# Tensor Decompositions for Modeling Inverse Dynamics

###### Abstract:

Modeling inverse dynamics is crucial for accurate feedforward robot control. The model computes the necessary joint torques, to perform a desired movement. The highly non-linear inverse function of the dynamical system can be approximated using regression techniques. We propose as regression method a tensor decomposition model that exploits the inherent three-way interaction of positions velocities accelerations. Most work in tensor factorization has addressed the decomposition of dense tensors. In this paper, we build upon the decomposition of sparse tensors, with only small amounts of nonzero entries. The decomposition of sparse tensors has successfully been used in relational learning, e.g., the modeling of large knowledge graphs. Recently, the approach has been extended to multi-class classification with discrete input variables. Representing the data in high dimensional sparse tensors enables the approximation of complex highly non-linear functions. In this paper we show how the decomposition of sparse tensors can be applied to regression problems. Furthermore, we extend the method to continuous inputs, by learning a mapping from the continuous inputs to the latent representations of the tensor decomposition, using basis functions. We evaluate our proposed model on a dataset with trajectories from a seven degrees of freedom SARCOS robot arm. Our experimental results show superior performance of the proposed functional tensor model, compared to challenging state-of-the art methods.

First]Stephan Baier Second]Volker Tresp

Ludwig Maximilian University of Munich, Oettingenstr. 67, 80538 Munich (e-mail: stephan.baier@campus.lmu.de).

Siemens AG and Ludwig Maximilian University of Munich, Otto-Hahn-Ring 6, 81739 Munich (e-mail: volker.tresp@siemens.com)

Keywords: Tensor modeling, tensor decomposition, inverse dynamics, robot dynamics, supervised machine learning

Within model-based robot control, an inverse dynamics model is used to compute the necessary joint torques of the robot’s motors for the execution of a desired movement. The feedforward control command can be calculated using the rigid-body formulation , with being vectors of joint positions, joint velocities, and joint accelerations. However, in practice many nonlinearities such as friction or actuator forces need to be taken into account. Thus, methods modeling using non-linear regression techniques have shown superior performance in inferring the required joint torques for feedforward robot control. The parameters of the function are estimated offline using collected trajectories of the robot. Craig (2005); Nguyen-Tuong et al. (2008); Nakanishi et al. (2005).

Tensor models have been applied successfully in many application areas, e.g., relational learning, multilinear time invariant systems, factor analysis, and spatio-temporal analysis, see Nickel et al. (2011); Pangalos et al. (2013b); Mørup et al. (2006); Bahadori et al. (2014). Most literature on tensor modeling, however, is concerned with the decomposition of dense tensors, i.e., most of the elements in the tensor are nonzero. Models for sparse tensors have mainly become popular for the application of modeling large knowledge graphs, such as Yago, DBpedia, and Freebase, see Nickel et al. (2011); Suchanek et al. (2007); Auer et al. (2007). In these models, the elements of the tensor represent all possible triple combinations of entities and relations in the knowledge graph. Only elements that represent known facts from the knowledge graph are set to one. This results in a very sparse tensor, where the vast majority of elements are zero. Recently, the approach has been extended to higher order tensors for the task of classifying discrete sensor data, see Baier et al. (2016). The tensor represents the space of all possible combinations of sensor values. By learning a representation for each possible value of all sensors, the decomposition allows for approximating highly non-linear functions.

In this paper we build upon the approach of decomposing sparse tensors, and apply it to inverse system identification. Our model exploits the inherent three-way interaction of positions velocities accelerations. We first show how the method can be applied to regression tasks. Furthermore, we extend the approach to continuous inputs, by including basis functions that map the continuous inputs to the latent representations of the tensor decompositions. In this way, we retrieve a functional version of tensor decompositions. The basis functions also imply smoothness on the inputs, such that the model is able to generalize well, in spite of the extreme sparsity. By using multivariate basis functions we can group inputs, such that the dimensionality of the tensor decomposition can be reduced. In our inverse dynamics model we group the joint positions, velocities, and accelerations of all degrees of freedom of the robot, resulting in a tensor of order three. This makes the powerful Tucker decomposition applicable to the problem.

