1 Introduction
Many machine learning algorithms, including neural networks, can be divided into three parts: the model, which is used to describe or approximate the structure present in the training data set; the loss function, which defines how well an instance of the model fits the samples; and the optimization method, which adjusts the model’s parameters to improve the error expressed by the loss function. Obviously these three parts are related, and the generalization capability of the obtained solution depends on the individual merit of each one of the three parts, and also on their interplay.
Most of current research in machine learning focuses on creating new models [1, 2], for the different applications and data types, and new optimization methods [3], which may allow faster convergence, more robustness, and a better chance to escape from poor local minima. However, focus is rarely attributed to the cost functions used in different applications.
Many cost functions come from statistical models [4], such as the quadratic error or crossentropy. These cost functions, in their linear formulations, usually define convex loss functions [4, 5]. Moreover, when building the statistical model of a sample set, frequently the total cost becomes the sum of losses for each sample. Although this methodology is sound, it can be problematic in realworld applications involving more complex models.
The first problem, discussed more extensively in Sec. 2, is that linear combination of losses can reduce the number of solutions achievable by the optimization algorithm. The second problem is that this cost may not reflect the behavior expected by the user.
Consider, for instance, a classification problem with samples and and two choices of parameters and . For
, the classifier has confidences
and for the correct classes of and , which means that is classified incorrectly. For , they are classified correctly with confidences and , respectively. Most machine learning users would prefer using , because has a highconfidence correct prediction in both cases while is classified incorrectly if is used. Nonetheless, the mean crossentropy loss, which is the standard cost for classification problems, places higher cost to than to , so the optimizer will choose .Although this toy example may not be representative of all the tradeoffs involved, it clearly reflects the conflict between what the user wants and what the cost function expresses. One possible solution to this problem is the use of boosting techniques to increase the weight of incorrectly classified samples and training multiple classifiers, as discussed in Sec. 3, but this requires multiple models and may not be robust to noise [6]. Moreover, this solution still restricts the solutions achievable by the optimization algorithm due to the linear combination of losses.
This paper presents a solution to this problem by evaluating the modelfitting problem in a multiobjective optimization perspective, of which the standard linear combination of losses is a particular case, and uses another method for transforming the problem into a singleobjective optimization, which allows the same standard optimization algorithms to be used. It will be shown that the gradient of the new objective automatically provides higher weights for samples with higher losses, like boosting methods, but uses a single model, with the current parameters working as the previous model in boosting. It is important to highlight that, although this paper focuses on optimization by gradient, the method proposed also works for hillclimbing on discrete problems.
The author’s conjecture that automatically placing more learning pressure on samples with high losses may improve the overall generalization, because it may remove small bumps in the error surface. Therefore, the multiobjective formulation allows better minima to be reached. Experimental results presented in this paper provide evidence for this conjecture.
The paper is organized as follows. Section 2 provides an overview of multiobjective optimization, properly characterizing the maximization of the hypervolume as a performance indicator for the learning system, given the data set and the loss function. Section 3 deals with a gradientbased approach to maximize the hypervolume, and shows that the gradient operator promotes the selfadjustment of the weights associated with each loss function. Sections 4 and 5 describe the experiments performed and the results obtained, respectively. Finally, Section 6 provides the concluding remarks and future research directions.
2 Multiobjective optimization
Multiobjective optimization is a generalization of the traditional singleobjective optimization, where the problem is composed of multiple objective functions , where [7]. Using the standard notation, the problem can be described by:
where is the decision space and includes all constraints of the optimization.
If some of the objectives have the same minima, then the redundant objectives can be ignored during optimization. However, if their minima are different, for example and , then there is not a single optimal point, but a set of different tradeoffs between the objectives. A solution that establishes an optimal tradeoff, that is, it is impossible to reduce one of the objectives without increasing another, is said to be efficient. The set of efficient solutions is called the Pareto set and its counterpart in the objective space is called the Pareto frontier.
A common approach in optimization used to deal with multiobjective problems is to combine the objectives linearly [7, 8], so that the problem becomes
(1) 
where the weight represents the importance given to objective .
Although the optimal solution of the linearly combined problem is guaranteed to be efficient, it is only possible to achieve any efficient solution when the Pareto frontier is convex [8]. This means that more desired solutions may not be achievable by performing a linear combination of the objectives.
As an illustrative example, consider the problem where and the Pareto frontier is almost linear, with and , but slightly concave. Any combination of weights and will only be able to provide or , depending on the specific weights, but a solution close to 0.5 might be more aligned with the tradeoffs expected by the user.
