Refactored Deep Learning Ideas
- examples: standard examples including the spiral dataset and mnist
- flow/dsl: setup of the neural computation graph
- flow/computation: all possible computations:
- add
- mul
- relu
- softmax
- flow/optimization: sgd and adam
- putting all together in neural net class
You will need jblas
- Accuracy is 97.4% on MNIST
- Confusion
| label | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 969 | 0 | 3 | 0 | 0 | 2 | 8 | 1 | 1 | 0 |
| 1 | 1 | 1129 | 3 | 0 | 0 | 0 | 3 | 8 | 2 | 3 |
| 2 | 1 | 1 | 1006 | 3 | 5 | 0 | 2 | 9 | 4 | 1 |
| 3 | 0 | 1 | 6 | 997 | 1 | 13 | 1 | 5 | 19 | 7 |
| 4 | 0 | 0 | 2 | 2 | 960 | 2 | 2 | 4 | 6 | 7 |
| 5 | 1 | 0 | 0 | 1 | 0 | 864 | 11 | 0 | 7 | 1 |
| 6 | 3 | 1 | 2 | 0 | 2 | 3 | 930 | 0 | 1 | 1 |
| 7 | 0 | 1 | 7 | 3 | 0 | 1 | 0 | 981 | 2 | 1 |
| 8 | 1 | 2 | 2 | 2 | 0 | 4 | 1 | 1 | 924 | 1 |
| 9 | 4 | 0 | 1 | 2 | 14 | 3 | 0 | 19 | 8 | 987 |