Learning Invariant Features Using Subspace Restricted Boltzmann Machine
 1.4k Downloads
 5 Citations
Abstract
The subspace restricted Boltzmann machine (subspaceRBM) is a thirdorder Boltzmann machine where multiplicative interactions are between one visible and two hidden units. There are two kinds of hidden units, namely, gate units and subspace units. The subspace units reflect variations of a pattern in data and the gate unit is responsible for activating the subspace units. Additionally, the gate unit can be seen as a pooling feature. We evaluate the behavior of subspaceRBM through experiments with MNIST digit recognition task and Caltech 101 Silhouettes image corpora, measuring crossentropy reconstruction error and classification error.
Keywords
Feature learning Unsupervised learning Invariant features Subspace features Deep model1 Introduction
The success of machine learning methods stems from appropriate data representation. Clearly this requires applying feature engineering, i.e., handcrafted proposition of a set of features potentially useful in the considered problem. However, it would be beneficial to propose an automatic features extraction to avoid any awkward preprocessing pipelines for handtuning of the data representation [2]. Deep learning turns out to be a suitable fashion of automatic representation learning in many domains such as object recognition [23], speech recognition [20], natural language processing [7], neuroimaging [13] multimodal learning from images and text annotations [27], pose recovery [29], or domain adaptation [9].
Fairly simple but still one of the most popular models for unsupervised feature learning is the restricted Boltzmann machine (RBM). Except automatic feature learning, RBMs can be stacked in a hierarchy to form a deep network [1]. The bipartie structure of the RBM enables block Gibbs sampling which allows formulating efficient learning algorithms such as contrastive divergence [10]. However, lately it has been argued that the RBM fails to properly reflect statistical dependencies [22]. One possible solution is to apply higherorder Boltzmann machine [17, 24] to model sophisticated patterns in data.
In this work we follow this line of thinking and develop a more refined model than the RBM to learn features from data. Our model introduces two kinds of hidden units, i.e., subspace units and gate units (see Fig. 1). The subspace units are hidden variables which reflect variations of a feature and thus they are more robust to invariances. The gate units are responsible for activating the subspace units and they can be seen as pooling features composed of the subspace features. The proposed model is based on an energy function with thirdorder interactions and maintains the conditional independence structure that can be readily used in simple and efficient learning.
2 The Model
3 Learning
In training, we take advantage of the Eqs. 4, 5, and 6 to formulate an efficient threephase blockGibbs sampling from the subspaceRBM (see Algorithm 1). First, for given data, we sample gate units from \(p(\mathbf {h}\mathbf {x})\) with \(\mathbf {S}\) marginalized out. Then, given both \(\mathbf {x}\) and \(\mathbf {h}\), we can sample subspace variables from \(p(\mathbf {S}\mathbf {x}, \mathbf {h})\). Eventually, the data can be sampled from \(p(\mathbf {x} \mathbf {h}, \mathbf {S})\).
4 Related Works
The standard RBM can reflect only the secondorder multiplicative interactions. However, in many reallife situations, higherorder interactions must be included if we want our model to be effective enough. Moreover, often the secondorder interactions themselves might represent little or no useful information. In the literature there were several propositions of how to extend the RBM to the higherorder Boltzmann machines. One such proposal is a thirdorder multiplicative interaction of two visible binary units \(x_{i}\), \(x_{i'}\) and one hidden binary unit \(h_{j}\) [11, 22], which can be used to learn a representation robust to spatial transformations [19]. Along this line of thinking, our model is the thirdorder Boltzmann machine but with different multiplicative interactions of one visible unit and two kinds of hidden units.
The proposed model is closely related to the special kind of spikeandslab restricted Boltzmann machine [6] called the subspace spikeandslab RBM (subspacessRBM) [5] where there are two kinds of hidden variables, namely, spike is a binary variable and slab is a realvalued variable. However, in our approach both the spike and slab variables are discrete. Additionally, in the subspaceRBM the hidden units \(\mathbf {h}\) behave as gates to subspace variables rather than spikes as in ssRBM.
Similarly to our approach, gating units were proposed in the Pointwise Gated Boltzmann Machine (PGBM) [25] where chosen units were responsible for switching on subsets of hidden units. The subspaceRBM is based on an analogous idea but it uses sigmoid units only whereas PGBM utilizes both sigmoid and softmax units.
