Next Federation
Next Federation trains one digit classifier across 10 clients whose data are distributed differently (Dirichlet label skew) and who sometimes drop out. It compares FedAvg, FedProx and local-only training on global accuracy, the worst client's accuracy on its own data, and communication traffic. This is a simulation: all clients run in one process; nothing was deployed to real devices.
Space: https://huggingface.co/spaces/thelamapi/Next-Federation — change skew, dropout, clients and rounds; all three methods retrain in the browser.
Results (5 seeds per setting, 40 rounds, 2 local epochs, lr 0.3, FedProx μ 0.01)
| Setting | Method | Global test accuracy | Worst client | Mean client | Traffic (MB) |
|---|---|---|---|---|---|
| alpha=0.1, dropout=0.0 | FedAvg | 0.934 ± 0.011 | 0.784 ± 0.100 | 0.940 | 7.7 |
| alpha=0.1, dropout=0.0 | FedProx | 0.934 ± 0.011 | 0.784 ± 0.100 | 0.940 | 7.7 |
| alpha=0.1, dropout=0.0 | Local only | 0.357 ± 0.011 | 0.448 ± 0.391 | 0.920 | 0.0 |
| alpha=0.1, dropout=0.3 | FedAvg | 0.928 ± 0.019 | 0.776 ± 0.090 | 0.934 | 6.5 |
| alpha=0.1, dropout=0.3 | FedProx | 0.929 ± 0.018 | 0.769 ± 0.090 | 0.932 | 6.5 |
| alpha=0.1, dropout=0.3 | Local only | 0.357 ± 0.011 | 0.448 ± 0.391 | 0.920 | 0.0 |
| alpha=0.5, dropout=0.0 | FedAvg | 0.953 ± 0.004 | 0.866 ± 0.031 | 0.947 | 7.7 |
| alpha=0.5, dropout=0.0 | FedProx | 0.952 ± 0.003 | 0.866 ± 0.031 | 0.948 | 7.7 |
| alpha=0.5, dropout=0.0 | Local only | 0.657 ± 0.019 | 0.674 ± 0.092 | 0.881 | 0.0 |
| alpha=0.5, dropout=0.3 | FedAvg | 0.952 ± 0.005 | 0.860 ± 0.025 | 0.949 | 6.5 |
| alpha=0.5, dropout=0.3 | FedProx | 0.952 ± 0.006 | 0.860 ± 0.025 | 0.949 | 6.5 |
| alpha=0.5, dropout=0.3 | Local only | 0.657 ± 0.019 | 0.674 ± 0.092 | 0.881 | 0.0 |
| alpha=5.0, dropout=0.0 | FedAvg | 0.961 ± 0.004 | 0.880 ± 0.013 | 0.956 | 7.7 |
| alpha=5.0, dropout=0.0 | FedProx | 0.960 ± 0.005 | 0.880 ± 0.013 | 0.955 | 7.7 |
| alpha=5.0, dropout=0.0 | Local only | 0.871 ± 0.006 | 0.755 ± 0.074 | 0.883 | 0.0 |
| alpha=5.0, dropout=0.3 | FedAvg | 0.960 ± 0.005 | 0.890 ± 0.013 | 0.955 | 6.5 |
| alpha=5.0, dropout=0.3 | FedProx | 0.959 ± 0.005 | 0.890 ± 0.013 | 0.955 | 6.5 |
| alpha=5.0, dropout=0.3 | Local only | 0.871 ± 0.006 | 0.755 ± 0.074 | 0.883 | 0.0 |
Global test set: 360 digits never used for training or tuning. "Worst client" is the lowest accuracy over clients on their own held-out 20%. Local-only global accuracy is the mean of the 10 client models. Traffic counts float32 model downloads to every client every round plus uploads that arrive.
Method
- Model: MLP 64→32→10 with ReLU, softmax cross-entropy, minibatch SGD (batch 20).
- FedAvg: sample-weighted average of the updates that arrive. FedProx adds (μ/2)‖w − w_global‖² to every local step.
- Dropout: a client's update is lost with probability p (its download still counts).
- lr and μ were chosen on a separate validation set with development seeds 100–101 (lr scores {0.05: 0.9027777777777778, 0.1: 0.9361111111111111, 0.3: 0.9527777777777777}, μ scores {0.01: 0.9527777777777777, 0.1: 0.95, 1.0: 0.9333333333333333}).
- The Space runs a line-by-line JavaScript port; Node parity on 4 fixed runs: max abs difference 0.
- Released weights (
model.json): FedProx global model from α=0.5, dropout 0.3, seed 0.
Usage
from inference import load, predict
spec = load()
probs = predict(spec, pixels) # pixels: (N, 64) ints 0..16
Limitations
- Simulated clients on a small, clean dataset; real federations have system heterogeneity, stragglers and privacy mechanisms (none here).
- Dropout is independent per client and round; real outages are correlated.
- Recorded on NVIDIA L4. The numbers describe this simulation only.

