| <Poster Width="1734" Height="2602"> |
| <Panel left="26" right="480" width="1671" height="396"> |
| <Text>Overview</Text> |
| <Text>DTF = Efficiently learnable non-parametric CRFs for discrete image labelling tasks</Text> |
| <Text>• All factors (unary, pairwise, higher-order) are represented by decision trees</Text> |
| <Text>• Decision trees are non-parametric</Text> |
| <Text>• Efficient training of millions of parameters using pseudo-likelihood</Text> |
| </Panel> |
|
|
| <Panel left="26" right="879" width="1677" height="578"> |
| <Text>Formally</Text> |
| <Figure left="26" right="1002" width="299" height="279" no="1" OriWidth="0.180507" OriHeight="0.131907 |
| " /> |
| <Text>Graphical Model:</Text> |
| <Text>Factor types</Text> |
| <Figure left="331" right="1028" width="303" height="243" no="2" OriWidth="0.196078" OriHeight="0.118093 |
| " /> |
| <Text>Factor Graph</Text> |
| <Text>Energy</Text> |
| <Text>Energy linear in w</Text> |
| <Text>Example pairwise factor</Text> |
| <Figure left="1382" right="1027" width="326" height="336" no="3" OriWidth="0.155133" OriHeight="0.111408 |
| " /> |
| </Panel> |
|
|
| <Panel left="33" right="1458" width="1666" height="347"> |
| <Text>Special Cases</Text> |
| <Text>• Unary factors only = Decision Forest, with learned leaf node distributions</Text> |
| <Text>Zero-depth trees (pairwise factors) = MRF</Text> |
| <Text>• Conditional (pairwise factors) = CRF</Text> |
| <Figure left="1289" right="1563" width="270" height="169" no="4" OriWidth="0.190888" OriHeight="0.114973 |
| " /> |
| <Figure left="784" right="1635" width="383" height="161" no="5" OriWidth="0.194925" OriHeight="0.114973 |
| " /> |
| </Panel> |
|
|
| <Panel left="26" right="1805" width="1671" height="399"> |
| <Text>Algorithm - Overview</Text> |
| <Text>Training</Text> |
| <Text>1.Define connective structure (factor types)</Text> |
| <Text>2.Train all decision trees (split functions) separately</Text> |
| <Text>3.Jointly optimize all weights</Text> |
| <Text>Testing (2 options)</Text> |
| <Text>•“Unroll” factor graph:</Text> |
| <Text>run: BP, TRW, QPBO, etc.</Text> |
| <Text>•Don’t “unroll” factor graph:</Text> |
| <Text>run Gibbs Sampling; Simulated Annealing</Text> |
| </Panel> |
|
|
| <Panel left="24" right="2203" width="1672" height="352"> |
| <Text>Training of weights “w”</Text> |
| <Text>•Maximum Pseudo-Likelihood training, convex optimization problem</Text> |
| <Text>Converges in practice after 150-200 L-BFGS iterations</Text> |
| <Text>Efficient even for large graphs (e.g. 12 connected, 1.47M weights, 22mins)</Text> |
| <Text>•Is parallel on the variable level</Text> |
| <Text>•Variable sub-sampling possible</Text> |
| <Text>Code will be made available next month!</Text> |
| </Panel> |
|
|
| </Poster> |
|
|
