Abstract
We propose a projected semi-stochastic gradient descent method with mini-batch for improving both the theoretical complexity and practical performance of the general stochastic gradient descent method (SGD). We are able to prove linear convergence under weak strong convexity assumption. This requires no strong convexity assumption for minimizing the sum of smooth convex functions subject to a compact polyhedral set, which remains popular across machine learning community. Our PS2GD preserves the low-cost per iteration and high optimization accuracy via stochastic gradient variance-reduced technique, and admits a simple parallel implementation with mini-batches. Moreover, PS2GD is also applicable to dual problem of SVM with hinge loss.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
- 2.
It is possible to finish each iteration with only b evaluations for component gradients, namely \(\{\nabla f_{i}(y_{k,t})\}_{i\in A_{kt}}\), at the cost of having to store {∇f i (x k )} i ∈ [n], which is exactly the way that SAG [14] works. This speeds up the algorithm; nevertheless, it is impractical for big n.
- 3.
We only need to prove the existence of β and do not need to evaluate its value in practice. Lemma 4 provides the existence of β.
- 4.
rcv1 and news20 are available at http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/.
- 5.
Available at http://users.cecs.anu.edu.au/~xzhang/data/.
- 6.
In practice, it is impossible to ensure that evaluating different component gradients takes the same time; however, Fig. 2 implies the potential and advantage of applying mini-batch scheme with parallelism.
- 7.
Note that this quantity is never computed during the algorithm. We can use it in the analysis nevertheless.
- 8.
For simplicity, we omit the E[⋅ | y k, t ] notation in further analysis.
- 9.
\(\bar{y}_{k,t+1}\) is constant, conditioned on y k, t .
References
Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J. Imag. Sci. 2(1), 183–202 (2009)
Calamai, P.H., Moré, J.J.: Projected gradient methods for linearly constrained problems. Math. Program. 39, 93–116 (1987)
Defazio, A., Bach, F., Lacoste-Julien, S.: SAGA: a fast incremental gradient method with support for non-strongly convex composite objectives. In: NIPS (2014)
Fercoq, O., Richtárik, P.: Accelerated, parallel and proximal coordinate descent. arXiv:1312.5799 (2013)
Fercoq, O., Qu, Z., Richtárik, P., Takáč, M.: Fast distributed coordinate descent for non-strongly convex losses. In: IEEE Workshop on Machine Learning for Signal Processing (2014)
Gong, P., Ye, J.: Linear convergence of variance-reduced projected stochastic gradient without strong convexity. arXiv:1406.1102 (2014)
Hoffman, A.J.: On approximate solutions of systems of linear inequalities. J. Res. Natl. Bur. Stand. 49(4), 263–265 (1952)
Jaggi, M., Smith, V., Takáč, M., Terhorst, J., Hofmann, T., Jordan, M.I.: Communication-efficient distributed dual coordinate ascent. In: NIPS, pp. 3068–3076 (2014)
Johnson, R., Zhang, T.: Accelerating stochastic gradient descent using predictive variance reduction. In: NIPS, pp. 315–323 (2013)
Karimi, H., Nutini, J., Schmidt, M.: Linear convergence of gradient and proximal-gradient methods under the polyak-łojasiewicz condition. In: ECML PKDD, pp. 795–811 (2016)
Kloft, M., Brefeld, U., Laskov, P., Müller, K.-R., Zien, A., Sonnenburg, S.: Efficient and accurate lp-norm multiple kernel learning. In: NIPS, pp. 997–1005 (2009)
Konečný, J., Liu, J., Richtárik, P., Takáč, M.: Mini-batch semi-stochastic gradient descent in the proximal setting. IEEE J. Sel. Top. Sign. Proces. 10, 242–255 (2016)
Konečný, J., Richtárik, P.: Semi-stochastic gradient descent methods. arXiv:1312.1666 (2013)
Le Roux, N., Schmidt, M., Bach, F.: A stochastic gradient method with an exponential convergence rate for finite training sets. In: NIPS, pp. 2672–2680 (2012)
Liu, J., Wright, S.J.: Asynchronous stochastic coordinate descent: parallelism and convergence properties. SIAM J. Optim. 25(1), 351–376 (2015)
Mareček, J., Richtárik, P., Takáč, M.: Distributed block coordinate descent for minimizing partially separable functions. In: Numerical Analysis and Optimization 2014, Springer Proceedings in Mathematics and Statistics, pp. 261–286 (2014)
