Solution | Simultrain

[ \tilde\nabla_k = \nabla \ell(w^(e)_k; x_k) + \alpha \cdot (w^(c)_k - w^(e)_k) ]

SimulTrain sends activations (lower dimension than raw data but higher than gradients). However, it enables bidirectional overlap , reducing total bandwidth-time product by 65% compared to SyncSGD. | Dataset | Centralized | SyncSGD | FedAvg (5 local steps) | SimulTrain | |-------------|-------------|---------|------------------------|------------| | UCF-101 | 84.2% | 83.9% | 81.1% | 83.7% | | WISDM | 91.5% | 91.3% | 88.9% | 91.1% | simultrain solution

where ( T_\textsend ) and ( T_\textrecv ) depend on bandwidth, and ( T_\textforward, T_\textbackward ) on model size. For large models (e.g., ResNet-50), ( T_\textsend \gg T_\textforward ) on typical 4G/5G networks. [ \tilde\nabla_k = \nabla \ell(w^(e)_k; x_k) + \alpha

of SimulTrain is that the forward pass of one batch and the backward pass of a previous batch can overlap in time, if we carefully manage parameter versions and gradients. This is analogous to CPU pipelining but applied to distributed training across heterogeneous compute nodes. For large models (e

[ \mathbbE[|\nabla \ell(w^(c)_K)|^2] \leq \frac2L(f(w^(c)_0) - f^*)K\eta + O(\eta \sigma^2) + O(\tau^2 \eta^2) ]

SimulTrain matches centralized accuracy within 0.5%, while FedAvg drops by ~3% due to local overfitting. Removing gradient forecast causes divergence after 500 steps (accuracy falls to 45%). Removing weight reconciliation increases staleness indefinitely, leading to 12% higher loss. 7. Discussion Why does SimulTrain work? The key is the forecast+reconciliation loop. Forecast reduces bias, reconciliation prevents catastrophic staleness. The pipeline ensures that both edge and cloud are always busy, achieving near-optimal utilization.

where ( \sigma^2 ) is gradient noise variance. This matches the rate of synchronous SGD when ( \tau ) is bounded.