Our model can be also related to RBM forests [15]. The RBM forests assume each hidden unit to be encoded by a complete binary tree. In our approach each gate unit is encoded by subspace units. Therefore, the subspaceRBM can be seen as a RBM forest but with flatter hierarchy of hidden units and hence easier learning and inference.
Lastly, the subspaceRBM but with the softmax hidden units \({\mathbf {h}}\) turns to be the implicit mixture of RBMs (imRBM) [21]. However, in our model the gate units can be seen as pooling features while in the imRBM they determine only one subset of subspace features to be activated. The subspaceRBM brings an important benefit over the imRBM because it allows the subspaceRBM to reflect multiple factors in data.
5 Experiment

Is the subspaceRBM preferable to the RBM in terms of reconstruction error and as a better feature extractor?
Data We performed the experiment using CalTech 101 \(28 \times 28\) Silhouettes^{3} (CalTech, for the sake of brevity), and MNIST.^{4} CalTech dataset consists of 4100 training images, 2264 validation images, and 2307 test images. In the dataset the objects are centered and scaled on a \(28 \times 28\) image plane and rendered as filled black regions on a white background [18]. MNIST consists of \(28 \times 28\) images representing handwritten digits from 0 through 9 [16]. The data is divided into 50,000 training examples, 10,000 validation images, and the test set contains 10,000 examples. In the experiments, we performed learning with different number of training images (10,100, and 1000 per digit) and the full training set.
Training protocol In the experiment, we compared the subspaceRBM with the RBM for the number of gate units equal \(M = 500\) and different number of subspace units \(K \in \{3, 5, 7\}\). The subspaceRBM was trained using the presented contrastive divergence (see Algorithm 1) and a minibatch of size 10 was used. In order to choose the value of the learning rate we performed the model selection using the validation set and the learning rate was \(\{0.001, 0.01, 0.1\}\). The number of iterations (epochs) over the training set was determined using early stopping according to the validation set crossentropy reconstruction error, with a look ahead of 5 iterations.
The RBM was used with 500, 1500, 2500 and 3500 hidden units, which corresponds to the same number of gates units multiplied by the number of subspace units in the subspaceRBM. The RBM was trained using the contrastive divergence with 1step Gibbs sampling. The learning rate was determined using the model selection using the validation set and its possible values were the same as in the case of the subspaceRBM. Similarly to the subspaceRBM, the early stopping procedure was used with looking ahead of 5 epochs.
We evaluated the subspaceRBM as a featureextraction scheme by plugging it into the classification pipeline developed by [4]. For classification the logistic regression^{5} used the probabilities of gate units, \(p(h_{j}=1\mathbf {x})\), as inputs. Analogously was done for the RBM.
We did 3 full runs for each dataset and averaged the results.
Average test classification error with one standard deviation for the RBM and different settings of the subspaceRBM evaluated on subsets of MNIST
Classification error (\(\%\))  

Model  \(N=100\)  \(N=1000\)  \(N=10{,}000\)  \(N=50{,}000\) 
RBM \(M=500\)  24.20 \(\pm \) 0.53  8.31 \(\pm \) 0.31  3.78 \(\pm \) 0.11  3.23 \(\pm \) 0.14 
RBM \(M=1500\)  25.83 \(\pm \) 0.75  8.17 \(\pm \) 0.17  3.06 \(\pm \) 0.02  2.12 \(\pm \) 0.02 
RBM \(M=2500\)  27.40 \(\pm \) 0.86  7.77 \(\pm \) 0.11  3.01 \(\pm \) 0.05  1.94 \(\pm \) 0.12 
RBM \(M=3500\)  30.65 \(\pm \) 0.44  8.28 \(\pm \) 0.13  2.93 \(\pm \) 0.04  1.8 \(\pm \) 0.01 
subspaceRBM \(M=500\), \(K=3\)  23.78 \(\pm \) 0.53  8.95 \(\pm \) 0.50  4.27 \(\pm \) 0.62  3.81 \(\pm \) 0.19 
subspaceRBM \(M=500\), \(K=5\)  24.40 \(\pm \) 0.37  8.33 \(\pm \) 0.13  3.69 \(\pm \) 0.03  3.18 \(\pm \) 0.14 
subspaceRBM \(M=500\), \(K=7\)  24.41 \(\pm \) 1.23  8.75 \(\pm \) 0.04  3.62 \(\pm \) 0.02  2.63 \(\pm \) 0.04 
We would like to highlight that in the experiment we aim at evaluating capabilities of the proposed model and comparing it with the RBM. Therefore, we resigned from applying sophisticated learning techniques, e.g., weight decay, momentum term, sparsity regularization [12]. We believe that application of more advanced training protocol could disrupt this comparison. As a consequence, we have obtained results that were worst than current stateoftheart but these allow to evaluate mainly models instead of learning algorithms.
