I Introduction
Deep neural networks (DNNs) have been shown to outperform shallow learning algorithms in applications such as computer vision
[1, 2], automatic speech recognition [3, 4]and natural language processing
[5, 6]; however DNNs also have large parameter sets, often making them prohibitive for smallfootprint devices [7]. For example, while the original LeNet5 network [8](a classification system based on convolutional neural network) has less than 100K parameters, the winner of the 2012 ImageNet competition
[9] has over 60M parameters. The memory access costs alone can make these larger networks unsuitable for lowpower settings.It has been posited that the expressive power of DNNs comes from their large parameter spaces and hierarchical structure; however recent studies have shown that there is often a great deal of parameter redundancy in DNNs [10, 11], making them unnecessarily complex. As a result, reducing the complexity of DNNs has been an area of great interest to the research community in recent years. For example, the authors in [12, 13] used low rank decomposition of the weights to reduce the parameter set and applied this method to a DNNbased acoustic model and to convolutional neural networks (CNN) for image classification. Similarly, the authors in [11] showed that over 95% parameters of DNNs can be predicted without any training and without impacting accuracy.
In addition to lowrank parameter decomposition, network pruning and quantization methods have also been proposed [14]. Neural network pruning has been investigated in early studies, including pruning weights with small magnitudes, optimal brain damage [15] and optimal brain surgeon [16]. The last two methods require estimation of the Hessian matrix of network parameters to decide on their importance; however, the sizes of existing networks make the estimation of this large matrix prohibitive. As a result, for largescale DNNs, magnitudebased weight pruning is still a popular method [17, 14, 18].
For fixedpoint implementations of DNNs, parameter quantization is also required. The studies in [19, 20] discretized the weights of a neural network according to the range of the weights. The methods in [21] and [22] used uniform scalar parameter quantization to implement fixedpoint versions of the networks. In [23]
, a new fixedpoint representation for DNN training was proposed, using stochastic rounding for the parameters. Vector quantization based schemes have been applied to CNNs for both computer vision and automatic speech recognition tasks
[24, 25, 26].In this paper, we propose a new method that ranks the parameters of a DNN for both network pruning and parameter quantization. We investigate an informationtheoretic approach to reduce the DNN parameter space by using the Fisher Information
as a proxy for parameter importance. The Fisher criterion is a natural metric for quantifying the relative importance of DNN parameters since it provides an estimate of how much information a random variable carries about a parameter of the distribution. In
[27], we introduced a new method to calculate the diagonal of the Fisher Information Matrix (FIM) and showed that it can be used to reduce the size of DNNs. In this paper, we extend this work by using a lowercomplexity estimate of the FIM diagonal and evaluating the technique on a much larger network. We validate the method on the MNIST dataset using a CNN and show that our method results in smaller networks with fewer parameters to quantize at a lower bit rate.Ii DNNs pruning and quantization
Iia Fisher Information and DNNs
We show a notional DNN architecture in Fig. 1. Let us consider the output
of the DNN as a conditional probability distribution
parameterized by the DNN input, , and its parameters, . The FIM evaluated at a particular value of is defined as:(1) 
We can see from eqn. (1) that the FIM is the covariance of the gradient of log likelihood with regards to its parameter . Thus, is a symmetric positive semidefinite matrix.
It is easy to see that the diagonal elements of the FIM can be calculated by the expectation of the elementwise multiplication of the gradient:
(2) 
where is the gradient of the loglikelihood and represents elementwise multiplication; is a vector, each element of which is the Fisher Information of a corresponding parameter.
The Fisher Information provides an estimate of the amount of information a random variable carries about a parameter of the distribution. In the context of a DNN, this provides a natural metric for quantifying the relative importance of any given parameter in the network. The less information an output variable carries about a parameter, the less important that parameter is to the output statistics of the network. As a result, we assume that removing parameters with low entries in will not greatly affect the output of the network. That is precisely the approach we take in this paper  we will rank the parameters in a DNN based on their corresponding entries in the FIM diagonal. In the ensuing section, we describe the method we use for approximating .
