A taste of Food 4.0 at EPFL’s Applied Machine Learning Days

A picture made with drone shows a woman working in kiwi fruit plantation near Guiyang, Guizhou province, China, 20 August 2020 (issued 28 August 2020). New technologies, including Artificial Intelligence, big data and 5G networks are transforming Chinese agriculture and are helping mitigate a mass rural exodus to the cities seen in recent decades. China has prioritised the application of new technologies to the rural sphere and the automation of harvesting procedures. (EPA/Alex Plavevski)

As the world scrambles to overcome the coronavirus pandemic, another serious challenge threatens to overshadow a post-pandemic recovery. Food insecurity is worsening globally, as conflict, climate change, Covid-19 and economic hardship disrupt production and limit access to food.

With over 10 billion mouths to feed by 2050 and rising living standards, research shows global food production will also need to expand by approximately 70 per cent in the next three decades.

That is where new technologies can play their part.  Over twenty scientists and scholars gathered online for the EPFL’s latest Applied Machine Learning Days (AMLD) to explore how artificial intelligence (AI) can improve food security, nutrition, and promote sustainable agriculture.

As AMLD states, “to achieve these goals, global food and agriculture systems will require profound changes, in which big data & AI technologies can play significant roles”.

Looming challenges. Surging worldwide demand for food goes hand in hand with environmental pressures. Agriculture is responsible for a quarter of global greenhouse gas emissions and the acceleration of biodiversity loss. Farming is also thirsty, accounting for 70 per cent of freshwater withdrawals while a quarter of crops grow in already water-stressed areas. Yet, a third of all food never makes it to our plates.

Christian Nils Schwab, executive director of the food and nutrition centre at EPFL, said in his opening speech:

“Such waste highlights the inefficiencies of the food systems in an environment where 3.5 billion people are undernourished, out of which 800 million are outright starving. Yet, at the other end of the scale, in developed countries, over 2.8 million people are over-nourished and found obese. More people are dying from excess nutrition than from starvation around the world.”

Such pressing sanitary and ecological issues call for rapid action - something that is not happening at present, says Schwab: “Despite the situation’s urgency, the agri-food sector has been extremely slow at embracing the possibilities and promises of machine learning and artificial intelligence.”

Embracing opportunities. On a more optimistic note, Schwab adds that “with challenges come opportunities”. In fact, the ensuing AI & Food and Nutrition talks showcase a wide range of state-of-the-art solutions. Among these, the prospect of an image-based food recognition app, allowing users to keep track of their nutritional intake by simply taking a snapshot of their meal. Over 2.5 terabytes of evolving data has already been collected via an open benchmark. As Sharada Mohanty, CEO & co-founder of AIcrowd, puts it: “if millions of users start posting images of their daily food, our models will only keep getting better”.

Read also: EPFL Robert West: ‘Machine learning is becoming a technology’

Artificial intelligence and machine learning render promising possibilities, such as tackling waste, enhancing food production, quality and safety, and seeking sustainability. For instance, precision agriculture – via targeted weeding applying machine learning – could reduce herbicide use by 95 per cent. For Yamine Bouzembrak, Researcher in Artificial Intelligence at Wageningen’s University Food Safety Research, AI can also help predict food fraud by scanning scientific literature and media, in order to identify unknown and potentially hazardous stimulants in food supplements.

Towards a healthy, personalized diet. Patrizia Catellani leads her talk with a simple question: “Can artificial intelligence and social psychology nudge towards a healthier diet?” Full Professor at the department of psychology at the Catholic University of the Sacred Heart, Milan, she emphasises the importance of personalised interactions:

“When communicating about nutrition, message framing can matter more than message content. For instance, I could say: ‘if you eat two or more servings of vegetables per day, you will improve your stomach’s health’, but I could also say that you will avoid damaging your stomach. What is quite interesting is that there is no “killer message”, nor any optimal option. The persuasiveness of each message depends on individual psychological dimensions. Via personalised communication, we can push consumers towards a deeper understanding of the risks surrounding unhealthy nutrition.”

In the same vein, Christoph Trattner, Associate Professor at the University of Bergen, Norway, and Centre Director of the Research Centre for Responsible Media Technology & Innovation, focusses his research on changing online eating habits:

“Algorithms on recipe websites generally produce unhealthy recommendations according to WHO and FSA guidelines. Moreover, user perception of healthiness is often biased. By understanding and modelling consumers’ eating habits on the web, we can then predict and alter them towards healthier food choices, for instance by manipulating images.”

Sola Shirai, a Doctoral Student at Rensselaer Polytechnic Institute, suggests utilising food knowledge graphs to find healthy ingredient substitutions:

“My focus has been on improving eating habits and nutritional intake aiming towards more context-aware and personalised health applications. Let’s imagine that a diabetic patient’s favourite dish is mashed potatoes. It is generally recommended that diabetic people avoid high carbohydrate intakes such as potatoes, since this can affect their blood sugar levels.

“Rather than preventing the patient from eating mashed potatoes, we could propose a similar low-carb alternative such as mashed cauliflower. By identifying which ingredients cause intolerance or allergies in an individual, artificial intelligence could automatically suggest substitutions. Hence we could preserve familiar recipes all by making them healthier.”

The discussions concluded with a pitch session showcasing six research projects, and will resume on 18 March, with a third track dedicated to Clinical Machine Learning. While there is still a long way to go before solving the world’s nutrition dilemma, AMLD’s second track provides food for thought.