MNIST Results
- I then tried it on the small MNIST task.
- This has 5620 items, and each is an 8x8 input.
- Trying to get just one neuron firing per instance, I decreased the
input strength (64 featues vs 4).
- Part of the problem was the system couldn't handle all 5620 inputs, so
I initially broke it into 20 parts, trained on one and tested on 19
(20 times).
- Even with this, it took a long time to run a test
- I did manage it with 10 folds, which did a bit better.
- If I increased the number of categorisation (and inhibition) neurons
from 100 to 200, it did a bit better. This did take the weekend for
a test (on my pretty weak PC).
| System | Train | Test |
|---|
| 20 Fold | 59.84% | 50.11% |
| 10 Fold | 60.05% | 52.98% |
| 10 Fold 200 Cat Neurons | 69.62% | 59.34% |
| Nearest Neighbor | | 98.25 |
- Nearest neighbor (on a two fold test) does a lot better, and takes
less than a second on the same PC.