We evaluate our model on a dataset of a seven degrees of freedom SARCOS robot arm that was introduced in Vijayakumar and Schaal (2000). An inverse dynamics model is learned based on collected trajectories, and its performance is evaluated on a 10 percent test set. The results show that our model outperforms a number of competitive baseline methods, such as linear regression, radial basis function networks (RBF-networks), and support vector regression. Furthermore, the Tucker model shows superior performance over a PARAFAC model.

The paper is structured as follows. The next section gives an overview of related work. Section Tensor Decompositions for Modeling Inverse Dynamics shows how the factorization of sparse tensors can be utilized for regression problems, and how the tensor decompositions can be extended to continuous inputs, using basis functions. In Section Tensor Decompositions for Modeling Inverse Dynamics we describe a functional Tucker decomposition for the task of modeling inverse dynamics. Related work is discussed in Section Tensor Decompositions for Modeling Inverse Dynamics. Section Tensor Decompositions for Modeling Inverse Dynamics presents the experimental evaluation. Finally, we conclude our work in Section Tensor Decompositions for Modeling Inverse Dynamics.

In this section we first show how the decomposition of sparse tensors can be applied to regression problems with discrete input variables. We then extend the model to continuous inputs, by using basis functions, which map the continuous input to the latent representations of the tensor decompositions.

Tensor decompositions are a generalization of low rank matrix factorizations to higher order tensors. There are multiple ways of decomposing a higher order tensor.

The full Tucker decomposition factorizes a tensor into matrices, including latent representations for all fibers in each mode. The tensor elements are expressed by the interaction of the latent representations, weighted by a core tensor such that

(1) |

with . The full Tucker decomposition does not scale to high dimensions, as the core tensor grows exponentially with the dimensionality of the tensor; see Tucker (1965).

A special case of the full Tucker decomposition is the PARAFAC decomposition, where the core tensor is diagonal. All other interactions are left out, such that

(2) |

with . As PARAFAC only models the diagonal of the core tensor, its parameters scale linearly with the order of the tensor; see Harshman (1970).

We consider a regression problem with discrete input variables. Each of the input variables for assumes one out of discrete values. Furthermore, we consider a dependent variable . We model a regression function for a dataset of training examples .

All training examples are mapped to a sparse tensor . The tensor is filled with

(3) |

The remaining entries of the tensor, which do not occur in the training data, are left unknown. This results in being a very sparse tensor.

The tensor is approximated using a low-rank tensor decomposition, e.g., the PARAFAC decomposition see equation Tensor Decompositions for Modeling Inverse Dynamics. Using low ranks for , the approximation results in a dense tensor . It describes the outcome for all combinations of the input variables . However, it would be impossible to compute and store the whole approximated tensor ; thus, only the parameters of the decomposition are stored. When predicting for a new set of input variables, the representations for that tuple are indexed, and the approximation is computed on demand. In principle any tensor decomposition can be used for the approximation. However, in practice only few decompositions, such as PARAFAC and Tensor Train are scale-able to many dimensions, see Harshman (1970); Oseledets (2011).

The proposed model so far only works for a discrete input space. Furthermore, it does not imply any smoothness on the values of the input variables. Although, this makes it a powerful, highly non-linear model, it is prone to overfitting. If the input values follow a natural ordering, or if they are discretized from a continuous scale, the model requires many more training examples to learn the smoothness implicitly. To introduce smoothness explicitly, and to extend the model to continuous inputs, we use smooth basis functions for the latent parameters of the decomposition. Instead of indexing the latent representation from a matrix, their values are computed using basis functions. For example, all in equation Tensor Decompositions for Modeling Inverse Dynamics can be modeled using a radial basis function

(4) |

This allows for continuous inputs . The latent representation is now modeled by the similarity of the input to the center of the radial basis function. In this way, similar inputs induce similar representations. The parameters of the basis function are optimized during training, to yield optimal regression results. Also a mixture of discrete and continuous inputs can easily be modeled, by applying the basis functions only to the continuous inputs, and learning representation matrices for the discrete input variables.

It is also possible to group multiple inputs together into one tensor mode, such that , where denotes the number of grouped inputs. In this way, the representation of a tensor mode is calculated given a vector of continous inputs. The grouping of input variables reduces the dimensionality of the tensor decomposition, and thus the number of free parameters.

In the following we describe how the continuous tensor decomposition proposed in section Tensor Decompositions for Modeling Inverse Dynamics can be applied to inverse dynamics modeling.