Necoara, I., Clipici, D.: Parallel random coordinate descent method for composite minimization: convergence analysis and error bounds. SIAM J. Optim. 26(1), 197–226 (2016)
Necoara, I., Patrascu, A.: A random coordinate descent algorithm for optimization problems with composite objective function and linear coupled constraints. Comput. Optim. Appl. 57(2), 307–337 (2014)
Necoara, I., Nesterov, Y., Glineur, F.: Linear convergence of first order methods for non-strongly convex optimization. arXiv:1504.06298 (2015)
Nemirovski, A., Juditsky, A., Lan, G., Shapiro, A.: Robust stochastic approximation approach to stochastic programming. SIAM J. Optim. 19(4), 1574–1609 (2009)
Nesterov, Y.: Introductory Lectures on Convex Optimization: A Basic Course. Kluwer, Boston (2004)
Nesterov, Y.: Efficiency of coordinate descent methods on huge-scale optimization problems. SIAM J. Optim. 22, 341–362 (2012)
Nesterov, Y.: Gradient methods for minimizing composite functions. Math. Program. 140(1), 125–161 (2013)
Nguyen, L.M., Liu, J., Scheinberg, K., Takáč, M.: SARAH: a novel method for machine learning problems using stochastic recursive gradient. arXiv:1703.00102 (2017)
Richtárik, P., Takáč, M.: Iteration complexity of randomized block-coordinate descent methods for minimizing a composite function. Math. Program. 144(1–2), 1–38 (2014)
Richtárik, P., Takáč, M.: Distributed coordinate descent method for learning with big data. J. Mach. Learn. Res. 17, 1–25 (2016)
Richtárik, P., Takáč, M.: Parallel coordinate descent methods for big data optimization. Math. Program. Ser. A 156, 1–52 (2016)
Shalev-Shwartz, S., Zhang, T.: Accelerated mini-batch stochastic dual coordinate ascent. In: NIPS, pp. 378–385 (2013)
Shalev-Shwartz, S., Zhang, T.: Stochastic dual coordinate ascent methods for regularized loss. J. Mach. Learn. Res. 14(1), 567–599 (2013)
Shalev-Shwartz, S., Singer, Y., Srebro, N., Cotter, A.: Pegasos: primal estimated sub-gradient solver for SVM. Math. Program. Ser. A, B Spec. Issue Optim. Mach. Learn. 127, 3–30 (2011)
Shamir, O., Zhang, T.: Stochastic gradient descent for non-smooth optimization: convergence results and optimal averaging schemes. In: ICML, pp. 71–79. Springer, New York (2013)
Takáč, M., Bijral, A.S., Richtárik, P., Srebro, N.: Mini-batch primal and dual methods for SVMs. In: ICML, pp. 537–552. Springer (2013)
Wang, P.-W., Lin, C.-J.: Iteration complexity of feasible descent methods for convex optimization. J. Mach. Learn. Res. 15, 1523–1548 (2014)
Xiao, L., Zhang, T.: A proximal stochastic gradient method with progressive variance reduction. SIAM J. Optim. 24(4), 2057–2075 (2014)
Zhang, T.: Solving large scale linear prediction using stochastic gradient descent algorithms. In: ICML, pp. 919–926. Springer (2004)
Zhang, H.: The restricted strong convexity revisited: analysis of equivalence to error bound and quadratic growth. Optim. Lett. 11(4), 817–833 (2016)
Zhang, L., Mahdavi, M., Jin, R.: Linear convergence with condition number independent access of full gradients. In: NIPS, pp. 980–988 (2013)
Acknowledgements
This research of Jie Liu and Martin Takáč was supported by National Science Foundation grant CCF-1618717. We would like to thank Ji Liu for his helpful suggestions on related works.
Author information
Authors and Affiliations
Corresponding authors
Editor information
Editors and Affiliations
Appendices
Appendix 1: Technical Results
Lemma 1.
Let set \(\mathcal{W}\subseteq \mathbb{R}^{d}\) be nonempty, closed, and convex, then for any \(x,y \in \mathbb{R}^{d}\),
Note that the above contractiveness of projection operator is a standard result in optimization literature. We provide proof for completeness.
Inspired by Lemma 1 in [34], we derive the following lemma for projected algorithms.
Lemma 2 (Modified Lemma 1 in [34]).
Let Assumption 1 hold and let \(w_{{\ast}}\in \mathcal{W}^{{\ast}}\) be any optimal solution to Problem (1). Then for any feasible solution \(w \in \mathcal{W}\) , the following holds:
Lemmas 3 and 4 come from [12] and [33], respectively. Please refer to the corresponding references for complete proofs.
Lemma 3 (Lemma 4 in [12]).
