Innovation & Enterprise


UCL spinout helps businesses solve problems through human-artificial intelligence (AI) cooperation

Humanloop makes it easier and quicker for people in businesses to develop AI programmes and then work cooperatively with AI to solve problems.

Two people sitting on opposite sides of the table, working

12 February 2021

The UCL spinout helps customers in a number of sectors with various problems, for example moderating content posted in online medical forums. 

The company’s vision is to enable workers to teach computers solve a task by explaining it to them, in natural language, as they would with a human colleague.

An innovative AI ecosystem

Humanloop is based on research from UCL’s AI Centre. It was spun-out with assistance from UCL Business (UCLB), the commercialisation arm of UCL and a part of UCL Innovation & Enterprise. The company benefitted from UCLB’s Portico Ventures scheme which offers a rapid process for licensing of non-patented UCL intellectual property into founder-driven spinouts.

The team comprises UCL Professors David Barber and Emine Yilmaz with PhD students Raza Habib and Peter Hayes. They are joined by entrepreneur and scientist Jordan Burgess, who was previously involved with another UCL spinout (Bloomsbury AI) as well as the Amazon Alexa AI team.

The company secured investment from UCL’s Technology Fund (UCLTF), which is managed by AlbionVC in collaboration with UCLB.

Raza said: "UCLB was very supportive of our goals of spinning out from the university and advised us on all the requirements relating to intellectual property.” Co-founder David added: "UCL Technology Fund’s vital investment means we can put our research into practice and take this exciting vision forward."

Humanloop was also accepted onto the 2020 cohort of Y Combinator, one of the world’s most prestigious business accelerators. This helped them define their offering and network with potential investors. 

Training the machine 

The power of AI, and specifically machine learning, lies in its ability to recognise patterns over time and solve problems automatically. Machine learning powers many of the services we use today. That includes everything from recommendations on Netflix, to banks’ credit decisions and increasingly more complex tasks like diagnosis of medical conditions from X-rays. 

But acquiring any new skill or ability requires a period of training and that’s the same for machine learning. This involves a human operator labelling or annotating large training datasets in a way that the machine learning programme can understand. This is a major, time consuming bottleneck for most companies.

“The way that machine learning is typically deployed is an extremely slow process and it takes companies a very long time to go from an idea to getting machine learning usefully deployed,” Raza says.

Towards a more collaborative and iterative AI

Humanloop’s solution, which involves natural language processing, makes it far easier and quicker to annotate data, then train and deploy machine learning. Crucially, it allows a more collaborative way of working between AI and human operator. Left entirely to its own devices AI tends to degrade in performance over time and makes mistakes.  

The team also emphasises that the machine learning their method builds never makes any decisions in a safety critical or important situation where it’s not confident. 

Future directions

One area where Humanloop is actively working with customers is in moderation of online communities that publish user generated content. It’s essential that toxic content or misleading information is quickly flagged, particularly in medical forums. That’s a huge undertaking, and something where machine learning can be of tremendous value, when working alongside a human. That’s where Humanloop’s iterative, cooperative solution is proving particularly beneficial. 

The company is also working on customer support ticket routing, triaging large volumes of customer questions and suggesting a response or forwarding to a human operator. Other areas of interest include extraction of information from lengthy legal and real estate contracts.

“We see this as a first step on a much longer journey,” says Raza. “The ultimate goal is for a knowledge worker, such as an online moderator or lawyer, to teach a computer to solve some task by explaining to them exactly like you would teach a colleague. That’s the vision of the company.”


Find out more about:

Photo © Humanloop

(facebook button)