Neural networks are clean

stage2.block7:481 → stage2.block8:63: mean routed attribution +0.102959; 8/8 examples share the mean sign.stage2.block7:766 → stage2.block8:63: mean routed attribution -0.028816; 8/8 examples share the mean sign.stage2.block7:481 → stage2.block8:792: mean routed attribution +0.059980; 7/8 examples share the mean sign.stage2.block7:766 → stage2.block8:792: mean routed attribution -0.104955; 8/8 examples share the mean sign.stage2.block8:63 → stage3.block0:883: mean routed attribution +0.335768; 8/8 examples share the mean sign.stage2.block8:792 → stage3.block0:883: mean routed attribution -0.197188; 6/8 examples share the mean sign.stage2.block8:63 → stage3.block0:1446: mean routed attribution +0.043502; 8/8 examples share the mean sign.stage2.block8:792 → stage3.block0:1446: mean routed attribution -0.099039; 8/8 examples share the mean sign.stage3.block0:883 → stage3.block1:1316: mean routed attribution +3.123215; 8/8 examples share the mean sign.stage3.block0:1446 → stage3.block1:1316: mean routed attribution +1.058888; 8/8 examples share the mean sign.stage3.block1:1316 → stage3.block2:1245: mean routed attribution +23.535583; 8/8 examples share the mean sign.stage2.block7:481; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Jigsaws & battlementsOptimized image, negative pole#1 jigsaw puzzle, activation -4.65625#2 jigsaw puzzle, activation -4.625#3 castle, activation -4.375#4 jigsaw puzzle, activation -4.375#481 ↗negative polestage2.block7:766; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Siding & windowsOptimized image, negative pole#1 picket fence, activation -6.9375#2 bookshop, activation -6.59375#3 fire engine, activation -6.53125#4 car wheel, activation -6.46875#766 ↗negative polestage2.block8:63; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Stone wallsOptimized image, negative pole#1 cliff dwelling, activation -7.59375#2 half track, activation -7.46875#3 safe, activation -7.1875#4 bell cote, activation -7.09375#63 ↗negative polestage2.block8:792; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Roofs & archesOptimized image, positive pole#1 tile roof, activation 3.9375#2 traffic light, activation 3.71875#3 palace, activation 3.5625#4 tile roof, activation 3.48438#792 ↗positive polestage3.block0:883; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Bell towersOptimized image, positive pole#1 monastery, activation 10.375#2 church, activation 10.1875#3 church, activation 9.3125#4 bell cote, activation 9.25#883 ↗positive polestage3.block0:1446; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.CastlesOptimized image, negative pole#1 castle, activation -9.25#2 castle, activation -7.96875#3 castle, activation -7.875#4 castle, activation -7.5#1446 ↗negative polestage3.block1:1316; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Monasteries & castlesOptimized image, positive pole#1 monastery, activation 23.25#2 monastery, activation 20.75#3 monastery, activation 20#4 monastery, activation 18.375#1316 ↗positive polestage3.block2:1245; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Monastery-responsive unitOptimized image, positive pole#1 monastery, activation 113.5#2 monastery, activation 111#3 monastery, activation 107.5#4 monastery, activation 101#1245 ↗positive poleBlock 14Block 15Block 16Block 17Block 18stage2.block7:481 → stage2.block8:63: mean routed attribution +0.102959; 8/8 examples share the mean sign.stage2.block7:766 → stage2.block8:63: mean routed attribution -0.028816; 8/8 examples share the mean sign.stage2.block7:481 → stage2.block8:792: mean routed attribution +0.059980; 7/8 examples share the mean sign.stage2.block7:766 → stage2.block8:792: mean routed attribution -0.104955; 8/8 examples share the mean sign.stage2.block8:63 → stage3.block0:883: mean routed attribution +0.335768; 8/8 examples share the mean sign.stage2.block8:792 → stage3.block0:883: mean routed attribution -0.197188; 6/8 examples share the mean sign.stage2.block8:63 → stage3.block0:1446: mean routed attribution +0.043502; 8/8 examples share the mean sign.stage2.block8:792 → stage3.block0:1446: mean routed attribution -0.099039; 8/8 examples share the mean sign.stage3.block0:883 → stage3.block1:1316: mean routed attribution +3.123215; 8/8 examples share the mean sign.stage3.block0:1446 → stage3.block1:1316: mean routed attribution +1.058888; 8/8 examples share the mean sign.stage3.block1:1316 → stage3.block2:1245: mean routed attribution +23.535583; 8/8 examples share the mean sign.stage2.block7:481; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Jigsaws & battlementsOptimized image, negative pole#1 jigsaw puzzle, activation -4.65625#2 jigsaw puzzle, activation -4.625#3 castle, activation -4.375#4 jigsaw puzzle, activation -4.375#481 ↗negative polestage2.block7:766; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Siding & windowsOptimized image, negative pole#1 picket fence, activation -6.9375#2 bookshop, activation -6.59375#3 fire engine, activation -6.53125#4 car wheel, activation -6.46875#766 ↗negative polestage2.block8:63; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Stone wallsOptimized image, negative pole#1 cliff dwelling, activation -7.59375#2 half track, activation -7.46875#3 safe, activation -7.1875#4 bell cote, activation -7.09375#63 ↗negative polestage2.block8:792; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Roofs & archesOptimized image, positive pole#1 tile roof, activation 3.9375#2 traffic light, activation 3.71875#3 palace, activation 3.5625#4 tile roof, activation 3.48438#792 ↗positive polestage3.block0:883; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Bell towersOptimized image, positive pole#1 monastery, activation 10.375#2 church, activation 10.1875#3 church, activation 9.3125#4 bell cote, activation 9.25#883 ↗positive polestage3.block0:1446; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.CastlesOptimized image, negative pole#1 castle, activation -9.25#2 castle, activation -7.96875#3 castle, activation -7.875#4 castle, activation -7.5#1446 ↗negative polestage3.block1:1316; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Monasteries & castlesOptimized image, positive pole#1 monastery, activation 23.25#2 monastery, activation 20.75#3 monastery, activation 20#4 monastery, activation 18.375#1316 ↗positive polestage3.block2:1245; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Monastery-responsive unitOptimized image, positive pole#1 monastery, activation 113.5#2 monastery, activation 111#3 monastery, activation 107.5#4 monastery, activation 101#1245 ↗positive pole