Let {ξ i } i = 1 n be a collection of vectors in \(\mathbb{R}^{d}\) and \(\mu \stackrel{{\it \mathit{\text{def}}}}{=} \frac{1} {n}\sum _{i=1}^{n}\xi _{ i} \in \mathbb{R}^{d}\) . Let \(\hat{S}\) be a τ-nice sampling. Then
Following from the proof of Corollary 3 in [34], by applying Lemma 3 with ξ i : = ∇f i (y k, t−1) −∇f i (w k ) = a i [∇g i (a i T y k, t−1) −∇g i (a i T w k )] and Lemma 2, we have the bound for variance as follows.
Theorem 3 (Bounding Variance).
Considering the definition of G k, t in Algorithm 1 , conditioned on y k, t , we have \(\mathbf{E}[G_{k,t}] = \frac{1} {n}\sum _{i=1}^{n}\nabla g_{ i}(y_{k,t}) + q = \nabla F(y_{k,t})\) and the variance satisfies,
Lemma 4 (Hoffman Bound, Lemma 15 in [33]).
Consider a non-empty polyhedron
For any w, there is a feasible point w ∗ such that
where θ(A, C) is independent of x,
Lemma 5 (Weak Strong Convexity).
Let \(w \in \mathcal{W}:=\{ w \in \mathbb{R}^{d}: Cw \leq c\}\) be any feasible solution (Assumption 3 ) and \(w_{{\ast}} =\mathop{ \mathrm{proj}}\nolimits _{\mathcal{W}^{{\ast}}}(w)\) which is an optimal solution for Problem (1). Then under Assumptions 2 – 3 , there exists a constant β > 0 such that for all \(w \in \mathcal{W}\) , the following holds,
where μ is defined in Assumption 2 . β can be evaluated by β = θ 2 where θ is defined in (17).
Appendix 2: Proofs
2.1 Proof of Lemma 1
For any \(x,y \in \mathbb{R}^{d}\), by Projection Theorem, the following holds:
similarly, by symmetry, we have
or equivalently,
and by Cauchy-Schwarz inequality, we have
when \(\mathop{\mathrm{proj}}\nolimits _{\mathcal{W}}(x) =\mathop{ \mathrm{proj}}\nolimits _{\mathcal{W}}(y)\) are distinct; in addition, when \(\mathop{\mathrm{proj}}\nolimits _{\mathcal{W}}(x) =\mathop{ \mathrm{proj}}\nolimits _{\mathcal{W}}(y)\), the above inequality also holds. Hence, for any \(x,y \in \mathbb{R}^{d}\), which is the same to
2.2 Proof of Lemma 2
For any i ∈ {1, …, n}, consider the function
then it should be obvious that ∇ϕ i (w ∗) = ∇f i (w ∗) −∇f i (w ∗) = 0, hence \(\min _{w\in \mathbb{R}^{d}}\phi _{i}(w) =\phi _{i}(w_{{\ast}})\) because of the convexity of f i . By Assumption 1 and Remark 1, ∇ϕ i (w) is Lipschitz continuous with constant L, hence by Theorem 2.1.5 from [21] we have
which, by (20), suggests that
By averaging the above equation over i = 1, …, n and using the fact that \(F(w) = \frac{1} {n}\sum _{i=1}^{n}f_{ i}(w)\), we have
which, together with ∇F(w ∗)T(w − w ∗) ≥ 0 indicated by the optimality of w ∗ for Problem (1), completes the proof for Lemma 2.
2.3 Proof of Lemma 5
First, we will prove by contradiction that there exists a unique r such that \(\mathcal{W}^{{\ast}} =\{ w \in \mathbb{R}^{d}: Cw \leq c,Aw = r\}\) which is non-empty. Assume that there exist distinct \(w_{1},w_{2} \in \mathcal{W}^{{\ast}}\) such that Aw 1 ≠ Aw 2. Let us define the optimal value to be F ∗ which suggests that F ∗ = F(w 1) = F(w 2). Moreover, convexity of function F and feasible set \(\mathcal{W}\) suggests the convexity of \(\mathcal{W}^{{\ast}}\), then \(\frac{1} {2}(w_{1} + w_{2}) \in \mathcal{W}^{{\ast}}\). Therefore,
Strong convexity indicated in Assumption 2 suggests that
which is a contradiction, so there exists a unique r such that \(\mathcal{W}^{{\ast}}\) can be represented by \(\{w \in \mathbb{R}^{d}: Cw \leq c,Aw = r\}\).
