Practical Bayesian Optimization of Machine Learning Algorithms

Introduction

Machine learning algorithms are rarely parameter-free; whether via the properties of a regularizer, the hyperprior of a generative model, or the step size of a gradient-based optimization, learning procedures almost always require a set of high-level choices that significantly impact generalization performance. As a practitioner, one is usually able to specify the general framework of an inductive bias much more easily than the particular weighting that it should have relative to training data. As a result, these high-level parameters are often considered a nuisance, making it desirable to develop algorithms with as few of these “knobs” as possible.

Another, more flexible take on this issue is to view the optimization of high-level parameters as a procedure to be automated. Specifically, we could view such tuning as the optimization of an unknown black-box function that reflects generalization performance and invoke algorithms developed for such problems. These optimization problems have a somewhat different flavor than the low-level objectives one often encounters as part of a training procedure: here function evaluations are very expensive, as they involve running the primary machine learning algorithm to completion. In this setting where function evaluations are expensive, it is desirable to spend computational time making better choices about where to seek the best parameters. Bayesian optimization (Mockus et al., 1978) provides an elegant approach and has been shown to outperform other state of the art global optimization algorithms on a number of challenging optimization benchmark functions (Jones, 2001). For continuous functions, Bayesian optimization typically works by assuming the unknown function was sampled from a Gaussian process (GP) and maintains a posterior distribution for this function as observations are made. In our case, these observations are the measure of generalization performance under different settings of the hyperparameters we wish to optimize. To pick the hyperparameters of the next experiment, one can optimize the expected improvement (EI) (Mockus et al., 1978) over the current best result or the Gaussian process upper confidence bound (UCB)(Srinivas et al., 2010). EI and UCB have been shown to be efficient in the number of function evaluations required to find the global optimum of many multimodal black-box functions (Bull, 2011; Srinivas et al., 2010).

Machine learning algorithms, however, have certain characteristics that distinguish them from other black-box optimization problems. First, each function evaluation can require a variable amount of time: training a small neural network with 10 hidden units will take less time than a bigger network with 1000 hidden units. Even without considering duration, the advent of cloud computing makes it possible to quantify economically the cost of requiring large-memory machines for learning, changing the actual cost in dollars of an experiment with a different number of hidden units. It is desirable to understand how to include a concept of cost into the optimization procedure. Second, machine learning experiments are often run in parallel, on multiple cores or machines. We would like to build Bayesian optimization procedures that can take advantage of this parallelism to reach better solutions more quickly.

In this work, our first contribution is the identification of good practices for Bayesian optimization of machine learning algorithms. In particular, we argue that a fully Bayesian treatment of the GP kernel parameters is of critical importance to robust results, in contrast to the more standard procedure of optimizing hyperparameters (e.g. Bergstra et al. (2011)). We also examine the impact of the kernel itself and examine whether the default choice of the squared-exponential covariance function is appropriate. Our second contribution is the description of a new algorithm that accounts for cost in experiments. Finally, we also propose an algorithm that can take advantage of multiple cores to run machine learning experiments in parallel.

Bayesian Optimization with Gaussian Process Priors

There are two major choices that must be made when performing Bayesian optimization. First, one must select a prior over functions that will express assumptions about the function being optimized. For this we choose the Gaussian process prior, due to its flexibility and tractability. Second, we must choose an acquisition function, which is used to construct a utility function from the model posterior, allowing us to determine the next point to evaluate.

2 Acquisition Functions for Bayesian Optimization

One intuitive strategy is to maximize the probability of improving over the best current value (Kushner, 1964). Under the GP this can be computed analytically as

Expected Improvement

Alternatively, one could choose to maximize the expected improvement (EI) over the current best. This also has closed form under the Gaussian process:

GP Upper Confidence Bound

A more recent development is the idea of exploiting lower confidence bounds (upper, when considering maximization) to construct acquisition functions that minimize regret over the course of their optimization (Srinivas et al., 2010). These acquisition functions have the form

with a tunable $\kappa$ to balance exploitation against exploration.

