This is an implementation of Memory Networks in Theano. The dataset can be found here.
Below are the results obtained from the current implementation on the set of toy tasks. Performance is better than LSTMs in most of the tasks, but it is still not quite at the level as the MemNN results reported in the original paper.
| Task# | N-gram Classifier | LSTM | MemNN (Weston 2014) | This Repo |
|---|---|---|---|---|
| 1 | 36 | 50 | 100 | 72 |
| 2 | 2 | 20 | 100 | 63 |
| 3 | 7 | 20 | 20 | 8 |
| 4 | 50 | 61 | 71 | 64 |
| 5 | 20 | 70 | 83 | 80 |
| 6 | 49 | 48 | 47 | 45 |
| 7 | 52 | 49 | 68 | N/A |
| 8 | 40 | 45 | 77 | N/A |
| 9 | 62 | 64 | 65 | 67 |
| 10 | 45 | 54 | 59 | 54 |
| 11 | 29 | 72 | 100 | 46 |
| 12 | 9 | 74 | 100 | 46 |
| 13 | 26 | 94 | 100 | 39 |
| 14 | 19 | 27 | 99 | 100 |
| 15 | 20 | 21 | 74 | 100 |
| 16 | 43 | 23 | 27 | 20 |
| 17 | 46 | 51 | 54 | 46 |
| 18 | 52 | 52 | 57 | 53 |
| 19 | 0 | 8 | 0 | N/A |
| 20 | 76 | 91 | 100 | 100 |
| Mean | 34 | 49 | 75 | 58.4 |