For any \(w \in \mathcal{W} =\{ x \in \mathbb{R}^{d}: Cw \leq c\},[Cw - c]^{+} = 0\), then by Hoffman’s bound in Lemma 4, for any \(w \in \mathcal{W}\), there exists \(w' \in \mathcal{W}^{{\ast}}\) and a constant θ > 0 defined in (17), dependent on A and C, such that
Being aware of that by choosing \(w_{{\ast}} =\mathop{ \mathrm{proj}}\nolimits _{\mathcal{W}_{{\ast}}}(w)\), we have that ∥w − w ∗∥ ≤ ∥w − w′∥, which suggests that
or equivalently,
where β = θ 2 > 0.
Optimality of w ∗ for Problem (1) suggests that
then we can conclude the following:
which, by considering F(w) = g(Aw) + q T w in Problem (1), is equivalent to
2.4 Proof of Theorem 1
The proof is following the steps in [12, 34]. For convenience, let us define the stochastic gradient mapping
then the iterate update can be written as
Let us estimate the change of ∥y k, t+1 − w ∗∥. It holds that
By the optimality condition of \(y_{k,t+1} =\mathop{ \mathrm{proj}}\nolimits _{\mathcal{W}}(y_{k,t} - hG_{k,t}) =\arg \min _{w\in \mathcal{W}}\{\tfrac{1} {2}\|w - (y_{k,t} - hG_{k,t})\|^{2}\}\), we have
then the update y k, t+1 = y k, t − hd k, t suggests that
Moreover, Lipschitz continuity of the gradient of F implies that
Let us define the operator Δ k, t = G k, t −∇F(y k, t ), so
Convexity of F suggests that
then equivalently,
Therefore,
In order to bound −Δ k, t T(y k, t+1 − w ∗), let us define the proximal full gradient update asFootnote 7
with which, by using Cauchy-Schwartz inequality and Lemma 1, we can conclude that
So we have
By taking expectation, conditioned on y k, t Footnote 8 we obtain
where we have used that E[Δ k, t ] = E[G k, t ] −∇F(y k, t ) = 0 and hence \(\mathbf{E}[-\varDelta _{k,t}^{T}(\bar{y}_{k,t+1} - w_{{\ast}})] = 0\).Footnote 9 Now, if we put (16) into (34) we obtain
where \(\alpha (b) = \frac{m-b} {b(m-1)}\).
Now, if we consider that we have just lower-bounds ν F ≥ 0 of the true strong convexity parameter μ F , then we obtain from (35) that
which, by decreasing the index t by 1, is equivalent to
Now, by the definition of w k we have that
By summing (36) multiplied by (1 − hν F )M−t for t = 1, …, M, we can obtain the left-hand side
and the right-hand side
Combining (38) and (39) and using the fact that LHS ≤ RHS we have
Now, using (37) we obtain
Note that all the above results hold for any optimal solution \(w_{{\ast}}\in \mathcal{W}^{{\ast}}\); therefore, they also hold for \(w_{{\ast}}' =\mathop{ \mathrm{proj}}\nolimits _{\mathcal{W}^{{\ast}}}(w_{k})\), and Lemma 5 implies that, under weak strong convexity of F, i.e., ν F = 0,
Considering E∥y k, M − w ∗′∥2 ≥ 0, y k, 0 = w k , and using (41), the inequality (40) with w ∗ replaced by w ∗′ gives us
or equivalently,
when 1 − 4hLα(b) > 0 (which is equivalent to \(h \leq \frac{1} {4L\alpha (b)}\) ), and when ρ is defined as
The above statement, together with assumptions of h ≤ 1∕L, implies
Applying the above linear convergence relation recursively with chained expectations and realizing that F(w ∗′) = F(w ∗) for any \(w_{{\ast}}\in \mathcal{W}^{{\ast}}\) since \(w_{{\ast}},w_{{\ast}}'\in \mathcal{W}^{{\ast}}\), we have
Rights and permissions
Copyright information
© 2017 Springer International Publishing AG
About this paper
Cite this paper
Liu, J., Takáč, M. (2017). Projected Semi-Stochastic Gradient Descent Method with Mini-Batch Scheme Under Weak Strong Convexity Assumption. In: Takáč, M., Terlaky, T. (eds) Modeling and Optimization: Theory and Applications. MOPTA 2016. Springer Proceedings in Mathematics & Statistics, vol 213. Springer, Cham. https://doi.org/10.1007/978-3-319-66616-7_7
Download citation
DOI: https://doi.org/10.1007/978-3-319-66616-7_7
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-66615-0
Online ISBN: 978-3-319-66616-7
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)