We describe a functional Tucker model for the approximation of the joint torques, necessary to perform a movement of a robot arm. Figure 1 shows the model schematically. We consider a robot with degrees of freedom (DoF). In the following we denote the vectors , describing the desired positions, velocities, and accelerations for each of the DoFs, as for syntactic reasons. The vector describes the corresponding joint torques.

We model the function using a functional tensor decomposition model. Each input vector is modeled by one dimension in the tensor decomposition, resulting in third-order tensors , which describe the joint torques. Each element of the vector is modeled in a separate model. The resulting three-dimensional tensors of the form positions velocities accelerations, are then factorized using the Tucker decomposition with limited rank, resulting in a tensor , such that

(5) |

to are functions, which map from the -dimensional input to the latent representations of the Tucker model. We model the representations using multivariate Gaussian kernels, such that

(6) |

with representing the centers and weighting the distance from the centers in the dimensional input space. The closer a data point is to the center of a basis function, the higher is its activation. Thus, the centers of the basis functions can be seen as landmarks in the input space. All three-way interactions, between the representations of the three input dimensions, are explicitly modeled and weighted by the elements of the core tensor .

Method | DoF 1 | DoF 2 | DoF 3 | DoF 4 | DoF 5 | DoF 6 | DoF 7 | Mean std in % |

Linear Regression | 6.80 | 11.62 | 10.82 | 5.81 | 12.81 | 22.59 | 6.73 | 11.03 0.26 |

RBF-Network Regression | 2.64 | 1.79 | 1.01 | 0.41 | 4.07 | 3.91 | 1.17 | 2.14 0.19 |

Support Vector Regression | 0.88 | 0.67 | 0.43 | 0.15 | 1.04 | 0.72 | 0.34 | 0.60 0.28 |

Functional-Tucker | 0.59 | 0.28 | 0.46 | 0.24 | 1.03 | 0.91 | 0.31 | 0.55 0.24 |

Functional-PARAFAC | 1.64 | 1.14 | 0.61 | 0.32 | 1.30 | 1.17 | 0.50 | 0.96 0.22 |

For training the model, we take a maximum likelihood approach. We minimize the negative log-likelihood of the collected dataset as

(7) |

where includes the parameters of the decomposition and the basis functions. Assuming a Gaussian distribution, we get the squared error cost function

(8) |

Note, that the cost function considers only nonzero elements of the tensor, i.e., the sparsity of the tensor is exploited. We minimize equation Tensor Decompositions for Modeling Inverse Dynamics using gradient descent. In experiments, we found the stochastic optimization algorithm Adam, see Kingma and Ba (2014), which adopts the learning rate automatically for each parameter, to work best for this task. The sampling of stochastic mini-batches for each update has also shown advantageous, for speeding up training.

We initialize the centers of the Gaussian kernel in a preprocessing step, using three k-means clusterings, such that

(9) |

are minimized for , see Lloyd (1982). All matrices are initialized with the identity matrix. The elements of the core tensor are initialized randomly with a Gaussian distribution of mean zero and standard deviation 0.05. While training all parameters are further optimized. We implemented the model using the computational python library Theano, see Theano Development Team (2016).

Multiway data analysis has found applications in a number of different areas, such as signal processing, neuroscience, and data mining, see Cichocki (2014); Mørup et al. (2006); Harshman (1970); Kolda and Bader (2009). Recently, tensor models also have found applications in control engineering such as for modeling hybrid systems, see Pangalos et al. (2013a) and multilinear dynamical systems, see Rogers et al. (2013). Furthermore, tensor methods have been applied to Boolean networks, see Cheng et al. (2010) and pneumatic models, see Gróf et al. (2010).

The factorization of sparse matrices has become popular in recommendation systems, especially due to its success in the Netflix challenge, see Koren et al. (2009). Extensions to higher order tensors can be found in the modeling of large knowledge bases, such as Yago, DBpedia, or Freebase, see Nickel et al. (2011, 2015). The multi-graphs have been modeled using sparse three-dimensional tensors and decompositions such as RESCAL, see Nickel et al. (2011, 2015). The approach of factorizing sparse tensors has further been exploited in Baier et al. (2016). Here, the decomposition of sparse tensors is applied to multi-class classification with discrete input features.