IiB Estimating the Fisher Information
A number of recent studies on natural gradient descent (NGD) [28] exploit the information geometry of the underlying parameter manifold and apply it to gradientbased optimization of DNNs. Natural gradient descent uses the inverse FIM to constrain the magnitude of the update steps such that the KullbackLeibler (KL) divergence between the output distribution of the network at iteration and iteration is constant [29, 30, 31]. This approach avoids large update steps and results in faster convergence.
For classification problems, DNNs are trained by minimizing the crossentropy loss function:
(3) 
where is the number of training samples, is the number of classes, is the true label of sample and is the indicator function. A recent paper proposed a new stochastic optimization method called “Adam” and showed we can efficiently estimate the FIM diagonal at each iteration while minimizing this loss function [32]. Similar to NGD, Adam uses the approximated FIM diagonal to adapt to the geometry of the data. As a result, in this study, we use Adam to train our DNN classification system. The details of the parameter update scheme for Adam are shown in algorithm 1. As the algorithm shows, after the training algorithm converges, it returns both the optimal and the approximated Fisher Information
. We should note that Adam is not the only choice as the optimizer because standard stochastic gradient descent can also be used; however this would require some other means of estimating
.IiC Network Pruning and Quantization
The simplest approach to network pruning is to rank the parameters by comparing their entries in the FIM diagonal and removing the ones with the lowest entries. However, as we will see in the results section, this method does not work well since estimating small values in the FIM diagonal is challenging and unreliable. When the model is overparameterized, the actual parameter space is much smaller than the number of parameters used in the network. As a result, after training, a number of parameters become close to zero and estimating their influence based on the Fisher Information is challenging [27]. To address this problem we use a combination of magnitudebased and FIMbased pruning. For example, if we want to prune parameters from the network, we first remove parameters with the smallest magnitude. Then, we rank the remaining parameters based on their FIM diagonal entries and remove the additional parameters with the smallest entries in the FIM diagonal. The parameter is between 0 and 1 and can be optimized using crossvalidation.
After network pruning, we want to quantize the remaining parameters with the lowest bit representation possible. After removing parameters, we rank the remaining parameters by comparing their entries in the FIM diagonal and then apply means clustering to separate the parameters into several groups. We quantize groups with higher Fisher Information values using more bits and groups with lower Fisher Information using fewer bits.
Iii Experiments and Results Analysis
In this section, we present the experiments and results in two parts: network pruning and network quantization. All the experiments were done using the Python neural network library Keras
[33]implemented using Theano on an NVIDIA GTX 760 GPU.
We evaluated our algorithms on the MNIST digits data set, which consists of 60K binary images for training and 10K for testing. There are 10 classes (digits from 0 to 9) in the data and the size of each image (and the input dimension of the neural network) is . We trained a convolutional neural network (CNN) with 2 convolutional layers each with 32 filters. The size of the convolutional kernel was
and the rectified linear unit (ReLU) activation function was used. There was a
maxpooling layer following the two convolutional layers with 0.25 dropout probability. Before the output layer, there was a fully connected layer with 128 nodes with ReLU activations and 0.5 dropout probability. The output layer had 10 nodes with softmax activations.The loss function used to train the network was the categorical crossentropy shown in eqn. (3). We used Adam as the optimizer with the following settings 1
: batch size = 256, number of epochs = 50, step size
= 0.001, = 0.9, = 0.999, . The accuracy of this model on the MNIST classification task without any pruning and quantization was 99.29%.Below we describe the performance of the proposed algorithm for both pruning and quantization tasks. While we focus on CNNs in this section, our method is in no way restricted to only CNNs or only classification networks. Indeed, our probabilistic interpretation of the DNN output makes the methodology applicable across all network types, provided that the Fisher Information can be accurately estimated. Since the fully connected layers of CNNs accounts for 90% of the total weights [34], we only focus on the weights (including bias terms) in the fully connected layers in this paper as others have done in [24] [25].
Iiia Network pruning
The trained network consisted of a total of 591,242 parameters in the fully connected layers. We removed different numbers of parameters using three different methods: (1) magnitudebased pruning where parameters with the smallest magnitude were removed; (2) Fisher Information based pruning where parameters with small entries in the FIM diagonal were removed; and (3) a combination of magnitude and Fisher Information based pruning, where we traded off between the two methods using the parameter (see Sec. IIC). The number of pruned parameters ranged from (1.69% of the total parameters) to (99.79% of the total parameters) with a step size . For the third method, we fixed the value to 0.05. On this network we found that values below 0.1 yield good results; however for other networks crossvalidation could be used to identify an appropriate value of .
