Understanding How Machines Understand Language
If Maziar saw the future of autonomous agents, Mojtaba learned how to build systems at scale, and Mehrdad learned how to connect technology with human needs, Souroush Mehri spent years working on another critical piece of the puzzle: How do machines understand language and information?
From 2014 to 2017, Souroush pursued his Master’s research at Mila, the University of Montreal’s renowned AI research institute, under Yoshua Bengio and Aaron Courville supervision.
This was not simply an education in AI. It was participation in the development of foundational AI technology.
Souroush became the first author on SampleRNN, an early neural architecture capable of generating raw audio end-to-end. He also co-authored Char2Wav, part of the early generation of neural text-to-speech systems. These were years when many of today’s AI capabilities were still research problems.
In 2016, Ubisoft Montreal’s La Forge research lab funded his Master’s thesis and brought him into the organization to rebuild the research as an industrial system. It was one of his first experiences crossing the boundary between academic research and technology that had to work in the real world.
In 2017, he joined Maluuba, an AI startup working on natural language understanding, question answering, machine reading, and reinforcement learning.
Soon afterward, Microsoft acquired Maluuba. Souroush became part of Microsoft Research Montreal, where his role was deliberately positioned between research and engineering. That combination became central to his career.
He worked with product teams, including Bing, on language models and retrieval. He researched why language models can become overly dependent on spurious correlations. In parallel, he maintained research collaborations in reinforcement learning, transfer learning, and imitation learning.
And his work did not remain confined to papers. With Xbox Cloud Gaming, he was part of a small team that trained and deployed models designed to mask network latency during gameplay. The work resulted in a U.S. patent.
The pattern throughout his career was consistent: Understand how intelligent systems work. Understand where they fail. Then build systems that work in the real world.
After years working with retrieval, language models, AI agents, and their failure modes, Souroush brought something essential to the eventual Elementera team: An understanding not merely of how to optimize content for algorithms, but of how intelligent systems interpret information in the first place.