Loading post…
AppleProblem & solution
A Practical Recipe for Semi-Supervised Federated ASR: Online Pseudo-Labels with Server Update Stabilization
Summary
Semi-supervised federated learning (SSFL) for ASR uses a teacher to generate pseudo-labels on unlabeled client data while the server trains on a small labeled seed. The paper finds that pseudo-label errors compound across the output sequence and training rounds, causing divergence and a large gap to fully-supervised FL. Two coupled design axes close this gap: the teacher type (per-client online, global broadcast, or transitioning) and the anchor (server-side labeled-data updates that stabilize training). A per-client online teacher diverges on its own but, once stabilized, matches or beats the global teacher in-domain and under domain shift. The server must keep training on labeled data between rounds; otherwise the online teacher drifts, and this interleaving governs convergence more than the seed model does. The proposed guidelines improve over the strongest prior method on 9 of 11 pairs, by 20.8% in-domain and 10.0% cross-domain on average.
Context
In semi-supervised federated learning (SSFL) for ASR, models train on client unlabeled data using a teacher-generated pseudo-labels, with a small labeled seed dataset on the server. Pseudo-label errors compound across the output sequence and training rounds, causing divergence and a large gap to fully-supervised FL. The gap can be closed by tuning the teacher and the server’s anchor updates jointly.
Approach / What changed
The paper studies semi-supervised federated learning for ASR, where client models use pseudo-labels from a teacher while the server trains on a small labeled seed. It identifies two coupled design axes: the teacher (per-client online, global broadcast, or a transitioning strategy) and the anchor (server-side labeled-data updates that stabilize training). Experiments vary the teacher type, anchor strength, seed size, data augmentation, and batch size to measure divergence and performance on in-domain and cross-domain tasks.
Takeaways
Closing the gap to fully-supervised FL in SSFL for ASR requires tuning the teacher type and the server anchor jointly, not independently.
A per-client online teacher diverges on its own but, when stabilized by strong server anchor updates, matches or beats the global teacher in-domain and under domain shift.
The server must interleave training on labeled data between rounds; otherwise, the online teacher drifts, and the anchor’s strength depends heavily on data augmentation and batch size.
FAQ
What are the teacher and anchor axes?
The teacher provides pseudo-labels for clients' unlabeled data. The anchor is the server-side training on labeled data between rounds, which stabilizes training and prevents the online teacher from drifting. The two are inseparable because aggressive teacher choices pay off only when the anchor is strong enough.
Which teacher strategy works best?
A per-client online teacher can match or beat the global teacher once stabilized, especially in-domain. As the labeled seed grows stronger, the online teacher's advantage narrows. The authors recommend a transitioning teacher that starts global and switches to online at a chosen round.
Source and freshness
- Source
- machinelearning.apple.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.