After removing unimportant parameters in the network, we evaluated the results on the test data and noted both the accuracy of the model and test scores (loss function evaluated on test data) as different numbers of parameters were removed. Since there was no obvious accuracy drop until parameters were removed (69.35% of the total parameters), we only show the accuracy and test score starting with 69.35% parameters removed. The results are shown in Fig. 3 and 3. In Fig. 3, we show the accuracy of the model on the test data as we remove an increased number of parameters; In Fig. 3, we show the same plot, but with the test score instead of the accuracy. In these figures, “mag” represents magnitude based pruning, “fisher” represents Fisher Information based pruning and “mag_fisher” represents a combination of magnitude and Fisher Information based pruning.
As Fig. 3 shows, as additional parameters are removed, the accuracy of the model eventually decreases. We find that using only Fisher Information based pruning, the accuracy of the model decreases quickly. This is because estimating the FIM for small parameter values is difficult as explained in Sec. IIC. This is consistent with our finding in [27]
, where we used the FIM criterion to remove parameters in an autoencoder. Using magnitudebased pruning, there is a clear dropoff in performance after
parameters (93.02% of the total parameters) are removed; however, by using our combination of magnitude and Fisher Information pruning there is no obvious accuracy drop until parameters (96.41% of the total parameters) are removed. The advantage of the combined method shows that about more parameters (3.38% of the total parameters) can be removed compared to magnitude based pruning with minimal impact on model performance. The test score in Fig. 3 follows the same trend as the accuracy plot in Fig. 3.To further highlight the differences in performance between “mag” and “mag_fisher”, we zoom in at the point where the accuracy starts to decrease by running the experiment with a smaller step size. These findings are shown in Fig. 4, where we show the accuracy (99.29%) without any pruning. The starting point of “mag_fisher” is 98.75% while for “mag” it is 98.02%. From this result, we can more clearly see the advantage of our combined method compared to magnitude based pruning. Consistent with findings from our previous work [27], the Fisher Information better captures the importance of larger parameters when compared to magnitudebased pruning.
IiiB Network quantization
After pruning, the remaining parameters in the network must be quantized for fixedpoint implementations. Our quantization method is a nonuniform quantization method based on means clustering. We rankorder the weights by an importance metric (“fisher” or “mag”) and cluster them into clusters. The clusters are then quantized using varying bit depths, from 1 bit/parameter (least important) to bits (most important).
To remove the effects of different pruning methods, we first removed parameters (91.33% of the total parameters) using magnitudebased pruning only. The resulting model had 51,242 (8.67%) parameters remaining and an overall accuracy of 98.67% (less than 1% accuracy loss). We quantized the remaining parameters using three different methods: (1) nonuniform quantization based on Fisher Information ranking; (2) nonuniform quantization based on magnitudebased ranking; and (3) uniform quantization. For methods (1) and (2), we varied the number of clusters (from 3 to 10) and estimated the accuracy of the model for each value of .
The results are shown in Fig. 6 and Fig. 6: “fisher_quant” represents nonuniform quantization based on Fisher Information, “mag_quant” represents nonuniform quantization based on magnitude ranking, “uniform” represents uniform quantization and “Original” represents the original nonquantized model after removing parameters (accuracy is 98.67%).
From the result, we see that nonuniform quantization based on the Fisher Information can achieve almost the same accuracy as the original model with only bits/parameter. This number is 3 for uniform quantization. Nonuniform quantization based on the magnitude never achieves the accuracy of the original model. The same pattern is seen for the test score. This shows that on this classification task, nonuniform quantization based on the Fisher Information achieves the highest compression ratio compared to nonuniform quantization based on magnitude ranking and uniform quantization.
