Federated Learning: Personalization vs. the Global Model

Federated Training

Personalization

Blend α (0 = global, 1 = local) 0%

Local fine-tune steps 6

Client data skew (non-IID) 35°

Focus Client

Live Metrics

Round

Global acc (avg)

—

Blended acc (avg)

—

Focus client acc

—

How it works

Six simulated devices each hold a private linear-classification task with its own slightly rotated decision boundary (non-IID data). Every round they run local logistic-regression gradient steps on their own points, and the server averages the resulting weights — FedAvg:

local:  w_k ← w_k − η·∇L_k(w_k)      (K steps, client k's data only)
global: w ← (1/N) Σ w_k                (server-side average)

A purely global model generalises across all clients but fits none of them perfectly; a purely local model overfits one client's skew. The α slider blends the two per client, the interpolated-personalization scheme studied in federated learning:

w_k(α) = (1 − α)·w_global + α·w_local_k
0 ≤ α ≤ 1  (0 = pure FedAvg, 1 = pure local fine-tune)

Run Rounds / Step Round — advances FedAvg: each client trains locally from the current global weights, then the server averages.
α slider — recolors every client's test points using the blended weights; watch a high-skew client's accuracy rise as α → 1 while the true global-only accuracy stays fixed.
Fine-tune steps — how many extra local gradient steps produce w_local_k from the current global weights.
Data skew — how far each client's true boundary is rotated from the others; at 0° every client is IID and personalization stops mattering.

This mirrors a real deployment trade-off: keyboard next-word prediction (Gboard-style) or hospital diagnostic models trained via FedAvg across sites with different populations, then fine-tuned locally so each site keeps some of its own signal.

Correctly classified Misclassified Global model

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.