Which framework should you choose for your federated learning projects?

Study Results: Flower Takes the Lead

A comprehensive study comparing 15 open-source FL frameworks across 15 evaluation criteria revealed Flower as the clear winner, scoring an impressive 84.75%! More here.

Why We’re Choosing Flower

• Top-Tier Partnerships: Collaborations with NVIDIA, Google, AWS, Intel, and Mozilla, to name a few.

• Cutting-Edge Methods: Quick adoption of newly published optimisation techniques.

• Beyond Basic FL: Supports additional methods like differential privacy.

• Reproducibility Focus: Runs a 3-month community sprint to improve reproducibility in FL with the goal of creating 50 high-quality baselines.

• Thriving Ecosystem: The growing community of developers and adopters thanks to its open-source and user-focused, and friendly approach.

At AI4Cosmetics, we are pioneering the use of federated learning for chemical safety assessment. Our goal is to help the industry with more accurate in silico predictions (not outside of the applicability domain) and regulatory-ready evidence. Our proof-of-concept use cases have already shown promising results, and we are now ready to bring them into real-world applications.

If this challenge resonates with you, we’d love to hear from you and join the consortium. We also offer bespoke workshops and can assist in setting up your federated models.

We gratefully acknowledge the Flower Pilot Program and the feasibility grant MKB Innovatiestimulering Topsectoren (MIT) Noord-Holland for their support in advancing our research.

Can distributed AI training actually help save our planet?

The Carbon Problem

For years, deep learning models have been growing increasingly complex and energy-hungry. Data centres now account for 0.3% of global carbon emissions — a number that keeps rising. But there’s hope on the horizon!

What Makes Federated Learning Different?

Federated Learning systems have three key components:

• Nodes: Each node represents an individual client device doing local training.

• Server: Coordinates the collaboration.

• Communication Network: Allows the transfer of data between the nodes and the server (not raw data!).

Communication plays a critical role in federated learning, and it can become a significant bottleneck in the process. The problem is not only the amount of communication bandwidth, but also the availability of a stable network.

The Green Potential

FL can significantly reduce carbon emissions compared to traditional centralised training. Here’s why:

1. Device Efficiency: Edge devices typically consume much less power than massive data centre GPUs.

2. No Cooling Costs: Data centres spend up to 40% of energy just on cooling. FL leverages existing devices that don’t need extra cooling.

3. Location Advantage: Training can happen in regions with cleaner energy grids.

The Environmental Trade-offs

FL isn’t automatically greener in all scenarios. Its carbon footprint depends on:

• Where You Train: Countries with cleaner energy show better emission profiles.

• What You Train: Lightweight models for tasks like speech recognition show the most promising results.

• How You Train: Strategies like FedADAM can be far more efficient than basic FedAVG.

• Data Transfer Costs: Communication between server and clients can account for anywhere from 0.4% to 93% of emissions!

The Bottom Line

For lightweight models deployed on energy-efficient devices in regions with clean electricity, FL can be dramatically greener than centralised alternatives.

References

• The Flower team showcased in one of their courses on DeepLearning.AI how a federated setup led to a 1063.5 times reduction in communication costs compared to a full centralised LLM model.

• A research summary done by Xinchi Qiu at CaMLSys, which we used for this edition.

At AI4Cosmetics, we are pioneering the use of federated learning for chemical safety assessment. Our goal is to help the industry with more accurate in silico predictions (not outside of the applicability domain) and regulatory-ready evidence. Our proof-of-concept use cases have already shown promising results, and we are now ready to bring them into real-world applications.

If this challenge resonates with you, we’d love to hear from you and join the consortium. We also offer bespoke workshops and can assist in setting up your federated models.

We gratefully acknowledge the Flower Pilot Program and the feasibility grant MKB Innovatiestimulering Topsectoren (MIT) Noord-Holland for their support in advancing our research.

Here are six examples of federated learning in real-world applications.

Example 1: MELLODDY – When Pharma Companies Federate

The MELLODDY project brought together 10 pharmaceutical companies to collaboratively train models, without ever sharing sensitive data. Over 2.6 billion experimental datapoints from more than 21 million small molecules were used to improve predictions in pharmacodynamics and pharmacokinetics.