2.1 Hypervolume indicator
Since the linear combination of objectives is not going to work properly on nonconvex Pareto frontiers, it is desirable to investigate other forms of transforming the multiobjective problem into a singleobjective one, which allows the standard optimization tools to be used.
One common approach in the multiobjective literature is to resort to the hypervolume indicator [9], which is frequently used to analyze a set of candidate solutions [10] and can be expensive in such cases [11]. However for a single solution, its logarithm can be written as:
(2) 
where is called the Nadir point and . The problem then becomes maximizing the hypervolume over the domain, and this optimization is able to achieve a larger number of efficient points, without requiring convexity of the Pareto frontier [12].
Among the many properties of the hypervolume, two must be highlighted in this paper. The first is that the hypervolume is monotonic in the objectives, which means that any improvement in any objective causes the hypervolume to increase, which in turn is aligned with the loss minimization. The maximum of the singlesolution hypervolume is a point in the Pareto set, which means that the solution is efficient.
The second property is that, like the linear combination, it also maintains some shape information from the objectives. If the objectives are convex, then their linear combination is convex and the hypervolume is concave, since is concave and the logarithm of a concave function is concave.
Therefore, the hypervolume has the same number of parameters as the linear combination, one for each objective, and also maintains optimality, but may be able to achieve solutions more aligned with the user’s expectations.
Consider the previous example with and . If , which indicates that both objectives are equally important a priori, then the hypervolume maximization will indeed find a solution close to . So the parameter defines prior inverse importances that are closer to what one would expect from weights: if they are equal, the solver should try to provide a balanced relevance to the objectives.
Therefore, the hypervolume maximization with a single model tries to find a good solution to all objectives at the same time, achieving the best compromise possible given the prior inverse importances . Note that, contrary to the weights in the linear combination, the values actually reflect the user’s intent for the solution found in the optimization.
2.2 Loss minimization
A standard objective in machine learning is minimization of some loss function over a given data set
. Note that this notation includes both supervised and unsupervised learning, as the space
can include both the samples and their targets.Using the multiobjective notation, the loss minimization problem is described as:
Just like in other areas of optimization, the usual approach to solve these problems in machine learning is the use of linear combination of the objectives, as shown in Eq. (1). Common examples of this method are using the mean loss as the objective to be minimized and defining a regularization weight [4], where the regularization characterizes other objectives.
However, as discussed in Sec. 2, this approach limits the number of solutions that can be obtained, which motivates the use of the hypervolume indicator. Since the objectives differ only in the samples used for the loss function and considering that all samples have equal importance^{1}^{1}1This is the same motivation for using the uniform mean loss. If prior importance is available, it can be used to define the value of , like it would be used to define in the weighted mean loss., the Nadir point can have the same value for all objectives so that the solution found is balanced in the losses. This value is given by the parameter , so that . Then the problem becomes
where
(3) 
which, as stated in Sec. 2.1, maintains the convexity information of the loss function but has a single parameter to be chosen along the iterative steps of the optimization.
As will be shown in Sec. 3, smaller values of place learning pressure on samples with large losses, improving the worstcase scenario, while large values of approximates learning to the uniform mean loss. And, of course, as learning progresses, the relative hardness of each sample may vary significantly.
3 Gradient as the operator for selfadjusting the weights
As exemplified in Sec. 1, the standard metric of average error can be problematic by providing excessive confidence to some samples while preventing lessconfident samples to improve. A solution to this problem is to use different weights, as in Eq. (1), with higher weights for harder samples instead of using the same weight for all samples.
However, one may not know a priori which samples are harder, which prevents the straightforward use of the weighted mean. One classical solution to this problem in classification is the use of boosting techniques [13], where a set of classifiers are learnt instead of a single one and they are merged into an ensemble. When training a new classifier, the errors of the previous ones on each sample are taken into account to increase the weights of incorrectly labeled samples, placing more pressure on the new classifier to predict these samples’ labels correctly.
The requirement of multiple models, to form the boosting ensemble, can be hard to train with very large models and may be too slow to be used in real problems, since multiple models have to be evaluated. Therefore, a single model that is able to selfadjust the weights so that higher weights are going to be assigned to currently harder samples is desired.
Taking the gradient of the losses hypervolume’s logarithm, shown in Eq. (3), one finds
(4) 
which is similar to a weighted mean loss cost with weights replacing the minimization of the loss by the maximization of the hypervolume. However, the weights are determined automatically as a function of , so that they do not have to be defined a priori and can change during the learning algorithm’s execution. Moreover, higher cost implies higher weight, that is,
(5) 
which is expected, as more pressure must be placed on samples with higher losses.
