🧑‍🔬 Faisal Ahmed

Faisal Ahmed's papers here, from 2017 to 2019, span two areas of machine learning for language: dialogue systems and vision and language. Both works ask how one model can be trained as a whole, without hand-built intermediate steps that block learning or tie the system to a single task. The methods are neural throughout: reinforcement learning for conversational agents, and Transformer pre-training on paired images and text.

The more cited of the two is UNITER (2019). It pre-trains one Transformer to produce a shared image-text embedding that can be reused across many vision-and-language tasks, rather than building a separate joint representation for each one. Image regions enter as detector features with their box positions, and words enter as WordPiece tokens. Both are mapped into a common space and trained on four public corpora with four objectives. One distinctive choice is to mask only one modality at a time, so the other stays intact as context.

The earlier paper, Towards End-to-End Reinforcement Learning of Dialogue Agents for Information Access (2017), deals with agents that help a user locate an entry in a knowledge base. Earlier systems sent an exact symbolic query to the database. That step throws away the parser's uncertainty and cannot be differentiated. The paper replaces it with a soft lookup that assigns a probability to every row, given the agent's beliefs about each slot, and passes that distribution to the policy network. This lets understanding and policy be learned jointly. The paper also handles missing values explicitly and tests robustness with a simulator whose replies can be made noisier.

Co-authors on these papers include Bhuwan Dhingra, Jianfeng Gao, Yen-Chun Chen, Jingjing Liu and Li Deng, along with others who each appear on one of the two works.

Faisal Ahmed has 2 papers in our library, in 2017–2019, most of them on Dialogue systems, Vision & language; often with Ahmed El Kholy, Bhuwan Dhingra, Jianfeng Gao.