Iv Discussion
To evaluate the compression ratio for the example shown here, we analyze the effects of both network reduction and quantization. As we previously saw in Fig. 3, if we limit our acceptable reduction in performance to 1%, then we can remove 92.18% parameters (accuracy is 98.37%) for magnitude based pruning. This results in a compression ratio of 12.8. For our proposed combination of magnitude and Fisher Information based pruning, we can remove 94.72% parameters (accuracy is 98.38%) and the compression ratio is 18.9.
For network quantization, we can choose between uniform or nonuniform quantization. If we assume the original parameters are saved in FLOAT32 format, as shown in Fig. 6 using uniform quantization, we can achieve a reduction of ; by using nonuniform quantization, we can achieve a reduction of . This means that the total compression ratio can be as high as 18.9 13.3 251.4 with less than 1% accuracy loss. This is likely an overestimate of the overall compression ratio because we need additional space to store the indices of the parameters that have been removed.
There is a relationship between our method and other previous methods based on estimation of the Hessian diagonal, namely optimal brain damage [15], optimal brain surgeon [16] and our previous work [27]. The first two methods use the entries in the Hessian diagonal of the resulting cost function to identify important and unimportant parameters. These approaches are closely related to our approach when the cost function is the loglikelihood since the second derivative of loglikelihood function (Hessian) evaluated at the maximum likelihood estimate is the observed Fisher Information [35]. Our approach in [27] made use of the relationship between Fisher Information and the family of divergences to estimate the FIM diagonal. The principal difference between those approaches and the one we use here is scalability  the stochastic optimization method we use to estimate the FIM diagonal can be scaled to much larger network sizes.
Finally, it is important to note that further gains in performance can be obtained by retraining the network after pruning and before quantization [14]. In this study, in an attempt to isolate the effects of network reduction and quantization, we elected not to retrain the network after the fact.
V Conclusion
In this paper, we propose a new network reduction and quantization scheme that uses a combination of the parameter magnitude and the Fisher Information as a measure of parameter importance. For network reduction, the proposed algorithm first removes parameters with small magnitude and then further reduces the network by removing additional parameters based on the Fisher Information. Following network reduction, we propose a nonuniform quantization scheme for remaining parameters based on the same Fisher criterion. The results show that the combination of network reduction and quantization results in large compression ratios. In future, our aim is to embed complexity reduction criteria in the training process instead of using it as a postprocessing step.
Acknowledgment
This research was supported in part by the Office of Naval Research grant N000141410722 (Berisha), an ASUMayo seed grant, and a hardware grant from NVIDIA.
References
 [1] K. Simonyan and A. Zisserman, “Very deep convolutional networks for largescale image recognition,” CoRR, vol. abs/1409.1556, 2014.

[2]
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich,
“Going deeper with convolutions,”
in
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
, 2015, pp. 1–9.  [3] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdelrahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al., “Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups,” Signal Processing Magazine, IEEE, vol. 29, no. 6, pp. 82–97, 2012.

[4]
Alex Graves and Navdeep Jaitly,
“Towards endtoend speech recognition with recurrent neural networks,”
inProceedings of the 31st International Conference on Machine Learning (ICML14)
, 2014, pp. 1764–1772.  [5] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton, Nature, vol. 521, no. 7553, pp. 436–444, 2015.
 [6] Yoav Goldberg, “A primer on neural network models for natural language processing,” arXiv preprint arXiv:1510.00726, 2015.
 [7] Vinayak Gokhale, Jonghoon Jin, Aysegul Dundar, Berin Martini, and Eugenio Culurciello, “A 240 gops/s mobile coprocessor for deep neural networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2014, pp. 682–687.
 [8] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner, “Gradientbased learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
 [9] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
 [10] Yu Cheng, Felix X Yu, Rogerio S Feris, Sanjiv Kumar, Alok Choudhary, and ShiFu Chang, “An exploration of parameter redundancy in deep networks with circulant projections,” in Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 2857–2865.
 [11] Misha Denil, Babak Shakibi, Laurent Dinh, Nando de Freitas, et al., “Predicting parameters in deep learning,” in Advances in Neural Information Processing Systems, 2013, pp. 2148–2156.

