Abstract
Learning rates in stochastic neural network training are currently determined a priori to training, using expensive manual or automated iterative tuning. Attempts to resolve learning rates adaptively, using line searches, have proven computationally demanding. Reducing the computational cost by considering mini-batch sub-sampling (MBSS) introduces challenges due to significant variance in information between batches that may present as discontinuities in the loss function, depending on the MBSS approach. This study proposes a robust approach to adaptively resolve learning rates in dynamic MBSS loss functions. This is achieved by finding sign changes from negative to positive along directional derivatives, which ultimately converge to a stochastic non-negative associated gradient projection point. Through a number of investigative studies, we demonstrate that gradient-only line searches (GOLS) resolve learning rates adaptively, improving convergence performance over minimization line searches, ignoring certain local minima and eliminating an otherwise expensive hyperparameter. We also show that poor search directions may benefit computationally from overstepping optima along a descent direction, which can be resolved by considering improved search directions. Having shown that GOLS is a reliable line search allows for comparative investigations between static and dynamic MBSS.
Similar content being viewed by others
References
Anguita, D., Ghio, A., Oneto, L., Parra, X., Reyes-Ortiz, J.L.: Human activity recognition on smartphones using a multiclass hardware-friendly support vector machine. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 7657 LNCS, pp. 216–223 (2012). https://doi.org/10.1007/978-3-642-35395-6_30
Anitescu, M.: Degenerate nonlinear programming with a quadratic growth condition. SIAM J. Optim. 10(4), 1116–1135 (2000). https://doi.org/10.1137/S1052623499359178
Arora, J.: Introduction to Optimum Design, 3rd edn. Academic Press Inc, Cambridge (2011)
Balles, L., Hennig, P.: Dissecting Adam: the sign, magnitude and variance of stochastic gradients, vol. 1, pp. 1–17 (2018). arXiv:1705.07774v2 [cs.LG]
Bergstra, J., Bardenet, R., Bengio, Y., Kégl, B.: Algorithms for hyper-parameter optimization. In: NIPS 2011, pp. 2546–2554 (2011). arXiv:1206.2944S
Bergstra, J., Bengio, Y.: Random search for hyper-parameter optimization. J. Mach. Learn. Res. 13(February), 281–305 (2012). https://doi.org/10.1162/153244303322533223
Bertsekas, D.P.: Massachusetts Institute of Technology: Convex Optimization Algorithms, 1st edn. Athena Scientific, Belmont (2015)
Bishop, C.M.: Pattern Recognition and Machine Learning, 1st edn. Springer, Berlin (2006)
Bollapragada, R., Byrd, R., Nocedal, J.: Adaptive sampling strategies for stochastic optimization, pp. 1–32 (2017). arXiv:1710.11258
Bottou, L.: Large-scale machine learning with stochastic gradient descent. In: COMPSTAT 2010, Keynote, Invited and Contributed Papers, vol. 19, pp. 177–186 (2010). https://doi.org/10.1007/978-3-7908-2604-3-16
Byrd, R.H., Chin, G.M., Nocedal, J., Wu, Y.: Sample size selection in optimization methods for machine learning. Math. Program. 134(1), 127–155 (2012). https://doi.org/10.1007/s10107-012-0572-5
Chen, T., Sun, Y., Shi, Y., Hong, L.: On sampling strategies for neural network-based collaborative filtering, pp. 1–14 (2017). arXiv:1706.07881 [cs.LG]
Choromanska, A., Henaff, M., Mathieu, M., Arous, G.B., LeCun, Y.: The loss surfaces of multilayer networks. In: AISTATS 2015, vol. 38, pp. 192–204 (2015)
Dauphin, Y., Pascanu, R., Gulcehre, C., Cho, K., Ganguli, S., Bengio, Y.: Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. ICLR 2014, 1–9 (2014)
Davis, C.: The norm of the Schur product operation. Numer. Math. 4(1), 343–344 (1962). https://doi.org/10.1007/BF01386329
Duchi, J., Hazan, E., Singer, Y.: Adaptive subgradient methods for online learning and stochastic optimization. J. Mach. Learn. Res. 12(July), 2121–2159 (2011)
Engelbrecht, A.P.: Fundamentals of Computational Swarm Intelligence, 1st edn. Wiley, Hoboken (2005)
Fisher, R.A.: The use of multiple measurements in taxonomic problems. Ann. Eugen. 7(2), 179–188 (1936). https://doi.org/10.1111/j.1469-1809.1936.tb02137.x
Floudas, C.A., Pardalos, P.M.: Encyclopedia of Optimization, 2nd edn. Springer, Berlin (2009)
