December 3, 2017
Members of the MultiComp Lab had four papers accepted at the Association of Advancement of Artificial Intelligence (AAAI 2018 https://aaai.org/Conferences/AAAI-18/) in New Orleans, Lousiana, USA.
“Memory Fusion Network for Multi-view Sequential Learning” by Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria and Louis-Philippe Morency studies the synchronization of multi-view sequences using a multi-view gated memory. The model achieves state of the art results on 7 publicly available datasets spanning multimodal sentiment analysis, emotion recognition and personality traits recognition.
“Multi-attention Recurrent Network for Human Communication Comprehension” by Amir Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, Prateek Vij and Louis-Philippe Morency introduces the Multi-attention Recurrent Network, a neural framework that models both view-specific and cross-view dynamics in multimodal human communication.
“Lattice Recurrent Unit: Improving Convergence and Statistical Efficiency for Sequence Modeling” by Chaitanya Ahuja, and Louis-Philippe Morency introduce a new recurrent unit which has two distinct flows of information along time and depth. This reduces the effect of vanishing gradients along the depth as well, while giving a boost to the computational and statistical efficiency.
“Using Syntax to Ground Referring Expressions in Natural Images” by Volkan Cirik, Taylor Berg-Kirkpatrick, and Louis-Philippe Morency explores the use of syntax for referring expression recognition – the task of identifying the target object in an image referred to by a natural language expression. The proposed model GroundNet effectively integrates syntax to achieve the balance between accurately identifying both the target object and supporting objects.