MemoryNetworks

• Jason Weston, Antoine Bordes, Sumit Chopra, Tomas Mikolov, Alexander M. Rush, "Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks", http://arxiv.org/abs/1502.05698
• Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, Rob Fergus, "End-To-End Memory Networks", http://arxiv.org/abs/1503.08895

Requirements

• PyTorch 0.3-
• Download the 20 QA bAbI tasks http://www.thespermwhale.com/jaseweston/babi/tasks_1-20_v1-2.tar.gz

Results

I tested my model with 10k dataset. () is original MemNNs's performance (1k, 3 hops, PE).

• Task 1: Acc 100.00% (99.9%)
• Task 2: Acc 97.78% (78.4%)
• Task 3: Acc 93.55% (35.8%)
• Task 4: Acc 78.63% (96.2%) ?
• Task 5: Acc 91.13% (85.9%)
• Task 6: Acc 93.55% (92.1%)
• Task 7: Acc 89.42% (78.4%)
• Task 8: Acc 95.56% (87.4%)
• Task 9: Acc 96.77% (76.7%)
• Task 10: Acc 87.90% (82.6%)
• Task 11: Acc 94.86% (95.7%)
• Task 12: Acc 100.00% (99.7%)
• Task 13: Acc 94.76% (90.1%)
• Task 14: Acc 100.00% (98.2%)
• Task 15: Acc 100.00% (100.0%)
• Task 16: Acc 48.39% (47.9%)
• Task 17: Acc 52.82% (49.9%)
• Task 18: Acc 56.65% (86.4%)
• Task 19: Acc 20.67% (12.6%)
• Task 20: Acc 100.00% (100.0%)

TODO

• Random noise (RN)
• Linear start (LS)
• joint training
• compare results with FAIR team (the performance of some tasks is very low)
• correct optimizer and learning rate