Therefore, hypervolume maximization using gradient adjusts the parameters in the same direction as weighted mean loss minimization with appropriate weights. Furthermore, as approaches , the worst case becomes the main influence in the gradient and the problem becomes similar to minimizing the maximum loss, while as approaches infinity, the problem becomes similar to minimizing the uniform mean loss. Thus a single parameter is able to control the problem’s placement between these two extremes.
Moreover, the hypervolume’s logarithm, show in Eq. (2), is just a sum of the different objectives, like the linear combination shown in Eq. (1
). This allows the same optimization methods to be used for maximizing the hypervolume, which includes methods that deal with large data sets and scalability issues, such as the stochastic gradient descent with minibatches.
4 Experiment
To evaluate the performance of maximizing the hypervolume, experiments were carried out using a denoising autoencoder [14], which is a neural network with a hidden layer that tries to reconstruct the input at the output, given a corrupted version of the input. The purpose is to reconstruct the digits in the MNIST data set, with 50000 training samples and 10000 validation and test samples each, as usual with this data set. Since the uniform mean loss is the usual objective for this kind of problem, it is used as a baseline, establishing a reference for comparison with the new proposed formulation, and as the indicator for comparison between the baseline and the hypervolume method.
In order to be able to use the same learning rate for the hypervolume and mean loss problems, and since the mean loss already has normalized weights, the parameter’s adjustment step was divided by the sum of the weights computed using the hypervolume gradient in order to normalize the weights.
The hypervolume maximization depends on the definition of the parameter , as discussed in Sec. 3. Since it must be larger than the objectives when the gradient is computed, the parameter used in the experiments is defined by:
(6) 
where
is the epoch number,
are the parameters at epoch , is the initial slack constant, and is a userchosen constant that can decrease the learning pressure over bad examples, as the number of epochs increase. For the stochastic gradient using minibatches, corresponds to the parameters at that point, but is only increased after the full epoch is completed.This way of defining the Nadir point removes the requirement of choosing an absolutely good value by allowing a value relative to the current losses to be determined, making its choice easier. The parameters were arbitrarily chosen as and for the experiments, so that the problem becomes more similar to the mean loss problem as the learning progresses, since the mean loss is used as the performance indicator.
The optimization was performed using gradient descent with learning rate and minibatch size of samples for epochs, with the crossentropy as the loss function. A total of runs with different initializations of the random number generator were used.
Saltandpepper noise was added to the training images with varying probability
, but with equal probabilities of black or white corruption, and hidden units were used in the denoising autoencoder. Notice that the number of inputs, and consequently the number of outputs, is , since the digits are represented by images. In order to evaluate the performance, the mean and maximum losses over the training, validation, and test sets were computed without corruption of the inputs.5 Results
Figure 1 shows the mean loss over the iterations for the training, validation, and test sets of the MNIST data set for different corruption probabilities . The plotted values are the difference between the standard mean loss minimization, used as a baseline, and the mean loss when maximizing the hypervolume. Positive values mean that the hypervolume maximization provides lower mean losses, which corresponds to better results. The same initial conditions and noises were used for the baseline and the proposed method, so that the difference in performance in a single run is caused by the difference of objectives and is not caused by different random numbers.
From Fig. 1, it is clear that the hypervolume method provides better results than the baseline for all corruption levels, indicating a better fit of the parameters. Moreover, the difference becomes larger as the noise level increases, indicating that the proposed method is more robust and guides to better generalization.
It is important to highlight that the hypervolume method achieves better performance than the baseline in a metric different from its optimization objective. This indicates that the higher weight imposed on samples with high loss improves the overall generalization, which is aligned with the conjecture proposed in Sec. 1.
Corruption level  Mean loss  Hypervolume 

0.0  53.604 (0.022)  53.523 (0.021) 
0.1  57.269 (0.056)  57.101 (0.057) 
0.2  64.227 (0.131)  63.723 (0.137) 
0.3  71.385 (0.207)  70.556 (0.181) 
0.4  78.160 (0.358)  77.136 (0.389) 
Test mean losses at the iteration in which the validation set had smallest mean loss. Mean and standard deviations over 50 runs are shown. All differences are statistically significant (
).Another effect noted in Fig. 1 is that the noise level influences the overall behavior of the difference between the methods. For , the hypervolume method always achieves better results faster than the baseline and most of the times for . As the number of iterations increase, the difference between the two methods decreases for , and increases for . Notice also that the difference in performance is more favorable to the hypervolume method when the noise level gets higher. If this occurred only for the training set, it would indicate an overfitting problem. But since it occurs on all sets, it means that the hypervolume method is able to cope better with higher noise levels, increasing the generalization.