[12]
Jian Xue, Jinyu Li, and Yifan Gong,
“Restructuring of deep neural network acoustic models with singular value decomposition.,”
in INTERSPEECH, 2013, pp. 2365–2369.  [13] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus, “Exploiting linear structure within convolutional networks for efficient evaluation,” in Advances in Neural Information Processing Systems, 2014, pp. 1269–1277.
 [14] Song Han, Jeff Pool, John Tran, and William Dally, “Learning both weights and connections for efficient neural network,” in Advances in Neural Information Processing Systems, 2015, pp. 1135–1143.
 [15] Yann LeCun, John S Denker, and Sara A Solla, “Optimal brain damage,” in Advances in Neural Information Processing Systems, 1990, pp. 598–605.
 [16] Babak Hassibi and David G Stork, “Second order derivatives for network pruning: Optimal brain surgeon,” in Advances in Neural Information Processing Systems, 1993, pp. 164–171.
 [17] Dong Yu, Frank Seide, Gang Li, and Li Deng, “Exploiting sparseness in deep neural networks for large vocabulary speech recognition,” in Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE International Conference on. IEEE, 2012, pp. 4409–4412.
 [18] Song Han, Huizi Mao, and William J Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,” arXiv preprint arXiv:1510.00149, 2015.

[19]
Fatih Köksal, Ethem Alpaydyn, and Günhan Dündar,
“Weight quantization for multilayer perceptrons using soft weight sharing,”
in Artificial Neural Networks (ICANN) 2001, pp. 211–216. Springer, 2001.  [20] Ryu Takeda, Naoyuki Kanda, and Nobuo Nukaga, “Boundary contraction training for acoustic models based on discrete deep neural networks,” in INTERSPEECH, 2014.
 [21] Xin Lei, Andrew Senior, Alexander Gruenstein, and Jeffrey Sorensen, “Accurate and compact large vocabulary speech recognition on mobile devices.,” in INTERSPEECH, 2013, pp. 662–665.
 [22] Vincent Vanhoucke, Andrew Senior, and Mark Z Mao, “Improving the speed of neural networks on cpus,” in Proc. Deep Learning and Unsupervised Feature Learning NIPS Workshop, 2011, vol. 1.
 [23] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan, “Deep learning with limited numerical precision,” arXiv preprint arXiv:1502.02551, 2015.
 [24] Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev, “Compressing deep convolutional networks using vector quantization,” arXiv preprint arXiv:1412.6115, 2014.
 [25] Guillaume Soulié, Vincent Gripon, and Maëlys Robert, “Compression of deep neural networks on the fly,” arXiv preprint arXiv:1509.08745, 2015.
 [26] Yongqiang Wang, Jinyu Li, and Yifan Gong, “Smallfootprint highperformance deep neural networkbased speech recognition using splitvq,” in Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on. IEEE, 2015, pp. 4984–4988.
 [27] Ming Tu, Visar Berisha, Martin Woolf, Jaesun Seo, and Yu Cao, “Ranking the parameters of deep neural network using the fisher information,” in Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on. IEEE, under publication, 2016.
 [28] ShunIchi Amari, “Natural gradient works efficiently in learning,” Neural computation, vol. 10, no. 2, pp. 251–276, 1998.
 [29] Razvan Pascanu and Yoshua Bengio, “Revisiting natural gradient for deep networks,” arXiv preprint arXiv:1301.3584, 2013.
 [30] Daniel Povey, Xiaohui Zhang, and Sanjeev Khudanpur, “Parallel training of dnns with natural gradient and parameter averaging,” arXiv preprint arXiv:1410.7455, 2014.
 [31] Guillaume Desjardins, Karen Simonyan, Razvan Pascanu, et al., “Natural neural networks,” in Advances in Neural Information Processing Systems, 2015, pp. 2062–2070.
 [32] Diederik Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.
 [33] Fran ois Chollet, “Keras,” https://github.com/fchollet/keras, 2015.
 [34] Matthew D Zeiler and Rob Fergus, “Visualizing and understanding convolutional networks,” in Computer vision–ECCV 2014, pp. 818–833. Springer, 2014.
 [35] Bradley Efron and David V Hinkley, “Assessing the accuracy of the maximum likelihood estimator: Observed versus expected fisher information,” Biometrika, vol. 65, no. 3, pp. 457–483, 1978.
Comments
There are no comments yet.