Using a multitask learning approach and a privacy-audited platform, MELLODDY showed that federated models outperformed local ones, especially in safety and pharmacokinetic predictions. More data, broader chemical space, and better models, without leaking IP.

📌 Outcome: Federated learning at scale is feasible.

Example 2: Predicting Compound Mechanism of Action via Cell Imaging

In drug discovery, image-based profiling (like cell painting) generates rich, insightful data, but much of it is locked behind institutional firewalls.

In this study, multiple parties trained a federated model to predict the mechanism of action (MoA) from fluorescent cell images. The model performed better than those trained in isolation, even when the data varied across partners.

📌 Outcome: FL enables data-rich collaboration without giving up data privacy.

Example 3: Roche’s Journey with Effiris

Roche participated in the Effiris challenge, a federated learning hackathon by Lhasa Limited, to test FL for toxicity prediction.

By comparing their federated models with in-house models, Roche confirmed that FL offers solid predictive performance while preserving the confidentiality of chemical structures and safety data.

The current version of Effiris only allows for the building of classification models. The authors highlighted that regression models would be highly desirable for the optimisation of on-target and off-target activities.

For the FL approach to function effectively, the authors highlighted, every partner should regularly provide the consortium with new data points, and the amounts supplied should be reasonably proportional to the size of the company.

📌 Outcome: Federated learning currently has no control over variations in the experimental setup and assay conditions (such as temperature, pH, or sampling) under which the data sets are supplied by different partners, among other limitations.

Example 4: FedECA in Clinical Trials by Owkin

FedECA was developed and proposed for estimating treatment effects in the context of external control arms (ECA) needed for efficacy validation in drug development and regulatory approval.

They highlighted that implementing FL in real-world healthcare environments presents major non-algorithmic and operational challenges, such as suboptimal per-site missing data imputation and the complexity of setting up networks that are compatible across heterogeneous IT systems.

📌 Outcome: They made the code publicly available to further advance their federated research network.

Example 5: Federated AI for Protein Properties Prediction by NVIDIA

Two teams at NVIDIA, mainly the NVIDIA BioNeMo and NVIDIA FLARE, combined their efforts and federated their AI model for protein properties prediction.

📌 Outcome: Their model is also made publicly available, easy for scientists to try out.

Example 6: The Launch of the Lilly TuneLab

Lilly announced the launch of the TuneLab. Although I assume not with free access, now everyone can make use of their proprietary computational models.

The interesting part is that they federated those models, and hence, can participate in federated networks.

📌 Outcome: More and more pharmaceutical companies are starting to commercialise their internally developed models, and Lilly set the tone for the adoption of FL as an add-on.

Additional examples in the medical field

• Pioneering the Future of Medical Imaging with Federated Learning

• Eye2Gene: Advancing Genetic Eye Disease Diagnosis with Federated Learning

• Scalable and low-cost Federated Learning in the NHS

• BloodCounts! real-world application of using Flower

Other resources

• The workshop organised by Owkin in 2021.

• Smajić et al (2023), Privacy-preserving techniques for decentralized and secure machine learning in drug discovery.

• Hanser (2023), Federated learning for molecular discovery.

• The Flower YouTube channel with recordings from the monthly events and their annual summits.

At AI4Cosmetics, we are pioneering the use of federated learning for chemical safety assessment. Our goal is to help the industry with more accurate in silico predictions (not outside of the applicability domain) and regulatory-ready evidence. Our proof-of-concept use cases have already shown promising results, and we are now ready to bring them into real-world applications.

If this challenge resonates with you, we’d love to hear from you and join the consortium. We also offer bespoke workshops and can assist in setting up your federated models.

We gratefully acknowledge the Flower Pilot Program and the feasibility grant MKB Innovatiestimulering Topsectoren (MIT) Noord-Holland for their support in advancing our research.

The next time someone mentions a federated system, you’ll know exactly which questions to ask!

Here’s your essential guide to federated strategies.

System Architectures

Federated systems come in four distinct types, each with unique characteristics:

• Cross-Silo, Multi-Org: Few participants (like hospitals) with large datasets and strong computational resources collaborating across organisational boundaries.

• Cross-Silo, Single-Org: One organisation with data distributed across regions (multinational bank).

