tp PDh784N GfpuCA=">A DOHicfVJdb9MwFHXC1xhfGzwiIYuKqZVQlQyk7WXSBJrE xoS3SbVoXJcp7UWO57t0EbGj/waXuGFf8Ib 4hXfgFOGomunbhSoutz 825PrmpzJk2UfQjCK9dv3Hz1sbtzTt3791/sLX98EQXpSJ0QIq8UGcp1jRng 4M zk9k4pinub0ND1/Xd P 1KlWSHem0rShO JYBkj2HhotB08QRybaZraIzeysjuHn2DVc8OsO+8lcOcAZQoTu8Jp61B2K0/3RHcF418RIgR3IMJSqmIO2y8279jZtw7pko+sOIjdB3+CGcp ZrpzDzmk2GRqenBWH9x/mhaEWqjpq+U04/CintKDswXY3kYuKTyHVSvi7MW6sht daJ+1ARcT+I26YA2jr2jARoXpORUGJ jrYdxJE1isTKM5NRtolJTick5ntChTwXmVCe2+ZMOPvPIG aF8o8wsEGXOyzmWlc89czacL1aq8GrasPSZPuJZUKWhgqyEMrKHJoC1msBx0xRYvLKJ5go5meFZIq9W8YvzyWVsa5H8/cQdEYKzrEYW3S0vAG1afGqRevJyW4/ftHf eyc/iqtW8DPAZPQRfEYA8cgjfgGAwACT4HX4Kvwbfwe/gz/BX+XlD oO15BC5F+Ocvo AKjA= </latexit> Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* Self-Normalized Importance Sampling (SNIS) E p ( x | y ) [ f ( x )] 9

eyc/iqtW8DPAZPQRfEYA8cgjfgGAwACT4HX4Kvwbfwe/gz/BX+XlD oO15BC5F+Ocvo AKjA= </latexit> <latexit sha1_base64="PUc2 1srs+tp PDh784N GfpuCA=">A DOHicfVJdb9MwFHXC1xhfGzwiIYuKqZVQlQyk7WXSBJrE xoS3SbVoXJcp7UWO57t0EbGj/waXuGFf8Ib 4hXfgFOGomunbhSoutz 825PrmpzJk2UfQjCK9dv3Hz1sbtzTt3791/sLX98EQXpSJ0QIq8UGcp1jRng 4M zk9k4pinub0ND1/Xd P 1KlWSHem0rShO JYBkj2HhotB08QRybaZraIzeysjuHn2DVc8OsO+8lcOcAZQoTu8Jp61B2K0/3RHcF418RIgR3IMJSqmIO2y8279jZtw7pko+sOIjdB3+CGcp ZrpzDzmk2GRqenBWH9x/mhaEWqjpq+U04/CintKDswXY3kYuKTyHVSvi7MW6sht daJ+1ARcT+I26YA2jr2jARoXpORUGJ jrYdxJE1isTKM5NRtolJTick5ntChTwXmVCe2+ZMOPvPIG aF8o8wsEGXOyzmWlc89czacL1aq8GrasPSZPuJZUKWhgqyEMrKHJoC1msBx0xRYvLKJ5go5meFZIq9W8YvzyWVsa5H8/cQdEYKzrEYW3S0vAG1afGqRevJyW4/ftHf Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* Self-Normalized Importance Sampling (SNIS) E p ( x | y ) [ f ( x )] = E p ( x ) [ f ( x ) p ( y | x )] E p ( x ) [ p ( y | x )] P N 1 n =1 f ( x n ) w n N ≈ P N 1 n =1 w n N x n ∼ q ( x ) w n = p ( x n , y ) q ( x n ) 10

Amortized Monte Carlo Integration Adam Goli ń ski*, Frank Wood, Tom Rainforth* 10 0 Relative Error 10 − 2 10 − 4 Traditional approach Its error lower bound 10 − 6 AMCI 10 1 10 2 10 3 10 4 10 5 Number of samples 11