In this work we will focus on the expected improvement criterion, as it has been shown to be better-behaved than probability of improvement, but unlike the method of GP upper confidence bounds (GP-UCB), it does not require its own tuning parameter. We have found expected improvement to perform well in minimization problems, but wish to note that the regret formalization is more appropriate for many settings. We perform a direct comparison between our EI-based approach and GP-UCB in Section 4.1.

Practical Considerations for Bayesian Optimization of Hyperparameters

Although an elegant framework for optimizing expensive functions, there are several limitations that have prevented it from becoming a widely-used technique for optimizing hyperparameters in machine learning problems. First, it is unclear for practical problems what an appropriate choice is for the covariance function and its associated hyperparameters. Second, as the function evaluation itself may involve a time-consuming optimization procedure, problems may vary significantly in duration and this should be taken into account. Third, optimization algorithms should take advantage of multi-core parallelism in order to map well onto modern computational environments. In this section, we propose solutions to each of these issues.

The power of the Gaussian process to express a rich distribution on functions rests solely on the shoulders of the covariance function. While non-degenerate covariance functions correspond to infinite bases, they nevertheless can correspond to strong assumptions regarding likely functions. In particular, the automatic relevance determination (ARD) squared exponential kernel

is often a default choice for Gaussian process regression. However, sample functions with this covariance function are unrealistically smooth for practical optimization problems. We instead propose the use of the ARD Matérn $5/2$ kernel:

This covariance function results in sample functions which are twice differentiable, an assumption that corresponds to those made by, e.g., quasi-Newton methods, but without requiring the smoothness of the squared exponential.

After choosing the form of the covariance, we must also manage the hyperparameters that govern its behavior (Note that these “hyperparameters” are different than the ones which are being subjected to the overall Bayesian optimization.), as well as that of the mean function. For our problems of interest, typically we would have ${D+3}$ Gaussian process hyperparameters: $D$ length scales $\theta_{1:D}$, the covariance amplitude $\theta_{0}$, the observation noise $\nu$, and a constant mean $m$. The most commonly advocated approach is to use a point estimate of these parameters by optimizing the marginal likelihood under the Gaussian process

However, for a fully-Bayesian treatment of hyperparameters (summarized here by $\theta$ alone), it is desirable to marginalize over hyperparameters and compute the integrated acquisition function:

2 Modeling Costs

Ultimately, the objective of Bayesian optimization is to find a good setting of our hyperparameters as quickly as possible. Greedy acquisition procedures such as expected improvement try to make the best progress possible in the next function evaluation. From a practial point of view, however, we are not so concerned with function evaluations as with wallclock time. Different regions of the parameter space may result in vastly different execution times, due to varying regularization, learning rates, etc. To improve our performance in terms of wallclock time, we propose optimizing with the expected improvement per second, which prefers to acquire points that are not only likely to be good, but that are also likely to be evaluated quickly. This notion of cost can be naturally generalized to other budgeted resources, such as reagents or money.

3 Monte Carlo Acquisition for Parallelizing Bayesian Optimization

Empirical Analyses

In this section, we empirically analyseAll experiments were conducted on identical machines using the Amazon EC2 service. the algorithms introduced in this paper and compare to existing strategies and human performance on a number of challenging machine learning problems. We refer to our method of expected improvement while marginalizing GP hyperparameters as “GP EI MCMC”, optimizing hyperparameters as “GP EI Opt”, EI per second as “GP EI per Second”, and $N$ times parallelized GP EI MCMC as “$N$x GP EI MCMC”.

2 Online LDA

Latent Dirichlet allocation (LDA) is a directed graphical model for documents in which words are generated from a mixture of multinomial “topic” distributions. Variational Bayes is a popular paradigm for learning and, recently, Hoffman et al. (2010) proposed an online learning approach in that context. Online LDA requires two learning parameters, $\tau_{0}$ and $\kappa$, that control the learning rate ${\rho_{t}=(\tau_{0}+t)^{-\kappa}}$ used to update the variational parameters of LDA based on the $t^{\rm th}$ minibatch of document word count vectors. The size of the minibatch is also a third parameter that must be chosen. Hoffman et al. (2010) relied on an exhaustive grid search of size ${6\times 6\times 8}$, for a total of 288 hyperparameter configurations.