Tensor regression methods are concerned with the regression of high dimensional data, structured in a multidimensional array. Tensor methods allow for efficient modeling where traditional methods are often insufficient, due to the complex structure of the data and the high input dimensionality. Tensor regression learns a linear mapping and deals with dense input tensors. Thus, their approach is fundamentally different from ours; see Zhou et al. (2013); Yu and Liu (2016).

Our approach shows some similarities to RBF-networks which are able to approximate any non-linear function by using radial basis functions. RBF-networks have been successfully applied to a number of tasks including control engineering, see Broomhead and Lowe (1988). The main difference to our proposed functional Tucker model is that RBF-networks learn one latent representation for the complete input, and map it to the output; whereas, the functional Tucker model learns a representation for each tensor mode and jointes them using the tensor decomposition model. In this way multi-way interactions are modeled explicitly.

Inverse dynamics are traditionally modeled using the rigid-body formulation, see Craig (2005). However, general regression techniques such as locally weighted projection regression (LWPR), Gaussian Processes, and RBF-networks, have shown advantageous for learning inverse dynamics, see Vijayakumar and Schaal (2000); Rasmussen (2006). The topic was subject to a number of studies, see Burdet and Codourey (1998); Nguyen-Tuong et al. (2008). Support vector regression has shown superior performance for this task.

In this section we evaluate our proposed method on an inverse dynamics dataset including movements from a seven degrees of freedom SARCOS robot arm. We compare against various other state-of-the-art regression techniques for this task.

The dataset was introduced by Vijayakumar and Schaal (2000).\@footnotemark\@footnotetexthttp://www.gaussianprocess.org/gpml/data/ It contains data from a SARCOS robot arm with seven degrees of freedom. The data was collected from the moving robot arm at 100Hz and corresponds to 7.5 minutes of movement. The dataset includes 21 input dimensions, consisting of 7 joint torques, 7 joint positions, 7 joint velocities, and 7 joint accelerations. The whole dataset consists of 42482 samples. We split the dataset randomly into 90 percent training and 10 percent test data. Additional 5 percent of the training set where used as a validation set. The task is to learn a model on the training data, which models the 7 joint torques, given the positions, velocities and accelerations. The offline learned model can then be applied in the forward controller of the robot. The dataset has been subject to some studies on the topic, see Vijayakumar and Schaal (2000); Rasmussen (2006). The regression task has been found to be highly non-linear. Non-linear regression techniques outperformed the rigid-body dynamics formulation by a large margin. The performance of the regression techniques is evaluated on the test set, which includes unseen movements. We repeated the random split 10 times and report the average results and the standard deviation of multiple trials.

We compare our model against various state-of-the art regression techniques, modeling the function . The baseline models we consider are linear regression, RBF-networks and support vector regression. In previous studies support vector regression has shown the best results on this task. For all baseline models a concatenated vector is built. The linear regression model learns a linear mapping from the inputs to the outputs, such that

(10) |

RBF-networks model the regression problem as,

(11) |

The parameters , and are learned using backpropagation. We initialized the parameters with the centroids of a k-means clustering on the training data, where is the number of centroids.

Support vector regression (see Smola and Vapnik (1997)) has shown state-of-the-art results in modeling inverse dynamics. It predicts as,

(12) |

with being a kernel function. In the experiments we use a Gaussian kernel. and are Lagrange multipliers, which are determined during optimization. In our experiments we use the libsvm library, see Chang and Lin (2011).

Furthermore, we compare the functional Tucker model proposed in Section Tensor Decompositions for Modeling Inverse Dynamics with a functional PARAFAC model. For the functional PARAFAC model we replace the tensor decomposition in equation Tensor Decompositions for Modeling Inverse Dynamics with a PARAFAC decomposition, as shown in equation Tensor Decompositions for Modeling Inverse Dynamics.

