REVIEW 3 cited by
Large Stepsize Gradient Descent for Logistic Loss: Non-Monotonicity of the Loss Improves Optimization Efficiency
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Signed reviews
abstract
We consider gradient descent (GD) with a constant stepsize applied to logistic regression with linearly separable data, where the constant stepsize $\eta$ is so large that the loss initially oscillates. We show that GD exits this initial oscillatory phase rapidly -- in $\mathcal{O}(\eta)$ steps -- and subsequently achieves an $\tilde{\mathcal{O}}(1 / (\eta t) )$ convergence rate after $t$ additional steps. Our results imply that, given a budget of $T$ steps, GD can achieve an accelerated loss of $\tilde{\mathcal{O}}(1/T^2)$ with an aggressive stepsize $\eta:= \Theta( T)$, without any use of momentum or variable stepsize schedulers. Our proof technique is versatile and also handles general classification loss functions (where exponential tails are needed for the $\tilde{\mathcal{O}}(1/T^2)$ acceleration), nonlinear predictors in the neural tangent kernel regime, and online stochastic gradient descent (SGD) with a large stepsize, under suitable separability conditions.
Forward citations
Cited by 3 Pith papers
-
Local Steps Speed Up Local GD for Heterogeneous Distributed Logistic Regression
A two-stage Local GD with learning-rate warmup achieves O(1/(K R)) convergence for heterogeneous distributed logistic regression, proving that local steps can provably reduce communication rounds.
-
Small steps no more: Global convergence of stochastic gradient bandits for arbitrary learning rates
The softmax gradient bandit converges almost surely to the optimal action for any constant learning rate, removing the small-learning-rate restriction of prior work.
-
Finite Horizon Optimization: Framework and Applications
A finite-horizon stepsize rule for the primal-dual method on LP, found via a 4x4 SDP, is claimed to accelerate convergence at the T-th iteration and to give about 3.9x speedup on Netlib instances.
Discussion (0). Continue with ORCID to comment.