Blog
Faculty Demo Day XVI: My pick of this year’s data science projects
We recently hosted our 16th Faculty Fellowship Demo Day, where for the first time our fellows gave virtual project presentations to the companies that hosted them, their peers and our wider community.
I’ve attended six Demo Days now, and every single time I’m astounded by what the fellows can achieve in just six weeks of placement with their host companies. The quality of the projects and the variety of techniques on display is a testament to both the talent of the fellows and the effort the Faculty team puts into mentorship and instruction.
But these presentations don’t just give an insight into the fellows’ hard work and talent; they’re also a fascinating cross-section of the way organisations are using AI.
Three projects really stood out to me as examples of the cutting-edge techniques, big ambitions, and diverse ways of working that characterise AI today:
1. Automated menu transcription for an online food delivery platform
Menu transcription is costly for the large online food delivery platform where Nicholas completed his fellowship project. Most importantly, it’s also expensive for the restaurants that use the service. Manual transcription costs 250 hours of labour a month and can lead to delays for restaurants waiting to get on the platform.
Nicholas worked with the company to develop a language processing algorithm that can automatically extract the item name, price and description from photographs of menus. Furthermore, Nicholas added a quality filter to his algorithm, so that transcriptions the model makes with high confidence are sent directly to the restaurant, while lower-confidence transcriptions go to the human transcription team.
As a result, the company can substantially reduce the number of menus that need hand transcription, meaning that new restaurants get access to the platform – and access to a huge pool of potential customers – faster. I found that Nicholas’s use of a human-in-the-loop approach was a great demonstration of how machine learning can be safely deployed to tackle difficult problems and how it can be used creatively to get around potential barriers.
2. Model explainability techniques for Attest
Surveys are a vital tool for understanding the key demographic drivers behind customer behaviour. But analysing the (often unstructured) content of these surveys can be time-consuming and costly.
At Attest, Francesca developed an automated pipeline that uses explainable machine learning techniques to extract data from surveys. Typically, using a complex model like a neural network means sacrificing explainability – our ability as humans to understand why and how a model makes its decisions. However, Francesca used a Shapley value based approach, which enables use of highly complex ‘black box’ models.
Although the concept of Shapley values dates back to 1953, their application to the explainability of machine learning problems is relatively new. In this respect, Francesca’s project is a great example of how quickly and effectively the latest techniques can be applied to real world problems.
3. Optimising standby crew allocation for easyJet
Airlines need to have standby crews available to cover staff shortages and minimise disruption to flight schedules. However, forecasting the fluctuations in the demand for standby crews at multiple locations is extremely challenging.
Matías’ fellowship took him to easyJet’s data science team, where he worked to predict the airline’s demand for reserve crew. He combined machine learning and simulation to build a framework for smart allocation of standby crew.
Impressively, Matías’ approach has the potential to both reduce standby costs while also reducing the number of cancelled flights. The unusual combination of simulation and machine learning made this project stand out for me.
What will you build?
You can watch recordings of all of the projects presented by our fellows – including those mentioned in this blog – on our Youtube channel.
If you’d like to stay up to date with the latest groundbreaking work from our fellows, you can join our next Demo Day on 26 November 2020. We’ll hear from fellows working with companies including London Fire Brigade, JustEat, Advertising Standards Authority, Birdie Care and Imperial College London.
To secure your place for the next Demo Day, get in touch.
If you’re interested in taking part in the fellowship – either as a fellow, or as a host company, you can find out more here.
