Federated Learning in Retail


In the field of Artificial Intelligence and more precisely of Machine Learning (the mechanism by which machines learn from data), Federated Learning is a collaborative learning mechanism by the nodes of a distributed architecture. Term that indicates peripheral systems, such as a Retail Point of Sale, with autonomous processing capacity.

One of the first and well-known study was published in 2017 https://ai.googleblog.com/2017/04/federated-learning-collaborative.html It is therefore a rather recent trend and different from the more consolidated Edge Computing, where computing and prediction models (based on algorithms or neural networks) reside in the peripheric devices to react faster to local stimuli and exchange less data with the Cloud, which, however, remains the only location for creating and updating models.

Federated Learning (FL) consists of creating models at the edge and sharing them without necessarily exchanging data, with advantages on privacy and network traffic. In medical research, for example, this method allows hospitals and clinics to improve their Artificial Intelligence (AI) models by comparing them with others without sharing patient data which, even if anonymised, is preferable not to leave the place where have been generated and/or stored.


Advantages in Retail

In Retail, the advantages of privacy and network traffic reduction are applicable to customer monitoring systems by video cameras, WiFi, Bluetooth ... Compared to centralized ones, AI systems inside each store are faster in reporting dangerous behavior for other customers or for the assets and can also identify the closest assistant, warning at first suspicions and not after the fact. About images interpretation, Edge Computing is already applied not only at the store level, but even inside video cameras for AI models execution or signal pre-processing.

Evolving to Federated Learning, new cases can be highlighted by local staff and just few minutes of video can be automatically used to update the model, without moving every day mountains of mostly useless data, clogging the Internet which is a finite and not free resource. A not less important advantage is that privacy becomes truly "by design".


Shopping experience personalization

AI allows handling each customer in the most pleasant and effective way by proposing, for example, only the promotions likely of interest or, in the stock management area, to suggest optimal prices and replenishment. A further significant advantage of FL in Retail is the creation of differential models, based on the local peculiarities of format, geographical area, competition and even climate, which use primarily the knowledge gathered locally, but also by other stores.

The data of each Store contain the greatest dose of "truth" but, unfortunately, also the least reliability due to the not negligible percentage of anomalous data (noise). Conversely, the models processed for the entire chain are less specific, but more reliable because, on large numbers, the anomalies tend to compensate each other.

New market trends usually appear as weak signals from few locations. FL allows these signals propagates to all stores. Each one will amplify or attenuate over time on the basis of their actual data.

FL amplifies the strengths and reduces the defects of the data collected from different locations spread throughout the territory. Compared to eCommerce giants that own Big Data, the physical stores have Small Data but much more specific, contextual and more effective for the physical presence of the products and the possibility of immediate satisfaction for the customer.



Federated Learning can be seen as a collaborative process of finding the optimal AI model for each individual store, or even for each individual device.

To give another example on Computer Vision, the fruit and vegetables department cameras could have slightly different models on board from those of the packaged products departments. Measurement of the crowding degree could be common, but the recognition of spoiled or missing products, as well as fraud, could be usefully differentiated. Managers and associates of the fruit and vegetables sector are the most suited to provide signals for the model improvement in their own sector and similarly for the other sectors.

As can be seen from these few examples, the march of AI in ferrying our society to the Fourth Revolution is unstoppable. Waiting too long to infuse this new form of cognitive energy into your Retail business can prove fatal.