We used the code made publically available by Hoffman et al. (2010) to run experiments with online LDA on a collection of Wikipedia articles. We downloaded a random set of 249,560 articles, split into training, validation and test sets of size 200,000, 24,560 and 25,000 respectively. The documents are represented as vectors of word counts from a vocabulary of 7,702 words. As reported in Hoffman et al. (2010), we used a lower bound on the per word perplixity of the validation set documents as the performance measure. One must also specify the number of topics and the hyperparameters $\eta$ for the symmetric Dirichlet prior over the topic distributions and $\alpha$ for the symmetric Dirichlet prior over the per document topic mixing weights. We followed Hoffman et al. (2010) and used 100 topics and ${\eta=\alpha=0.01}$ in our experiments in order to emulate their analysis and repeated exactly the grid search reported in the paperi.e. the only difference was the randomly sampled collection of articles in the data set and the choice of the vocabulary. We ran each evaluation for 10 hours or until convergence.. Each online LDA evaluation generally took between five to ten hours to converge, thus the grid search requires approximately 60 to 120 processor days to complete.

In Figures 4a and 4b we compare our various strategies of optimization over the same grid on this expensive problem. That is, the algorithms were restricted to only the exact parameter settings as evaluated by the grid search. Each optimization was then repeated one hundred times (each time picking two different random experiments to initialize the optimization with) and the mean and standard error are reported. Figures 4a and 4b respectively show the average minimum loss (perplexity) achieved by each strategy compared to the number of times online LDA is evaluated with new parameter settings and the duration of the optimization in days. Figure 4c shows the average loss of 3 and 5 times parallelized GP EI MCMC which are restricted to the same grid as compared to a single run of the same algorithms where the algorithm can flexibly choose new parameter settings within the same range by optimizing the expected improvement.

In this case, integrating over hyperparameters is superior to using a point estimate. While GP EI MCMC is the most efficient in terms of function evaluations, we see that parallelized GP EI MCMC finds the best parameters in significantly less time. Finally, in Figure 4c we see that the parallelized GP EI MCMC algorithms find a significantly better minimum value than was found in the grid search used by Hoffman et al. (2010) while running a fraction of the number of experiments.

3 Motif Finding with Structured Support Vector Machines

In this example, we consider optimizing the learning parameters of Max-Margin Min-Entropy (M3E) Models (Miller et al., 2012), which include Latent Structured Support Vector Machines (Yu and Joachims, 2009) as a special case. Latent structured SVMs outperform SVMs on problems where they can explicitly model problem-dependent hidden variables. A popular example task is the binary classification of protein DNA sequences (Miller et al., 2012; Kumar et al., 2010; Yu and Joachims, 2009). The hidden variable to be modeled is the unknown location of particular subsequences, or motifs, that are indicators of positive sequences.

Setting the hyperparameters, such as the regularisation term, $C$, of structured SVMs remains a challenge and these are typically set through a time consuming grid search procedure as is done in Miller et al. (2012) and Yu and Joachims (2009). Indeed, Kumar et al. (2010) report that hyperparameter selection was avoided for the motif finding task due to being too computationally expensive. However, Miller et al. (2012) demonstrate that classification results depend highly on the setting of the parameters, which differ for each protein.

M3E models introduce an entropy term, parameterized by $\alpha$, which enables the model to significantly outperform latent structured SVMs. This additional performance, however, comes at the expense of an additional problem-dependent hyperparameter. We emulate the experiments of Miller et al. (2012) for one protein with approximately 40,000 sequences. We explore 25 settings of the parameter $C$, on a log scale from $10^{-1}$ to $10^{6}$, 14 settings of $\alpha$, on a log scale from 0.1 to 5 and the model convergence tolerance, ${\epsilon\in\{10^{-4},\,10^{-3},\,10^{-2},\,10^{-1}\}}$. We ran a grid search over the 1,400 possible combinations of these parameters, evaluating each over 5 random 50-50 training and test splits.

