Interns have impact. This couldn't be more true than on our Salesforce Research team, where our PhD interns are helping to shape the future of Artificial Intelligence.
Under the guidance of Salesforce Chief Scientist Richard Socher, three recent Futureforce interns worked on research papers that were recognized by the International Conference on Learning Representations (ICLR) and will be presented at the annual ICLR conference. These papers focused on the latest deep learning models, from image recognition to automated RNN Architecture Generation.
We caught up with one of our interns, Martin Schrimpf, to discuss his work on the team and reaction to having his paper selected for ICLR.
What made you choose Salesforce as a place to do your PhD research?
I met Richard Socher at MIT’s summer school in Woods Hole and became really interested in the strong work on natural language processing that he was presenting. At the time, I had mostly worked on vision and I found the team's work on visual question answering very intriguing as it was one of the first approaches towards integrating the domains of vision and language.
What sort of research did you work on once at Salesforce?
I was focused on machine learning; it often takes us years to find models that perform well on tasks such as machine translation. Our research aimed to cut down this expensive human search by automating the process and using a neural network to find other neural network architectures. Instead of years, it now took us days to find many models that out-performed the commonly used model on a machine translation and language modeling task. Through this approach, we were also able to remove the human priors to some extent and incorporate operations in our search space that have not been given much human attention, such as division and sine curves. And (much to our surprise) every single one of these uncommon operators found its way into an architecture that beat the standard architecture in their respective tasks, suggesting that the space of usable architectures is far larger than previously assumed.
In regards to how this research can impact our business, I am hoping that architecture search will further enable the use of machine learning in currently overlooked use cases.
How did you react when you heard the research you worked on at Salesforce had been accepted to ICLR?
My first reaction was to message Stephen Merity, who I worked with on the project, to ask if this meant that we really got accepted. We both had to re-read the email twice because we couldn't believe it, but after finally convincing ourselves, we were very happy that our work had paid off!
What would tell someone interested in interning with our Salesforce Research team?
I really enjoyed my time at Salesforce and was able to make a lot of progress on developing and applying my skills in machine learning research. The Salesforce Research team is very supportive and I felt like they were always more than happy to help with whatever obstacle I was facing. Aside from developing and building on my research and applying machine learning at scale across dozens of GPUs, I also made great friends that I’m really looking forward to connecting with again.
Learn more about our Salesforce Research Team and available PhD Internships at www.SalesforceResearch.com.
You can also follow our Salesforce Research Team adventures from ICLR on Facebook at https://www.facebook.com/SalesforceResearch/.