Digital Summit replay: how to bring your datasets to life

Jenni Baker, assistant editor at The Drum, chaired the session and was joined by Alessandra Sala, director of AI and data science, Shutterstock; Will Lowe, chief data officer, Engine UK; Sorcha Gilroy, data science team lead, Peak; and David Olesnevich, head of product, IBM Watson Advertising.

Sala opened the conversation on key trends in the marketplace: “The key challenges in the AI and data space today are around data quality and diversity. We know that computer vision models, for example, need vast amounts of quality data to help train, develop and test their effectiveness. At Shutterstock, companies work with us to expand their data capability to train those models. We take quality very seriously. Every one of our 1.7 million data assets goes through a rigorous process of human and AI review to guarantee its quality.

“On the diversity side, biased data can lead to inaccurate and, often, unethical AI models, so we constantly expand our data collection across many dimensions of diversity. We tap into a network of more than a million contributors around 150 countries. We’re constantly sourcing data, looking at diversity from a visual perspective, from a cultural perspective, from a linguistic perspective. We source fresh data every day as the world constantly evolves around us.”

IBM’s Olesnevich said: “There's so much data that's being created and coming into our systems all the time. Often, that data needs a lot of cleaning and transformation. To really gather value from an AI system, you’ll need to leverage automation to get that ‘time to value’ in a shorter window. We're focused on shortening the cycle, working with our partners to deliver value from these models and the feedback loop they help create.”

Gilroy added that, alongside data quality, marketers need to ensure they have the infrastructure they need to make sense of a diverse dataset. She said: “We work with a lot of retailers and a big trend we've been seeing recently is retailers wanting to pull lots of data sources together. So, for example, an ecommerce retailer wanting to combine their transactional data with web browsing data and app data, or social media interactions. A broad view of their customers is super useful for marketing teams, but achieving it comes with major challenges as well. Marrying up this data can be difficult if the different systems used different Customer ID for the same individual. Marketing teams really need to engage with their data team early in the process to get the infrastructure they need to go forward.”

AI in action

Lowe of interagency consultancy Engine UK said: “We help UK government departments and brands in the UK on their digital and data transformation journey. A lot of the time these organizations have had data for decades, but data quality has never risen to the top of their agenda before. In my experience, unless you get the board excited, it's very hard for them to know why they should invest in data quality. The problem is, to get a board excited, you have to show them the value that can be created by having better quality data! So, it's a bit ‘chicken and egg’ for a lot of the clients and organizations we work with at first.

“One really exciting development is how we're now getting so much better at using unstructured data sources. There are projects underway in the NHS, for example, looking at how to process millions of handwritten doctor’s notes and helping to create predictive models for early diagnosis of various conditions. Likewise, one of the big pet companies is doing something similar with the medical notes written by vets.”

Sala said: “At Shutterstock, we're focusing on the way that marketers can use AI to improve their creative process. Content decisions can be guided and supported through smart use of data. Marketers need to reach huge audiences and every single person is a constantly evolving individual with different needs at different times. Being able to predict how a piece of creative will actually be consumed and received by a diverse range of customer-types in practice is invaluable.

“We use ‘computer vision’ and natural language processing to break down all the elements of a creative, to understand what composition will resonate best with specific people in different contexts and settings. Computer vision is when machine learning is used to mimic the human vision system. We apply it to our 380 million images, 22 million videos and vast library of 3D objects, with the addition of a highly curated metadata. We also recently acquired three leading AI companies to incorporate with our solution and we’re growing our Data Science Division to dramatically accelerate our speed to market with Shutterstock AI. We predict big things in this space.”

You can watch the full panel session here.