Friedlander, M.P., Schmidt, M.: Hybrid deterministic-stochastic methods for data fitting, pp. 1–26 (2011). https://doi.org/10.1137/110830629. arXiv:1104.2373 [cs.LG]
Gong, P., Ye, J.: Linear convergence of variance-reduced stochastic gradient without strong convexity (2014). arXiv:1406.1102
Goodfellow, I.J., Vinyals, O., Saxe, A.M.: Qualitatively characterizing neural network optimization problems. ICLR 2015, 1–11 (2015)
Jaderberg, M., Dalibard, V., Osindero, S., Czarnecki, W.M., Donahue, J., Razavi, A., Vinyals, O., Green, T., Dunning, I., Simonyan, K., Fernando, C., Kavukcuoglu, K.: Population based training of neural networks, pp. 1–13 (2017). arXiv:1711.09846
Johnson, B., Tateishi, R., Xie, Z.: Using geographically weighted variables for image classification. Remote Sens. Lett. 3(6), 491–499 (2012). https://doi.org/10.1080/01431161.2011.629637
Johnson, B.A., Tateishi, R., Hoan, N.T.: A hybrid pansharpening approach and multiscale object-based image analysis for mapping diseased pine and oak trees. Int. J. Remote Sens. 34(20), 6969–6982 (2013). https://doi.org/10.1080/01431161.2013.810825
Karimi, H., Nutini, J., Schmidt, M.: Linear convergence of gradient and proximal-gradient methods under the Polyak–Łojasiewicz Condition. In: ECML PKDD: Joint European Conference on Machine Learning and Knowledge Discovery in Databases, vol. 9851, pp. 795–811. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46128-1_50
Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization. ICLR 2015, 1–15 (2015). https://doi.org/10.1145/1830483.1830503
Kingma, D.P., Welling, M.: Auto-encoding variational Bayes, pp. 1–14 (2013). https://doi.org/10.1051/0004-6361/201527329. arXiv:1312.6114v10
Krizhevsky, A., Hinton, G.E.: Learning Multiple Layers of Features from Tiny Images. University of Toronto, Toronto (2009)
Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proc. IEEE 86(11), 2278–2324 (1998). https://doi.org/10.1109/5.726791
Li, M., Zhang, T., Chen, Y., Smola, A.J.: Efficient mini-batch training for stochastic optimization. In: Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1–10 (2014). https://doi.org/10.1145/2623330.2623612
Liu, J., Wright, S.J., Ré, C., Bittorf, V., Sridhar, S.: An asynchronous parallel stochastic coordinate descent algorithm. J. Mach. Learn. Res. 16, 285–322 (2015)
Lucas, D.D., Klein, R., Tannahill, J., Ivanova, D., Brandon, S., Domyancic, D., Zhang, Y.: Failure analysis of parameter-induced simulation crashes in climate models. Geosci. Model Dev. 6(4), 1157–1171 (2013). https://doi.org/10.5194/gmd-6-1157-2013
Luo, Z.Q., Tseng, P.: Error bounds and convergence analysis of feasible descent methods: a general approach. Ann. Oper. Res. 46–47(1), 157–178 (1993). https://doi.org/10.1007/BF02096261
Mahsereci, M., Hennig, P.: Probabilistic line searches for stochastic optimization. J. Mach. Learn. Res. 18, 1–59 (2017)
Mansouri, K., Ringsted, T., Ballabio, D., Todeschini, R., Consonni, V.: Quantitative structure–activity relationship models for ready biodegradability of chemicals. J. Chem. Inf. Model. 53(4), 867–878 (2013). https://doi.org/10.1021/ci4000213
Martens, J.: Deep learning via Hessian-free optimization. ICML 2010, 1–6 (2010). https://doi.org/10.1155/2011/176802
Marwala, T.: Bayesian training of neural networks using genetic programming. Pattern Recogn. Lett. 28(12), 1452–1458 (2007). https://doi.org/10.1016/J.PATREC.2007.03.004
Montana, D.J., Davis, L.: Training feedforward neural networks using genetic algorithms (1989)
Nash, W.J., Sellers, T.L., Talbot, S.R., Cawthorn, A.J., Ford, W.B.: The Population Biology of Abalone (\_Haliotis\_ species) in Tasmania. I. Blacklip Abalone (\_H. rubra\_) from the North Coast and Islands of Bass Strait. Technical report, Sea Fisheries Division (1994)
Nesterov, Y.: Primal–dual subgradient methods for convex problems. Math. Program. Ser. B 120, 221–259 (2009). https://doi.org/10.1007/s10107-007-0149-x
Paschke, F., Bayer, C., Bator, M., Mönks, U., Dicks, A., Enge-Rosenblatt, O., Lohweg, V.: Sensorlose Zustandsüberwachung an Synchronmotoren. In: Conference: 23. Workshop Computational Intelligence (VDI/VDE-Gesellschaft Mess- und Automatisierungstechnik (GMA)). Dortmund (2013)
Prechelt, L.: PROBEN1—a set of neural network benchmark problems and benchmarking rules (Technical Report 21-94). Technical report, Universität Karlsruhe (1994)
