AssayPilot research / Federated learning
A downloadable, inspectable research pipeline for laboratories that want to test collaborative learning on their own annotated plate images.
Describe annotated images, declare permitted use, identify physical dishes and acquisition groups, and record image hashes. Keep related observations in the same split.
Choose held-out laboratories, rounds, thresholds and hyperparameters before evaluating the test set.
Each client extracts documented image features and fits countability and count heads. Transfer the update files using a channel your laboratories approve. The coordinator validates and aggregates them.
Compare with local-only training, retain per-lab outcomes, and create an auditable candidate bundle. Candidates do not automatically replace the website detector.
What is implemented
Hand-engineered OpenCV image features, locally fitted countability and count models, sample-weighted FedAvg and a FedProx option. This is not a pretrained image encoder.
The kit includes manifests, commands, checkpoint validation and an executable generated-image protocol test. Its README gives the complete local workflow.
python -m training.federated --helpBefore using real lab data
The kit records a lab's authorization declaration; it cannot independently verify ownership or permission. Use consistent expert labels and retain an untouched held-out laboratory.
Updates and hash commitments may reveal information. FedAvg alone is not secure aggregation or differential privacy. Agree on recipients, transfer protection, retention and withdrawal handling before exchanging any real updates.
Images remain in each client's local folder in this workflow. The toolkit has no automatic network upload. Copying images into a shared folder changes that boundary.
The engineering loop can be run and inspected. Whether it improves colony counting on unfamiliar real laboratories is an experiment that still needs to be performed. Synthetic protocol checks establish neither accuracy nor privacy.