Relational memory lets a recurrent model reason about relationships among stored items instead of merely recalling them. In the RMC, memory slots and attention support comparisons across items: finding the nth-farthest vector requires comparing distances and ranking them. On this task, LSTMs and DNCs stayed below 30% accuracy, while the RMC reached about 91%; results were less consistent with higher-dimensional vectors. section 7.4.1
On program-evaluation tasks derived from Learning to Execute, the RMC matched or exceeded strong baselines. It achieved near-perfect per-character accuracy on simpler operations such as addition, copying, and reversing. section 7.4.1
Table 7.4 Accuracy on program evaluation and memorization tasks
|
Model
|
Add
|
Control
|
Program
|
Copy
|
Reverse
|
Double
|
|---|---|---|---|---|---|---|
|
LSTM
|
99.8
|
97.4
|
66.1
|
99.8
|
99.7
|
99.7
|
|
EntNet
|
98.4
|
98.0
|
73.4
|
91.8
|
100.0
|
62.3
|
|
DNC
|
99.4
|
83.8
|
69.5
|
100.0
|
100.0
|
100.0
|
|
RMC
|
99.9
|
99.6
|
79.0
|
100.0
|
100.0
|
99.8
|
Compare reported program-evaluation accuracies for the RMC and recurrent baselines. section 7.4.1