5.1 Results
MNIST The averaged results with one standard deviation of the subspaceRBM and the RBM are presented in Table 1 (for test classification error), in Table 2 (for test reconstruction error), and the average number of active units calculated on test data is outlined in Table 3. A random subset of subspace features for the subspaceRBM (\(M=500\), \(K=7\)) trained on 50,000 images is shown in Fig. 2.
Average test reconstruction error with one standard deviation for different settings of the RBM and the subspaceRBM evaluated on subsets of MNIST
Reconstruction error  

Model  \(N=100\)  \(N=1000\)  \(N=10{,}000\)  \(N=50{,}000\) 
RBM \(M=500\)  140.35 \(\pm \) 9.31  87.15 \(\pm \) 3.29  75.03 \(\pm \) 2.57  73.41 \(\pm \) 0.59 
RBM \(M=1500\)  144.47 \(\pm \) 16.76  89.88 \(\pm \) 0.59  75.13 \(\pm \) 0.057  72.37 \(\pm \) 0.22 
RBM \(M=2500\)  161.75 \(\pm \) 11.32  88.00 \(\pm \) 0.51  75.98 \(\pm \) 0.11  71.98 \(\pm \) 0.39 
RBM \(M=3500\)  230.72 \(\pm \) 8.48  91.71 \(\pm \) 0.08  75.76 \(\pm \) 0.27  71.17 \(\pm \) 0.01 
subspaceRBM \(M=500\), \(K=3\)  123.28 \(\pm \) 2.35  82.26 \(\pm \) 1.43  71.76 \(\pm \) 0.89  71.57 \(\pm \) 0.54 
subspaceRBM \(M=500\), \(K=5\)  121.27 \(\pm \) 1.26  81.66 \(\pm \) 0.75  71.86 \(\pm \) 0.43  69.35 \(\pm \) 1.54 
subspaceRBM \(M=500\), \(K=7\)  123.70 \(\pm \) 0.77  82.67 \(\pm \) 0.76  71.23 \(\pm \) 1.06  68.64 \(\pm \) 1.56 
Number of active units for the RBM and different settings of the subspaceRBM evaluated on subsets of MNIST
Number of active units  

Model  \(N=100\)  \(N=1000\)  \(N=10{,}000\)  \(N=50{,}000\) 
RBM \(M=500\)  70  65  50  37 
RBM \(M=1500\)  60  60  43  36 
RBM \(M=2500\)  45  60  44  36 
RBM \(M=3500\)  32  63  44  34 
subspaceRBM \(M=500\), \(K=3\)  62  93  85  120 
subspaceRBM \(M=500\), \(K=5\)  58  74  78  107 
subspaceRBM \(M=500\), \(K=7\)  41  66  66  79 
Average test results with one standard deviation for different settings of the RBM and the subspaceRBM evaluated on CalTech
Model  Classification  Reconstruction  Number of 

error (\(\%\))  error  active units  
RBM \(M=500\)  35.03 \(\pm \) 0.38  103.05 \(\pm \) 2.16  85 
RBM \(M=1500\)  37.34 \(\pm \) 1.06  115.68 \(\pm \) 6.28  85 
RBM \(M=2500\)  40.29 \(\pm \) 1.71  105.26 \(\pm \) 7.03  82 
RBM \(M=3500\)  46.97 \(\pm \) 7.82  110.58 \(\pm \) 8.34  75 
subspaceRBM \(M=500\), \(K=3\)  34.51 \(\pm \) 0.16  69.09 \(\pm \) 10.93  202 
subspaceRBM \(M=500\), \(K=5\)  35.97 \(\pm \) 0.81  66.76 \(\pm \) 8.42  175 
subspaceRBM \(M=500\), \(K=7\)  37.37 \(\pm \) 1.44  67.99 \(\pm \) 5.53  112 
5.2 Discussion
We notice that application of subspace units is beneficial for better reconstruction capabilities (see Tables 2 and 4). For classification it is advantageous to use subspaceRBM in the case of small sample size regime (for MNIST dataset with N equal 100 and 1000, and for CalTech data, see Tables 1 and 4) with smaller number of subspace units. However, this result is rather not surprising because for overcomplete representations simpler classifiers work better. On the other hand, for the small sample size there is a big threat of overfitting. Introducing subspace units to the hidden layer restricts the variability of the representation and thus preventing from learning noise in data. In the case of classification for larger number of observations (for MNIST data with N equal 10,000 and 50,000, see Table 1), best results were obtained for K equal 5 and 7. This result suggests that indeed the subspace units lead to features that are more robust to small perturbations.
