Food Distribution Engine
The main components of the Food Distribution Engine were a FoodAngels-facing interface for entering available rescued food items (Pick up tab on the interface), a participant-facing interface for entering food type and quantity available in kitchens (Kitchen tab), data labelling (Labelling tab) and training distribution models (Training tab), and choosing the distribution model to use for next day's distribution (Model tab).

You can explore the Food Distribution Engine's interface here.

The data labelling function emerged as the key feature of the Food Distribution Engine. Through enabling participants to generate their training data for the Engine’s classification algorithm, community-based labelling increased their autonomy over the models, a key aspect of inviting algorithms into commoning.

The SMUC team developed the Food Distribution Engine and corresponding smuc.kitchen interface in collaboration with George Simms and Mahinya Mkwawa.

Rescued Food Deliveries
In Basel, unsold food is abundantly available for collection. Rescued food is a resource ready to be turned into commons. FoodAngels made daily collections from their partnering supermarkets and sorted collected food in its depot based on appearance and smell. FoodAngels then prepared food boxes containing canned goods, fresh produce, bread, or chocolate. Using an electric cargo bike, they delivered the boxes to participants' doorsteps, guided by the responses from the Food Distribution Engine.

Machine Teaching in the Assembly
The Field Test Assemblies were sessions of story-sharing (cooking habits, nutritional preferences, recipes, views about food waste, health) and decision-making, similarly to Assemblies of housing cooperatives or community gardens. Additionally, the research team provided inputs about the use and background of the Food Distribution Engine.

The Assembly negotiated principles of food distribution, which it captured in the ‘Machine Teaching Protocol’. The Machine Teaching Protocol was a living document containing about half a dozen rules such as ‘If a kitchen is low on a type of food (for example, bread) and the people in the household eat bread, that kitchen gets the available food of that type.’ or ‘Substitutability between food types matters when considering which kitchen is low on stock to prioritise kitchens for receiving available food items’.

The Machine Teaching Protocol provided the basis for collectively producing training data through manual data labelling, a key component of collective machine teaching.

Training Data By Collective Labelling
Participants used the data labelling section of the Food Distribution Engine’s interface to generate training data. The interface showed the individual kitchens with a visualisation of the amount of ingredients they had in their fridges and pantries, like a snapshot of every kitchen’s inventory at a given moment. Additionally, the interface showed the amount and type of food (for example, ‘legumes 150 grams’) to be assigned to one of the kitchens. (Before enough real-time data had accumulated during the Field Test, a Data Generator fed the interface with synthetic data). Data labelling meant that participants assigned the rescued food items to kitchens based on the snapshot of the inventories and their understanding of how the Machine Teaching Protocol would apply. The inventory snapshots labelled with kitchens constituted the training data for the classification algorithm behind the Food Distribution Engine’s distribution model. During the Field Test, participants labelled slightly more than a thousand data points altogether.

During the week, participants exchanged experiences or recipes (such as mango pies, tomato paste bruschetta, and asparagus curries) or asked for assistance via a dedicated group in a messaging app. After the first Delivery Week, the group gathered again for a second Assembly meeting to iterate the Machine Teaching Protocol and re-teach the machine by generating updated training data.

Takeaways
Collective manual labelling to generate training data emerged as a key aspect of the Field Test. Participants translated the collectively negotiated protocols of food distribution into algorithmic classification through labelling which constituted machine teaching.

The participant-generated data and the collective machine teaching enhanced participants' autonomy over the algorithm, and mitigated external data biases. It also revealed a pathway towards countering extractivist algorithms, as participants captured the value generated by the Food Distribution Engine within the commons.

However, the Field Test also revealed the labour intensity of generating data about kitchens and the labelling activity itself, posing further questions about small data, implicit and affective aspects of negotiations in the Assembly getting lost in translation during labelling large amounts of data.

On a hopeful note, the Field Test also suggested that under conditions of serendipitous food deliveries participants prided themself in skillfully substituting ingredients in their favourite recipes. The joy of flexible cooking hints at pathways towards principles of good enough prediction in collective machine teaching efforts that counter the drive to ultimate optimisation.