• Cross-Device, Multi-User: Many clients with limited resources (like mobile phones) are connecting to a single organisation.

• Cross-Device, Single-User: One organisation connecting multiple internal machines (factory equipment).

Data Partitioning Methods

• Horizontal FL: Participants have similar features but different data samples.

• Vertical FL: Participants have different features but shared identifiers.

• Federated Transfer Learning: Pre-trained general models fine-tuned locally.

Popular Algorithms

• FedAvg: The foundational algorithm that aggregates model updates from selected clients in each round.

• FedProx: Adds constraints to keep local models from diverging too far from the global model

Key Questions to Ask when analysing any federated system:

1. Which architecture type is being used?

2. How is the data partitioned?

3. Which algorithm coordinates the learning?

Federated learning isn’t one-size-fits-all; the strategy must match the specific privacy, communication, and computational constraints of the ecosystem, which is orchestrated.

At AI4Cosmetics, we are pioneering the use of federated learning for chemical safety assessment. Our goal is to help the industry with more accurate in silico predictions (not outside of the applicability domain) and regulatory-ready evidence. Our proof-of-concept use cases have already shown promising results, and we are now ready to bring them into real-world applications.

If this challenge resonates with you, we’d love to hear from you and join the consortium. We also offer bespoke workshops and can assist in setting up your federated models.

We gratefully acknowledge the Flower Pilot Program and the feasibility grant MKB Innovatiestimulering Topsectoren (MIT) Noord-Holland for their support in advancing our research.

That’s where Federated Analytics (FA) comes in.

Wait... Learning Without Learning?

Yep! Federated Analytics uses the same setup as FL: multiple clients, private local data, a central coordinator, but instead of optimising a model, we compute statistics [1]. Think: mean values, counts, histograms… without data ever leaving the device.

In fact, federated learning is technically a subset of federated analytics, just with a very fancy “statistic”: the model.

An example

In a case study from Flower [2], the team demonstrates federated analytics in action, without any model training. Instead, it shows how to analyse decentralised data using Pandas across 100 clients (i.e., hypothetical organisations), each holding a private portion of the well-known Iris dataset, which contains 150 samples of Iris flowers from three species (setosa, versicolor, virginica).

The twist?
Each client locally computes statistics, like the frequency of sepal lengths and sepal widths and only the aggregated results are shared with the server. No raw data ever leaves the clients, illustrating a simple but powerful privacy-preserving analytics workflow.

Why this matters

FA opens the door to decentralised dashboards, reports, and exploratory data analysis in regulated or privacy-sensitive domains. It’s not only about smarter AI, it’s about smarter decisions without compromising trust.

References

1. Federated Analytics: Collaborative Data Science without Data Collection by Google

2. The Federated Analytics example by Flower

At AI4Cosmetics, we are pioneering the use of federated learning for chemical safety assessment. Our goal is to help the industry with more accurate in silico predictions (not outside of the applicability domain) and regulatory-ready evidence. Our proof-of-concept use cases have already shown promising results, and we are now ready to bring them into real-world applications.

If this challenge resonates with you, we’d love to hear from you and join the consortium. We also offer bespoke workshops and can assist in setting up your federated models.

We gratefully acknowledge the Flower Pilot Program and the feasibility grant MKB Innovatiestimulering Topsectoren (MIT) Noord-Holland for their support in advancing our research.

Let’s break it down.

Federated Learning was introduced by Google researchers in 2017 to solve a growing problem: how to train powerful AI models without centralising private data [1]. Instead of sending your data to a server, FL brings the training to you — on your phone, in your lab, or across your organisation.

Think of it like this: your data stays at home, the model travels. It learns from local data, sends only the updates back to a central server, and combines everyone’s learnings into one smarter global model.

How it works (in 3 steps):

1. Your device (or lab) downloads the global model.

2. It trains locally using your own data (chemical assay results, lab notes, sensor readings — you name it).

3. Only the updates (not your data, but the parameters of the model!) are sent back to the global model to be averaged with others.

Different data modalities, different computational modelling methods, different types of devices — all can be trained in a federated way.

The result? A high-performing AI model that learns from many sources, without compromising data ownership or privacy.