Comparing the subspaceRBM to the RBM with comparable size, i.e., \(M \in \{1500, 2500, 3500\}\), it turns out that in terms of the classification error RBMs with larger number of hidden units obtained much better results. However, this result follows from the fact that it is easier to discriminate if there are more available features. Of course, this statement is true only if the features represent reasonable patterns (i.e., different than noise), and the sample size is appropriate (see Table 1 for \(N=100\) and Table 4 where larger RBMs tend to be heavily overfitted). Nonetheless, the reconstruction error for any dataset is in favor of the subspaceRBM. This effect can be explained as follows. During reconstructing data lots of features are useless but they still contribute to the reconstruction but rather as a source of noise. Therefore, the more features are in a model, the more noise is incorporated to the reconstruction. However, it seems that the larger number of subspace units results in better reconstruction. This means that the subspaceRBM indeed captures different forms of a feature and incorporating more subspace units is beneficial.
Eventually, it is worth noticing that on average the number of active hidden units is higher for the subspaceRBM in comparison to the RBM. This result may be explained by the sum of softplus terms used in calculating the conditional probability (see Eq. 6). The effect of increased activity of hidden units is especially apparent in the case of CalTech where on average about half of gate units are active (see Table 2).
6 Conclusion
In this paper, we have proposed an extension of the RBM by introducing subspace hidden units. The formulated model can be seen as the thirdorder Boltzmann machine with thirdorder multiplicative interactions. We have showed that the subspaceRBM does not reduce to a vanilla version of the RBM (see Eq. 7). The carriedout experiments have revealed that the proposed model is advantageous over the RBM in terms of reconstruction and classification error.
We see several possible extensions of the outlined approach. In our opinion, the examination of the effect of high activity of gate units is very appealing. It has been advocated [9] that sparse activity of hidden units provides more robust representation, therefore, we plan to apply some kind regularization enforcing sparsity [14] or features robustness [26]. Moreover, it would be beneficial to utilize other learning algorithms instead of the contrastitve divergence, such as, sampling methods [3], score matching [28] and other inductive principles, e.g., Maximum PseudoLikelihood [18]. Last but not least, subspaceRBM can be used as a building block in a deep model. However, we leave investigation of stated issues as future research.
Footnotes
Notes
Acknowledgments
The research conducted by the authors has been partially cofinanced by the Ministry of Science and Higher Education, Republic of Poland, namely, Jakub M. Tomczak: Grant No. B50106W8/K3, Adam Gonczarek: Grant No. B50137W8/K3.