pytorch.org: PyTorch. https://pytorch.org/ (2019). Version: 1.0
Radiuk, P.M.: Impact of training set batch size on the performance of convolutional neural networks for diverse datasets. Inf. Technol. Manag. Sci. 20(1), 20–24 (2017). https://doi.org/10.1515/itms-2017-0003
Robbins, H., Monro, S.: A stochastic approximation method. Ann. Math. Stat. 22(3), 400–407 (1951). https://doi.org/10.1214/aoms/1177729586
Ruder, S.: An overview of gradient descent optimization algorithms, pp. 1–14 (2016). https://doi.org/10.1111/j.0006-341X.1999.00591.x. arXiv:1609.04747v2 [cs.LG],
Saxe, A.M., McClelland, J.L., Ganguli, S.: Exact solutions to the nonlinear dynamics of learning in deep linear neural networks, pp. 1–22 (2013). CoRR arXiv:1312.6120
Shor, N.Z.: Minimization Methods for Non-Differentiable Functions, 1st edn. Springer, Berlin (1985)
Shor, N.Z.: The subgradient method. In: Minimization Methods for Non-Differentiable Functions, pp. 22–47. Springer, Berlin (1985)
Snoek, J., Larochelle, H., Adams, R.: Practical Bayesian optimization of machine learning algorithms. In: NIPS, pp. 1–9 (2012). arXiv:1206.2944S
Snyman, J.A., Wilke, D.N.: Practical Mathematical Optimization. Springer Optimization and Its Applications, vol. 133. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-77586-9
Tong, F., Liu, X.: Samples selection for artificial neural network training in preliminary structural design. Tsinghua Sci. Technol. 10(2), 233–239 (2005). https://doi.org/10.1016/S1007-0214(05)70060-2
Vurkaç, M.: Clave-direction analysis: a new arena for educational and creative applications of music technology. J. Music Technol. Educ. 4(1), 27–46 (2011). https://doi.org/10.1386/jmte.4.1.27_1
Werbos, P.J.: The Roots of Backpropagation: From Ordered Derivatives to Neural Networks and Political Forecasting. Wiley, New York, NY (1994)
Wilke, D.N., Kok, S., Snyman, J.A., Groenwold, A.A.: Gradient-only approaches to avoid spurious local minima in unconstrained optimization. Optim. Eng. 14(2), 275–304 (2013). https://doi.org/10.1007/s11081-011-9178-7
Yeh, I.C., Lien, C.H.: The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients. Expert Syst. Appl. 36(2), 2473–2480 (2009). https://doi.org/10.1016/J.ESWA.2007.12.020
Zhang, C., Öztireli, C., Mandt, S., Salvi, G.: Active mini-batch sampling using repulsive point processes (2018). ArXiv:1804.02772
Zhang, H., Yin, W.: Gradient methods for convex minimization: better rates under weaker conditions. ArXiv e-prints (2013)
Ziȩba, M., Tomczak, S.K., Tomczak, J.M.: Ensemble boosted trees with synthetic features generation in application to bankruptcy prediction. Expert Syst. Appl. 58, 93–101 (2016). https://doi.org/10.1016/J.ESWA.2016.04.001
Zuo, X., Chintala, S.: Basic VAE example. https://github.com/pytorch/examples/tree/ master/vae (2018). Accessed on 7 May 2018
Acknowledgements
This work was supported by the Centre for Asset and Integrity Management (C-AIM), Department of Mechanical and Aeronautical Engineering, University of Pretoria, Pretoria, South Africa. We would also like to thank NVIDIA for sponsoring the Titan X Pascal GPU used in this study.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
A Artificial neural networks
The single and double hidden layer feedforward neural network architectures are expressed mathematically by Eqs. (23) and (24) respectively below. The optimization vector, \({\textit{\textbf{x}}}\) is sectioned and transformed into matrices \({\textit{\textbf{X}}}^{(c)}\) for the relevant weights in connection layers c of the network. The given data observation pair \({\textit{\textbf{t}}}_b\) is separated to give the input data, \({\textit{\textbf{T}}}^i_b\), and output data, \({\textit{\textbf{T}}}^o_b\). Suppose a given dataset has an input domain, \({\textit{\textbf{T}}}^i\) with \(|{\mathcal {B}}|\) observations and D dimensions (features). The respective output domain, \({\textit{\textbf{T}}}^o\), has corresponding observations \(|{\mathcal {B}}|\) and output dimensions E (classes). Then for every observation b and every output dimension e, a prediction of the output data \(\hat{{\textit{\textbf{T}}}}^o\) can be constructed from the original data input domain \({\textit{\textbf{T}}}^i\), given by
for a single hidden layer neural network and
for a double hidden layer neural network.