Where it started & where it’s going

FL started with mobile keyboards (yes, the one predicting your next word), but it’s now making waves in finance, insurance, self-driving cars, life sciences, agriculture, and even robotics. Wherever sensitive data lives, FL can help unlock its potential safely.

Now, it’s powering Large Language Models too

And FL isn’t stopping at keyboards or lab data, it’s now taking on the giants of AI. For example, the Flower team released FlowerLLM-3B in 2024 [2], the world’s first large language model with 3 billion parameters trained entirely using federated learning. That means even models as complex as GPT-like LLMs can now be trained across decentralised GPUs and data sources, without ever pooling raw data into a central server. It’s a massive step toward democratising access to AI, enabling safer, more inclusive development for science, healthcare, and beyond.

References

1. Kairouz et al. (2019), Advances and Open Problems in Federated Learning

2. Introducing FlowerLLM (Flower AI Summit 2024)

Additional learning resources

• An online comic on federated learning by Google AI

• Martin Keen at IBM explains in a video how federated learning works [6:27min]

• Li et al. (2019), Federated Learning: Challenges, Methods, and Future Directions

At AI4Cosmetics, we are pioneering the use of federated learning for chemical safety assessment. Our goal is to help the industry with more accurate in silico predictions (not outside of the applicability domain) and regulatory-ready evidence. Our proof-of-concept use cases have already shown promising results, and we are now ready to bring them into real-world applications.

If this challenge resonates with you, we’d love to hear from you and join the consortium. We also offer bespoke workshops and can assist in setting up your federated models.

We gratefully acknowledge the Flower Pilot Program and the feasibility grant MKB Innovatiestimulering Topsectoren (MIT) Noord-Holland for their support in advancing our research.

Why? Because the AI world is waking up to a truth long known by privacy-first pioneers like Prof. Nic Lane (Cambridge, Co-Founder of Flower [1, 2]):

“The first AGI will be federated.”

Here’s why this matters now

Privacy by design

The age of AI regulation has officially begun. With the EU’s AI Act pushing for responsible, rights-preserving systems, FL is positioned as a core solution. A recent paper [3] argues FL’s privacy-first architecture aligns better with regulatory demands than traditional centralised approaches.

Context is king

When data stays local, users feel safer sharing richer, more relevant context, leading to smarter, more personalised AI without compromising trust.

Data access = power

Edd Wilder-James once said, “The biggest obstacle to using advanced data analysis isn’t skill-based or technology; it’s plain old access to the data [4].” FL flips this on its head by enabling AI without extracting data. That’s a revolution in motion.

Down with the data deal

As Hannah Fry points out, the “free tech for your data” model is capitalism’s most brilliant and terrifying trick [5]. FL offers a new deal - keep your data, still get great AI.

Let’s build fairer AI

Katherine Jarmul reminds us, “the AI revolution is to make it usable for people on their devices or with friends in ways that they want. AI is a different type of computing, and until we have an AI that everybody can use, we just have a centralised, corporate-controlled device that is maybe really good at SEO spam, but it’s not good at anything else [6].” Decentralised AI offers the path toward more democratic, equitable, and inclusive AI ecosystems.

Check out where FL stands in the hype cycle [7].

References

1. Flower: The Friendly Federated AI Framework

2. Lane Nicholas, The first AGI will be Federated, NeurIPS (2024)

3. Woisetschläger et al. (2024), Federated Learning Priorities Under the European Union Artificial Intelligence Act

4. Wilder-James Edd, Breaking Down Data Silos (2016)

5. Fry Hannah, Hello World (2019)

6. Jarmul Katherine, Keynote DIY Personalization: How, when why to offer self-made AI (2024)

7. Gartner Hype Cycle (2023)

At AI4Cosmetics, we are pioneering the use of federated learning for chemical safety assessment. Our goal is to help the industry with more accurate in silico predictions (not outside of the applicability domain) and regulatory-ready evidence. Our proof-of-concept use cases have already shown promising results, and we are now ready to bring them into real-world applications.

If this challenge resonates with you, we’d love to hear from you and join the consortium. We also offer bespoke workshops and can assist in setting up your federated models.

We gratefully acknowledge the Flower Pilot Program and the feasibility grant MKB Innovatiestimulering Topsectoren (MIT) Noord-Holland for their support in advancing our research.