We have already discussed that AI-based applications are the future and offer great advantages in many industries and come with many business scenarios. But while we accept this as fact, how does AI perform in real-world applications, and to what extent can we trust its results? The inconvenient truth is that the results of such applications are only as good as the quality of the data available. In most cases and industries, there is limited quantity and quality of data that pose barriers to the wider adoption of these technologies. Other obstacles? Of course, we have a lot of them! Industry competition, privacy and security issues, complex data integration, and sharing processes are some of the problems faced by many users of artificial intelligence technologies. But there’s no reason to be doubtful about AI potential since new approaches have already emerged!
Another buzzword in the field of technology? We prefer to stay to the essence so let us give you a simple definition! Federated learning is a Machine technique that enables the use of decentralized data, e.g., residing on devices. As an example, we can refer to healthcare organizations where there is an increasing amount of data providing both advantages and challenges. While Machine Learning provides the tools needed to analyze big data Federated learning attempts to solve the data dilemma faced by traditional ML methods by enabling the possibility to train a shared global model with a central server while keeping all the sensitive data in local institutions like hospitals.
Is everything perfect in federated Learning? It’s not our intention to disappoint you but there are some drawbacks! Actual, there are 3 main categories:
Statistical: There is a doubt if locally available data points can be considered as a representative sample of the overall distribution.
Communication: The number of clients may be large and can be much larger than the average number of training samples stored in the activated clients.
Privacy and security: How can we be sure that none of the clients are malicious?
Existing frameworks for Federated Learning are Tensorflow, PySyft, Substra, and FATE. Of course, each one of them comes with certain pros and cons. Overall, existing federated learning approaches for healthcare applications consider federated learning in controlled settings in which data federation partners (clinics in our case) must be completely known in advance.
Going above and beyond
ASCAPE includes research teams with extensive expertise in the area of Federated Learning. As a result in the last two and a half years, we have managed to achieve significant improvements in the area. We all love making lists, that's what we do for our progress! Let's take a look:
- Democratization of federated models: we developed a federated learning scheme in which models are learned incrementally or semi-concurrently as clinics join the ASCAPE platform.
- Continuous models’ performance evaluation: Global models are downloaded at the edge node and evaluated against the performance of personalized local models having also the ability of feature selection
- Semi-supervised cancer-care predictive models: we support semi-supervised and unsupervised learning of federated machine and deep learning models combined with feature selection techniques
- Unsupervised cancer-care data analytics and outlier detection: ASCAPE detects data instances strongly deviating either from the rest of local data or unknown data instances that were previously used to train the global model and enables measuring similarity between data instances coming from different data federation partners without any data exchange.
- An extendable base of models for cancer-care predictions: the cancer-care knowledge contained in the can continuously grow in two directions: (a) by updating existing predictive models (knowledge growth in terms of accumulated medical evidence for a particular QoL indicator or intervention), and (b) by creating new models (knowledge growth in terms of broader coverage of QoL indicators and interventions).