\(M_{n}\), \(n \in [1,2]\) gives number of nodes in the respective hidden layers. The nodal activation function is denoted by a and \({\textit{\textbf{X}}}^{(c)}\), \(c \in [1,2,3]\) denotes the set of weights connecting sequential layers in the network between the input layer and the output layer in a forward direction. Thus the single hidden layer network has two sets of weights, \({\textit{\textbf{X}}}^{(c)}\), and the double hidden layer network has three respectively [8].
The nodal weights \({\textit{\textbf{x}}}\) are optimized to a configuration which best captures the relationship between the input and output data spaces. The loss-function used is the mean squared error (MSE), determined over every b in batch size \(|{\textit{\textbf{B}}}|\) and every class \(e \in E\) according to the Proben1 dataset guidelines [43] as:
where \(\hat{{\textit{\textbf{T}}}}^o({\textit{\textbf{x}}})\) is the output estimation of the current network configuration as a function of the weights, and \({\textit{\textbf{T}}}^o\) is the target output of the corresponding training dataset samples.
B Exact line search: gradient-only line search with bisection (GOLS-B)
The directional derivative values used in this method are defined as \(F'_n(\alpha ) = {{\textit{\textbf{g}}}}({\textit{\textbf{x}}}_n + \alpha \cdot {\textit{\textbf{d}}}_n)^T{\textit{\textbf{d}}}_n \) and the search direction, \({\textit{\textbf{d}}}_n\), at the respective values for \(\alpha \) at the different points.
C Inexact line search: gradient-only line search that is inexact (GOLS-I)
Parameters used for this method are: \(\eta = 2\), \(c_2 = 0.9\), \(\alpha _{min} = 10^{-8}\) and \(\alpha _{max} = 10^7\). \(F'_n(\alpha ) = {{\textit{\textbf{g}}}}({\textit{\textbf{x}}}_n + \alpha \cdot {\textit{\textbf{d}}}_n)^T{\textit{\textbf{d}}}_n \).
D Inexact line search: gradient-only line search maximizing step size (GOLS-Max)
Parameters used for this method are: \(\eta = 2\), \(c_2 = 0.9\), \(\alpha _{min} = 10^{-8}\) and \(\alpha _{max} = 10^7\). \(F'_n(\alpha ) = {{\textit{\textbf{g}}}}({\textit{\textbf{x}}}_n + \alpha \cdot {\textit{\textbf{d}}}_n)^T{\textit{\textbf{d}}}_n \).
E Inexact line search: gradient-only line search with backtracking (GOLS-Back)
Parameters used for this method are: \(\eta = 2\), \(c_2 = 0\), \(\alpha _{min} = 10^{-8}\) and \(\alpha _{max} = 10^7\). \(F'_n(\alpha ) = {{\textit{\textbf{g}}}}({\textit{\textbf{x}}}_n + \alpha \cdot {\textit{\textbf{d}}}_n)^T{\textit{\textbf{d}}}_n \).
Rights and permissions
About this article
Cite this article
Kafka, D., Wilke, D.N. Resolving learning rates adaptively by locating stochastic non-negative associated gradient projection points using line searches. J Glob Optim 79, 111–152 (2021). https://doi.org/10.1007/s10898-020-00921-z
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10898-020-00921-z