← Go back

Research Colloquium Presentation by N'Dah Jean Kouagou

4 months ago

Neural Logic Reasoning

N'Dah Jean Kouagou - ESR 11, Paderborn University

I presented the paper by Shaoyung, Hanxiong et al. titled neural logic reasoning. The paper was pulished in CIKM20, and it tackles the problem of building logic integrated neural networks to solve real world tasks. The authors proposed LINN---Logic Integrated Neural Network, a dynamic neural network that represents logic operations AND, OR, NOT as a two layer multi-layer perceptron, and vector-like learnable parameters for logic variables. The network was trained end-to-end via the minization of a task specific loss function together with regularizers, including logic regularizers. LINN was first evaluated on logic equations. To this end, thousands of logic variables were generated to form logic equations. The evaluation results showed that LINN outperformed state-of-the-art approaches, including CNN and Bi-LSTM, with an accuracy of up to 0.95 on the test data. Next, LINN was applied as a recommender system to predict relationships between users and items on ML-100k and Amazon Electronics. The results obtained showed that LINN was better than some previously existing approaches, including GRU4Rec and NARM, on the two datasets.

The slides of the presentation can be found here