Ladybug

Block 14Block 15Block 16Block 17Block 18stage2.block7:751 → stage2.block8:630: mean routed attribution -0.008386; 7/8 examples share the mean sign.stage2.block7:42 → stage2.block8:630: mean routed attribution -0.004706; 5/8 examples share the mean sign.stage2.block7:751 → stage2.block8:592: mean routed attribution -0.002110; 4/8 examples share the mean sign.stage2.block7:42 → stage2.block8:592: mean routed attribution -0.001332; 5/8 examples share the mean sign.stage2.block8:630 → stage3.block0:1436: mean routed attribution +0.003532; 7/8 examples share the mean sign.stage2.block8:592 → stage3.block0:1436: mean routed attribution -0.002162; 7/8 examples share the mean sign.stage2.block8:630 → stage3.block0:566: mean routed attribution -0.014763; 8/8 examples share the mean sign.stage2.block8:592 → stage3.block0:566: mean routed attribution -0.012087; 7/8 examples share the mean sign.stage3.block0:1436 → stage3.block1:1519: mean routed attribution +0.356008; 8/8 examples share the mean sign.stage3.block0:566 → stage3.block1:1519: mean routed attribution -0.277160; 7/8 examples share the mean sign.stage3.block1:1519 → stage3.block2:34: mean routed attribution +7.082140; 8/8 examples share the mean sign.stage2.block7:751; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Spots & holesOptimized image, positive pole#1 ocarina, activation 4.90625#2 ladybug, activation 4.84375#3 saltshaker, activation 4.34375#4 ocarina, activation 4.1875#751 ↗positive polestage2.block7:42; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Small colored formsOptimized image, negative pole#1 hip, activation -10.75#2 American chameleon, activation -10.3125#3 goldfish, activation -10.125#4 monarch, activation -9.875#42 ↗negative polestage2.block8:630; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Small red dottedOptimized image, negative pole#1 ladybug, activation -5.125#2 Chihuahua, activation -4.3125#3 curly-coated retriever, activation -3.6875#4 Great Pyrenees, activation -3.57812#630 ↗negative polestage2.block8:592; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Red dots and whiteOptimized image, negative pole#1 bulbul, activation -8.5625#2 hot pot, activation -8.125#3 ice cream, activation -7.0625#4 drake, activation -6.96875#592 ↗negative polestage3.block0:1436; positive activation pole. Four selected positive responses, at ranks 2, 8, 76, and 256. Short label is a visual interpretation.Red dots & circlesOptimized image, positive pole#2 ladybug, activation 5.96875#8 tree frog, activation 4.84375#76 ladybug, activation 3.4375#256 ladybug, activation 2.70312#1436 ↗positive polestage3.block0:566; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Spotted red surfacesOptimized image, positive pole#1 agaric, activation 10.3125#2 bottlecap, activation 9.5#3 agaric, activation 8.375#4 mushroom, activation 8.3125#566 ↗positive polestage3.block1:1519; negative activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Beetle-responsive unitOptimized image, negative pole#1 ladybug, activation -22.125#2 leaf beetle, activation -21.375#3 leaf beetle, activation -21#4 leaf beetle, activation -20.5#1519 ↗negative polestage3.block2:34; positive activation pole. Four consecutive natural examples at this pole. Short label is a visual interpretation.Ladybug-responsive unitOptimized image, positive pole#1 ladybug, activation 21.5#2 ladybug, activation 15.25#3 ladybug, activation 14#4 ladybug, activation 12.1875#34 ↗positive poleBlock 14Block 15Block 16Block 17Block 18stage2.block7:751 → stage2.block8:630: mean routed attribution -0.008386; 7/8 examples share the mean sign.stage2.block7:42 → stage2.block8:630: mean routed attribution -0.004706; 5/8 examples share the mean sign.stage2.block7:751 → stage2.block8:592: mean routed attribution -0.002110; 4/8 examples share the mean sign.stage2.block7:42 → stage2.block8:592: mean routed attribution -0.001332; 5/8 examples share the mean sign.stage2.block8:630 → stage3.block0:1436: mean routed attribution +0.003532; 7/8 examples share the mean sign.stage2.block8:592 → stage3.block0:1436: mean routed attrib