We report the normalized mean squared error (nMSE) for the regression task, which is defined as the mean squared error of all data points divided by the variance of the target variable in the training data. Table 1 summarizes the mean nMSE for all seven degrees of freedom in percent. In the rightmost column the mean of all seven degrees of freedom is shown. All results, as well as the standard deviation are referring to the average result of 10 random data splits. The performance of the regression techniques varies across the DoFs. The linear model reaches an nMSE of 11.03% in average. The nonlinear RBF-networks performs much better with an nMSE of 2.14% in average. The number of of hidden neurons for the RBF-network was set to 1000. With larger numbers the predictive performance did not increase. The support vector regression model yields a very good result of 0.60%. Here, we set the parameter C to 600 and to 0.1. All hyperparameters were evaluated on a separate validation set. Our proposed functional Tucker model resulted in a slightly better nMSE of 0.55%. Especially, for the first two DoFs the functional Tucker model performs significantly better than support vector regression. For the other DoFs the results of support vector regression and functional Tucker decomposition are very close to each other. The parameter efficient functional PARAFAC model reaches an nMSE of 0.96% in average. Figure 2 shows the performance of the two functional tensor decomposition models in dependence of the rank of the decompositions. For the Tucker model, the performance converges at a rank of 30 and for the PARAFAC model at a rank of 40. It is also notable that both methods already perform relatively well with a very small rank of 5. The nMSE of the Tucker model is 2.09% with a rank of 5 and the nMSE of the PARAFAC model is 2.43%. Both functional tensor models show clearly better results than RBF-networks. This indicates that the explicit modeling of the three-way interaction, yields a significant improvement.

In this paper we apply a tensor model, that is based on the Tucker decomposition, to inverse dynamics. Our proposed model exploits the inherent three-way interaction of positions velocities accelerations. We show how the decomposition of sparse tensors can be applied to regression tasks. Furthermore, we propose to augment the tensor decompositions with basis functions for allowing continuous input variables. In this way, a functional version of a tensor decomposition can be derived. Representations for each tensor mode are induced through the basis functions and fused by the tensor model. The parameters of the basis functions are learned using backpropagation. Experiments on an inverse dynamics dataset, derived from a seven degrees of freedom robot arm, show promising results of our proposed model for the application of learning inverse dynamics. The proposed functional Tucker model outperforms RBF-networks, and even support vector regression, which has shown state-of-the-art performance on this task. Our extension of tensor decomposition models to continuous inputs enables a wide range of application areas. Especially if an inherent multi-way structure exists in the data, functional tensor models can be advantageous over traditional techniques, by explicitly modeling the multi-way interaction. Within automatic robot control the approach might be further extended to learning also a functional Tucker model for a feedback controller based on the tracking errors.

