Mapping Stacked Decision Forests toDeep and Sparse ConvolutionalNeural Networks for Semantic

Segmentation

Presented by Daniela MassicetiReading Group - January 2016

v2 – 22 December 2015

David L Richmond1, Dagmar Kainmueller1, Michael Y

Yang2, Eugene W Myers1, and Carsten Rother2 1MPI-CBG, 2TU Dresden

23/01/16 Daniela Massiceti – RG Jan 2016 2

Random Forests vs Neural Networks

• RFs:– Easy to train; multi-class; training requires little data– Non-differentiable; each tree trained greedily

• NNs:– End-to-end learning of all parameters; feature learning– Training requires a lot of data & time

• Goal: bridge gap between RFs & NNs

23/01/16 Daniela Massiceti – RG Jan 2016 4

Sethi and Welbl’s mapping

23/01/16 Daniela Massiceti – RG Jan 2016 5

Sethi and Welbl’s mapping

-1-1 +1

0 1 0 0

-ω+ω

y1

y2

23/01/16 Daniela Massiceti – RG Jan 2016 6

Sethi and Welbl’s mapping

● Replicate T times for T trees in the RF● Fully connect output layer

23/01/16 Daniela Massiceti – RG Jan 2016 7

Relationship between RFs and CNNs

● RF with contextual features = special case of CNN with sparseconvolutional kernels & no max pooling

● Difference:● RFs: pixel's feature vector is precomputed● CNNs: features learnt

23/01/16 Daniela Massiceti – RG Jan 2016 8

Mapping forward: RF Stack to Deep CNN

23/01/16 Daniela Massiceti – RG Jan 2016 9

Mapping back: Deep CNN to RFStack

● Map back #1

– Keep weights and bias between H1 and H2 fixed, and re-train bybackprop

– Trivial mapping back to original RFwith updated thresholds and leafdistributions

● Map back #2

– Distributed activation on H2 = manyleafs contributing to prediction

– New leaf distributions:

23/01/16 Daniela Massiceti – RG Jan 2016 10

Results● Kinect Body Part Classification

Depth input Ground truth Stacked RF

82%

Retrained CNN

91%

Groundtruth

Stacked RF

RetrainedCNN

23/01/16 Daniela Massiceti – RG Jan 2016 11

Results● Zebrafish Somite Classification

Input GT Stacked RF RetrainedCNN

Mapback #1 Mapback #2

Method RF FCN CNN MB1 MB2

Dicescore

0.60 0.18 0.66 0.59 0.63

23/01/16 Daniela Massiceti – RG Jan 2016 12

Conclusions & comments

● Mapping from RF to CNN and back again

– Works with minimal training data

– Distributed activation on H2 – soft DTs

– Hand-selected features – representation learning

– Weights between H1 and H2 fixed – Decision Jungles

● Methodology-focussed rather than results-focussed

● Map-backed RF – improved performance, fastevaluation at test time?