1 of 21

Lab 1: Drug Concentration

1

2 of 21

Lab 1: Drug Concentration

  •  

2

 

 

3 of 21

Same MLP, Different Constraint Strategies

  • Shared neural network
    • 2 hidden layers
    • 32 neurons per layer
    • Tanh activation

3

4 of 21

Baseline - Learn from data only

4

 

 

5 of 21

Soft - Add constraint penalties

5

 

 

 

 

6 of 21

Hard - Transform the model output

6

 

 

 

7 of 21

Result Comparison

7

8 of 21

Lab 2: SIR Conservation

8

9 of 21

Lab 2: SIR Conservation

  •  

9

 

 

 

10 of 21

Soft - Penalize Constraint Violations

10

 

 

 

11 of 21

Projection - Correct the Raw Output

11

 

 

 

12 of 21

Hard - Apply Softmax to the Output

12

 

 

13 of 21

Result Comparison

13

14 of 21

Lab 3: Pairwise Interaction

14

15 of 21

Lab 3: Pairwise Interaction

  •  

15

 

 

 

 

 

Rotation / Translation

16 of 21

Test Equivalent Configurations

16

 

17 of 21

Raw Coordinates - No Prior Knowledge

17

18 of 21

Data Augmentation

18

 

 

 

19 of 21

Augmented + Pooling

19

 

Mean

20 of 21

Invariant Feature

20

 

 

21 of 21

Result Comparison

21