差分
このページの2つのバージョン間の差分を表示します。
| 両方とも前のリビジョン 前のリビジョン 次のリビジョン | 前のリビジョン | ||
| playground:playground [2021/08/22 14:33] – Hideaki IIDUKA | playground:playground [2021/08/22 14:34] (現在) – Hideaki IIDUKA | ||
|---|---|---|---|
| 行 15: | 行 15: | ||
| | Proposed |$\displaystyle{\mathcal{O}\left( \frac{1}{T} \right)} + C_1 \alpha + C_2 \beta$|$\displaystyle{\mathcal{O}\left( \frac{1}{\sqrt{T}} \right)}$|$\displaystyle{\mathcal{O}\left( \frac{1}{n} \right)} + C_1 \alpha + C_2 \beta$|$\displaystyle{\mathcal{O}\left( \frac{1}{\sqrt{n}} \right)}$| | | Proposed |$\displaystyle{\mathcal{O}\left( \frac{1}{T} \right)} + C_1 \alpha + C_2 \beta$|$\displaystyle{\mathcal{O}\left( \frac{1}{\sqrt{T}} \right)}$|$\displaystyle{\mathcal{O}\left( \frac{1}{n} \right)} + C_1 \alpha + C_2 \beta$|$\displaystyle{\mathcal{O}\left( \frac{1}{\sqrt{n}} \right)}$| | ||
| - | Note: $C$, $C_1$, and $C_2$ are constants independent of learning rates $\alpha, \beta$, number of training examples $T$, and number of iterations $n$. The convergence rate for convex optimization is measured in terms of regret as $R(T)/T$, and the convergence rate for nonconvex optimization is measured as the expectation of the squared gradient norm $\min_{k\in [n]} \mathbb{E}[\|\nabla f(\bm{x})\|^2]$. In the case of using constant learning rates, SGD \cite{sca2020} and Algorithm \ref{algo: | + | Note: $C$, $C_1$, and $C_2$ are constants independent of learning rates $\alpha, \beta$, number of training examples $T$, and number of iterations $n$. The convergence rate for convex optimization is measured in terms of regret as $R(T)/T$, and the convergence rate for nonconvex optimization is measured as the expectation of the squared gradient norm $\min_{k\in [n]} \mathbb{E}[\|\nabla f(x)\|^2]$. In the case of using constant learning rates, SGD \cite{sca2020} and Proposed |
| \end{table*} | \end{table*} | ||