Figure 2 shows the maximum loss over the different parts of the data set as the learning advances. For all parts of the data set, in some cases the worstcase loss is increased by the hypervolume method, but the median shows that improvement happens in most cases, with possibly large improvements, as shown by the upper bounds. The increase in the worstcase loss in the training data set is explained by the fact that this performance is measured on the noiseless data, while the training happens on the noisy data.
Table 1
shows the resulting mean losses for both methods on the test set. The point of evaluation was chosen as the one where the noiseless validation set presented the smallest mean loss. Clearly the hypervolume method provides better results than the commonly used mean loss for all corruption levels, and the differences are significant at less than 0.1% using the paired ttest.
6 Conclusion
This paper presented the problem of minimizing the loss function over a data set from a multiobjective optimization perspective. It discussed the issues of using linear combination of loss functions as the target of minimization in machine learning, proposing the use of another metric, the hypervolume indicator, as the target to be optimized. This metric preserves optimality conditions of the losses and may achieve tradeoffs between the losses more aligned to the user’s expectations.
It is also shown that the gradient of this metric is equivalent to the gradient of a weighted mean loss, but without requiring the weights to be determined a priori. Instead, the weights are determined automatically by the gradient and possesses a boostinglike behavior, with the losses for the current set of parameters establishing the weights’ values.
Experiments with the MNIST data set and denoising autoencoders show that the new proposed objective achieves better mean loss over the training, validation, and test data. This increase in performance occurs even though the performance indicator is different from the objective being optimized, which could cause the problem to converge to an optimal solution to the problem but with low performance in the mean loss metric. This indicates that the hypervolume maximization is similar to the mean loss minimization, but may be able to achieve better minima due to the change in the loss landscape, as conjectured in Sec. 1.
Results also show that the maximum loss is almost always reduced in the training, validation, and test sets, with a slight reduction in some cases and large improvements on others. Moreover, by using the validation set to find the best set of parameters, the hypervolume method achieved statistically significant smaller mean loss in the test set, indicating that it was able to generalize better.
Future research directions involve analyzing the effect of including regularization terms, which usually are added to the linear combination of loss functions, as new objectives. The use of this method for larger problems and models and in other learning settings, such as multitask learning, where separating tasks in different objectives is natural, should also be pursued.
Acknowledgments
The authors would like to thanks CNPq for the financial support.
References
References
 [1] D. Koller and N. Friedman. Probabilistic Graphical Models: Principles and Techniques. MIT press, 2009.
 [2] Y. Bengio. Learning deep architectures for AI. Foundations and trends® in Machine Learning, 2(1):1–127, 2009.
 [3] K. P. Bennett and E. ParradoHernández. The interplay of optimization and machine learning research. The Journal of Machine Learning Research, 7:1265–1281, 2006.
 [4] C. M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
 [5] A. Banerjee, S. Merugu, I. S Dhillon, and J. Ghosh. Clustering with Bregman divergences. The Journal of Machine Learning Research, 6:1705–1749, 2005.
 [6] P. M. Long and R. A. Servedio. Random classification noise defeats all convex potential boosters. Machine Learning, 78(3):287–304, 2010.
 [7] K. Deb. Multiobjective optimization. In Search methodologies, pages 403–449. Springer, 2014.
 [8] S. P. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004.
 [9] E. Zitzler, D. Brockhoff, and L. Thiele. The hypervolume indicator revisited: On the design of Paretocompliant indicators via weighted integration. In Evolutionary multicriterion optimization, pages 862–876. Springer, 2007.
 [10] E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G. Da Fonseca. Performance Assessment of Multiobjective Optimizers: An Analysis and Review. Evolutionary Computation, IEEE Transactions on, 7(2):117–132, 2003.
 [11] N. Beume, C. M. Fonseca, M. LópezIbáñez, L. Paquete, and J. Vahrenhold. On the complexity of computing the hypervolume indicator. Evolutionary Computation, IEEE Transactions on, 13(5):1075–1082, 2009.

[12]
A. Auger, J. Bader, D. Brockhoff, and E. Zitzler.
Theory of the hypervolume indicator: optimal distributions and
the choice of the reference point.
In
Proceedings of the tenth ACM SIGEVO workshop on Foundations of genetic algorithms
, pages 87–102. ACM, 2009.  [13] R. E. Schapire. The strength of weak learnability. Machine learning, 5(2):197–227, 1990.
 [14] P. Vincent, H. Larochelle, Y. Bengio, and P.A. Manzagol. Extracting and Composing Robust Features with Denoising Autoencoders. In International Conference on Machine Learning, pages 1096–1103. ACM, 2008.
Comments
There are no comments yet.