## References

- Auer et al. (2007) Auer, S., Bizer, C., Kobilarov, G., Lehmann, J., Cyganiak, R., and Ives, Z. (2007). Dbpedia: A nucleus for a web of open data. In The semantic web, 722–735. Springer.
- Bahadori et al. (2014) Bahadori, M.T., Yu, Q.R., and Liu, Y. (2014). Fast multivariate spatio-temporal analysis via low rank tensor learning. In Advances in neural information processing systems, 3491–3499.
- Baier et al. (2016) Baier, S., Krompass, D., and Tresp, V. (2016). Learning representations for discrete sensor networks using tensor decompositions. International Conference on Multisensor Fusion and Integration for Intelligent Systems.
- Broomhead and Lowe (1988) Broomhead, D.S. and Lowe, D. (1988). Radial basis functions, multi-variable functional interpolation and adaptive networks. Technical report, DTIC Document.
- Burdet and Codourey (1998) Burdet, E. and Codourey, A. (1998). Evaluation of parametric and nonparametric nonlinear adaptive controllers. Robotica, 16(01), 59–73.
- Chang and Lin (2011) Chang, C.C. and Lin, C.J. (2011). Libsvm: a library for support vector machines. ACM Transactions on Intelligent Systems and Technology (TIST), 2(3), 27.
- Cheng et al. (2010) Cheng, D., Qi, H., and Li, Z. (2010). Analysis and control of Boolean networks: a semi-tensor product approach. Springer Science & Business Media.
- Cichocki (2014) Cichocki, A. (2014). Tensor networks for big data analytics and large-scale optimization problems. arXiv preprint arXiv:1407.3124.
- Craig (2005) Craig, J.J. (2005). Introduction to robotics: mechanics and control, volume 3. Pearson Prentice Hall Upper Saddle River.
- Gróf et al. (2010) Gróf, P., Takarics, B., Petres, Z., and Korondi, P. (2010). Tensor product model type polytopic decomposition of a pneumatic system with friction phenomena taken into account. In Applied Machine Intelligence and Informatics (SAMI), 2010 IEEE 8th International Symposium on, 153–158. IEEE.
- Harshman (1970) Harshman, R.A. (1970). Foundations of the parafac procedure: Models and conditions for an” explanatory” multi-modal factor analysis.
- Kingma and Ba (2014) Kingma, D. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
- Kolda and Bader (2009) Kolda, T.G. and Bader, B.W. (2009). Tensor decompositions and applications. SIAM review, 51(3), 455–500.
- Koren et al. (2009) Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, (8), 30–37.
- Lloyd (1982) Lloyd, S. (1982). Least squares quantization in pcm. IEEE transactions on information theory, 28(2), 129–137.
- Mørup et al. (2006) Mørup, M., Hansen, L.K., Herrmann, C.S., Parnas, J., and Arnfred, S.M. (2006). Parallel factor analysis as an exploratory tool for wavelet transformed event-related eeg. NeuroImage, 29(3), 938–947.
- Nakanishi et al. (2005) Nakanishi, J., Farrell, J.A., and Schaal, S. (2005). Composite adaptive control with locally weighted statistical learning. Neural Networks, 18(1), 71–90.
- Nguyen-Tuong et al. (2008) Nguyen-Tuong, D., Peters, J., Seeger, M., and Schölkopf, B. (2008). Learning inverse dynamics: a comparison. In European Symposium on Artificial Neural Networks, EPFL-CONF-175477.
- Nickel et al. (2015) Nickel, M., Murphy, K., Tresp, V., and Gabrilovich, E. (2015). A review of relational machine learning for knowledge graphs: From multi-relational link prediction to automated knowledge graph construction. arXiv preprint arXiv:1503.00759.
- Nickel et al. (2011) Nickel, M., Tresp, V., and Kriegel, H.P. (2011). A three-way model for collective learning on multi-relational data. In Proceedings of the 28th international conference on machine learning (ICML-11), 809–816.
- Oseledets (2011) Oseledets, I.V. (2011). Tensor-train decomposition. SIAM Journal on Scientific Computing, 33(5), 2295–2317.
- Pangalos et al. (2013a) Pangalos, G., Eichler, A., and Lichtenberg, G. (2013a). Tensor systems - multilinear modeling and applications. In Proceedings of the 3rd International Conference on Simulation and Modeling Methodologies, Technologies and Applications, 275–285. doi:10.5220/0004475602750285.
- Pangalos et al. (2013b) Pangalos, G., Eichler, A., and Lichtenberg, G. (2013b). Tensor systems-multilinear modeling and applications. In SIMULTECH, 275–285.
- Rasmussen (2006) Rasmussen, C.E. (2006). Gaussian processes for machine learning.
- Rogers et al. (2013) Rogers, M., Li, L., and Russell, S.J. (2013). Multilinear dynamical systems for tensor time series. In Advances in Neural Information Processing Systems, 2634–2642.
- Smola and Vapnik (1997) Smola, A. and Vapnik, V. (1997). Support vector regression machines. Advances in neural information processing systems, 9, 155–161.
- Suchanek et al. (2007) Suchanek, F.M., Kasneci, G., and Weikum, G. (2007). Yago: a core of semantic knowledge. In Proceedings of the 16th international conference on World Wide Web, 697–706. ACM.
- Theano Development Team (2016) Theano Development Team (2016). Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints, abs/1605.02688. URL http://arxiv.org/abs/1605.02688.
- Tucker (1965) Tucker, L.R. (1965). Some Mathematical Notes on Three-mode Factor Analysis [by]. Urbana, Department of Psychology, University of Illinois.
- Vijayakumar and Schaal (2000) Vijayakumar, S. and Schaal, S. (2000). Locally weighted projection regression: An o (n) algorithm for incremental real time learning in high dimensional space. In International conference on machine learning, proceedings of the sixteenth conference.
- Yu and Liu (2016) Yu, R. and Liu, Y. (2016). Learning from multiway data: Simple and efficient tensor regression. In Proceedings of the 33nd International Conference on Machine Learning (ICML-16), 238–247.
- Zhou et al. (2013) Zhou, H., Li, L., and Zhu, H. (2013). Tensor regression with applications in neuroimaging data analysis. Journal of the American Statistical Association, 108(502), 540–552.