References
 1.Bengio Y (2009) Learning deep architectures for AI. Found Trends Mach Learn 2(1):1–127CrossRefzbMATHGoogle Scholar
 2.Bengio Y, Courville A, Vincent P (2013) Representation learning: a review and new perspectives. IEEE Trans Pattern Anal Mach Intell 35(8):1798–1828CrossRefGoogle Scholar
 3.Brügge K, Fischer A, Igel C (2013) The flipthestate transition operator for restricted Boltzmann machines. Mach Learn 93(1):53–69MathSciNetCrossRefzbMATHGoogle Scholar
 4.Coates A, Ng AY, Lee H (2011) An analysis of singlelayer networks in unsupervised feature learning. In: International conference on artificial intelligence and statistics, pp 215–223Google Scholar
 5.Courville A, Desjardins G, Bergstra J, Bengio Y (2014) The spikeandslab RBM and extensions to discrete and sparse data distributions. IEEE Trans Pattern Anal Mach Intell 36(9):1874–1887CrossRefGoogle Scholar
 6.Courville AC, Bergstra J, Bengio Y (2011) A spike and slab restricted Boltzmann machine. In: International conference on artificial intelligence and statistics, pp 233–241Google Scholar
 7.Dahl GE, Adams RP, Larochelle H (2012) Training restricted Boltzmann machines on word observations. In: Proceedings of the 29th international conference on machine learningGoogle Scholar
 8.Fischer A, Igel C (2014) Training restricted Boltzmann machines: an introduction. Pattern Recognit 47(1):25–39CrossRefzbMATHGoogle Scholar
 9.Glorot X, Bordes A, Bengio Y (2011) Deep sparse rectifier networks. In: Proceedings of the 14th international conference on artificial intelligence and statistics, vol 15, pp 315–323Google Scholar
 10.Hinton GE (2002) Training products of experts by minimizing contrastive divergence. Neural Comput 14(8):1771–1800CrossRefzbMATHGoogle Scholar
 11.Hinton GE (2010) Learning to represent visual input. Philos Trans R Soc B 365(1537):177–184CrossRefGoogle Scholar
 12.Hinton GE (2012) A practical guide to training restricted Boltzmann machines. In: Montavon G, Orr GB, Müller KR (eds) Neural networks: tricks of the trade. Springer, Berlin, pp 599–619Google Scholar
 13.Hjelm RD, Calhoun VD, Salakhutdinov R, Allen EA, Adali T, Plis SM (2014) Restricted Boltzmann machines for neuroimaging: an application in identifying intrinsic networks. NeuroImage 96:245–260CrossRefGoogle Scholar
 14.Ji N, Zhang J, Zhang C, Yin Q (2014) Enhancing performance of restricted Boltzmann machines via logsum regularization. Knowl Based Syst 63:82–96CrossRefGoogle Scholar
 15.Larochelle H, Bengio Y, Turian J (2010) Tractable multivariate binary density estimation and the restricted Boltzmann forest. Neural Comput 22(9):2285–2307MathSciNetCrossRefzbMATHGoogle Scholar
 16.LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradientbased learning applied to document recognition. Proc IEEE 86(11):2278–2324CrossRefGoogle Scholar
 17.Leisink MA, Kappen HJ (2000) Learning in higher order Boltzmann machines using linear response. Neural Netw 13(3):329–335CrossRefGoogle Scholar
 18.Marlin BM, Swersky K, Chen B, Freitas ND (2010) Inductive principles for restricted Boltzmann machine learning. In: International conference on artificial intelligence and statistics, pp 509–516Google Scholar
 19.Memisevic R, Hinton GE (2010) Learning to represent spatial transformations with factored higherorder Boltzmann machines. Neural Comput 22(6):1473–1492CrossRefzbMATHGoogle Scholar
 20.Mohamed AR, Dahl GE, Hinton G (2012) Acoustic modeling using deep belief networks. IEEE Trans Audio Speech Lang Process 20(1):14–22CrossRefGoogle Scholar
 21.Nair V, Hinton GE (2008) Implicit mixtures of restricted Boltzmann machines. NIPS 21:1145–1152Google Scholar
 22.Ranzato M, Krizhevsky A, Hinton GE, et al (2010) Factored 3way restricted Boltzmann machines for modeling natural images. In: International conference on artificial intelligence and statistics, pp 621–628Google Scholar
 23.Salakhutdinov R, Tenenbaum JB, Torralba A (2013) Learning with hierarchicaldeep models. IEEE Trans Pattern Anal Mach Intell 35(8):1958–1971CrossRefGoogle Scholar
 24.Sejnowski TJ (1986) Higherorder Boltzmann machines. AIP Conf Proc 151:398–403CrossRefGoogle Scholar
 25.Sohn K, Zhou G, Lee C, Lee H (2013) Learning and selecting features jointly with pointwise gated Boltzmann machines. In: Proceedings of The 30th international conference on machine learning, pp 217–225Google Scholar
 26.Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R (2014) Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res 15(1):1929–1958MathSciNetzbMATHGoogle Scholar
 27.Srivastava N, Salakhutdinov R (2014) Multimodal learning with deep Boltzmann machines. J Mach Learn Res 15:2949–2980MathSciNetzbMATHGoogle Scholar
 28.Swersky K, Ranzato M, Buchman D, Marlin BM, Freitas ND (2011) On autoencoders and score matching for energy based models. In: Proceedings of the 28th international conference on machine learning (ICML11), pp 1201–1208Google Scholar
 29.Tompson J, Stein M, LeCun Y, Perlin K (2014) Realtime continuous pose recovery of human hands using convolutional networks. ACM Trans Gr 33(5):169:1–169:10CrossRefGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.