In Figures 5a and 5b, we compare the randomized grid search to GP EI MCMC, GP EI per Second and their 3x parallelized versions, all constrained to the same points on the grid, in terms of minimum validation error vs wallclock time and function evaluations. Each algorithm was repeated 100 times and the mean and standard error are shown. We observe that the Bayesian optimization strategies are considerably more efficient than grid search which is the status quo. In this case, GP EI MCMC is superior to GP EI per Second in terms of function evaluations but GP EI per Second finds better parameters faster than GP EI MCMC as it learns to use a less strict convergence tolerance early on while exploring the other parameters. Indeed, 3x GP EI per second is the least efficient in terms of function evaluations but finds better parameters faster than all the other algorithms.

Figure 5c compares the use of various covariance functions in GP EI MCMC optimization on this problem. The optimization was repeated for each covariance 100 times and the mean and standard error are shown. It is clear that the selection of an appropriate covariance significantly affects performance and the estimation of length scale parameters is critical. The assumption of the infinite differentiability of the underlying function as imposed by the commonly used squared exponential is too restrictive for this problem.

4 Convolutional Networks on CIFAR-10

Neural networks and deep learning methods notoriously require careful tuning of numerous hyperparameters. Multi-layer convolutional neural networks are an example of such a model for which a thorough exploration of architechtures and hyperparameters is beneficial, as demonstrated in Saxe et al. (2011), but often computationally prohibitive. While Saxe et al. (2011) demonstrate a methodology for efficiently exploring model architechtures, numerous hyperparameters, such as regularisation parameters, remain. In this empirical analysis, we tune nine hyperparameters of a three-layer convolutional network, described in Krizhevsky (2009) on the CIFAR-10 benchmark dataset using the code providedAvailable at: http://code.google.com/p/cuda-convnet/ using the architechture defined in http://code.google.com/p/cuda-convnet/source/browse/trunk/example-layers/layers-18pct.cfg. This model has been carefully tuned by a human expert (Krizhevsky, 2009) to achieve a highly competitive result of 18% test error, which matches the published state of the artWithout augmenting the training data. result (Coates and Ng, 2011) on CIFAR-10. The parameters we explore include the number of epochs to run the model, the learning rate, four weight costs (one for each layer and the softmax output weights), and the width, scale and power of the response normalization on the pooling layers of the network.

We optimize over the nine parameters for each strategy on a withheld validation set and report the mean validation error and standard error over five separate randomly initialized runs. Results are presented in Figure 6 and contrasted with the average results achieved using the best parameters found by the expert. The best hyperparametersThe optimized parameters deviate interestingly from the expert-determined settings; e.g., the optimal weight costs are asymmetric (the weight cost of the second layer is approximately an order of magnitude smaller than the first layer), a learning rate two orders of magnitude smaller, a slightly wider response normalization, larger scale and much smaller power. found by the GP EI MCMC approach achieve an error on the test set of $14.98\%$, which is over 3% better than the expert and the state of the art on CIFAR-10.

Conclusion

In this paper we presented methods for performing Bayesian optimization of hyperparameters associated with general machine learning algorithms. We introduced a fully Bayesian treatment for expected improvement, and algorithms for dealing with variable time regimes and parallelized experiments. Our empirical analysis demonstrates the effectiveness of our approaches on three challenging recently published problems spanning different areas of machine learning. The code used will be made publicly available. The resulting Bayesian optimization finds better hyperparameters significantly faster than the approaches used by the authors. Indeed our algorithms surpassed a human expert at selecting hyperparameters on the competitive CIFAR-10 dataset and as a result beat the state of the art by over 3%.

This work was supported by a grant from Amazon Web Services and by a DARPA Young Faculty Award.

References