Bloomsbury AI
Bloomsbury AI received early access to cutting edge hardware and expertise from Digital Catapult, resulting in increased productivity and faster product development. The Bloomsbury AI team has recently been recruited by Facebook as part of its efforts to grow natural language processing research, and further understand its applications.
The challenge
Founded in 2015, as a spin off from a machine reading group at University College London, Bloomsbury AI was a startup with a strong research pedigree. It kept its commitment to ongoing research which fed into the products it was developing using natural language processing (NLP), which is a strand of AI that seeks to understand how people write or speak in a natural way.
Bloomsbury AI had created an algorithm and tool to read and understand documents. The vision was to enable large companies, such as law firms, to ask questions and receive accurate answers from an AI solution based on the information contained in large quantities of written documents.
The system would interrogate huge quantities of a documents and understand the context sufficiently to answer the question accurately. It would extract meaningful information from text it was fed.
Bloomsbury AI had many potential directions to take the the future development of its product, but as with many startups working with AI and machine learning, access to compute power posed a significant barrier to market entry.The algorithms the company was developing required a large amount of computing power, which lead to prohibitively high running costs for a startup.
Although it already had some test users, Bloomsbury AI didn’t have the capital to scale its idea. It wanted to train a bigger model and to do this it needed more compute power.
It also needed help to select, configure and integrate the multiple pieces of software required to train its algorithm using highly specialised, leading edge hardware that has been developed specifically for deep machine learning.
The solution
Through its AI programme, Machine Intelligence Garage, Digital Catapult provided Bloomsbury AI with access to NVIDIA DGX-1 and Cray CS-Storm systems that would provide the required compute power to train the algorithm on a much larger data set.
Machine Intelligence Garage is designed to help startups with a well defined business idea and technical capability for whom access to computation power is a barrier to growth. Successful applicants are given access to best-in-class cloud compute, hosted hardware or UK HPC from our partners and collaborators.
The programme also hosts workshops and experimentation-days to help companies of all sizes get to grips with systems for machine intelligence. The programme is independent and provider-agnostic. Via the programme Bloomsbury AI was also able to access specialist DevOps support. This helped to eliminate the technical challenges that come with using advanced hardware such as NVIDIA DGX-1 and a Cray CS-Storm system.
The results
Having access to specialist expertise and compute power was a game changer for Bloomsbury AI as it enabled them to fine tune the solution, test new ideas and accelerate product development. The resulting solution, called Cape, can now process data much more accurately, on a much greater variety of documents.
A common challenge in NLP is finding a way to accurately extract information not only from small paragraphs, but also, and especially from long documents too. Digital Catapult’s support has allowed Bloomsbury AI to experiment with models for understanding documents of different lengths. The added compute power enabled it to combine these models and train a single model that performs well on both, small and very large data sets simultaneously.
Before this, a model would analyse a data set to review very small documents to, for example, answer questions from a single paragraph with very high accuracy. Then another model was used to answer more general questions on longer docs (circa 16-30 paragraphs in length).
A single model to achieve both tasks gives the best of both worlds. With the help of Machine Intelligence Garage, Bloomsbury AI was able to increase the accuracy of its solution by a significant margin when trained on publicly available reference datasets.
Through it’s Machine Intelligence Garage programme Digital Catapult has enabled Bloomsbury AI to compete with big established players in the tech industry.
Bloomsbury AI’s expertise and success brought it to the attention of tech giant Facebook, which announced recently that the team would join the London lab in August. The Bloomsbury AI team will be working on strengthening Facebook’s efforts in natural language processing research, and help it further understand natural language and its applications.
Access to an NVIDIA DGX-1 and a Cray CS-Storm cluster, highly specialised, leading edge hardware
Sped up product development
Delivered greater accuracy
Successfully sold product