I Introduction
Restricted Boltzmann Machines (RBMs) and Deep Belief Networks (DBNs) (Fig. 1) are stochastic graphical models that are commonly used in both discriminative and generative applications such as image classification, sequence completion, and speech recognition. An RBM is a symmetrically connected, bipartite Markov Random Field (MRF) which is composed of neuronlike units– these partitions are called the visible layer and the hidden layer . In an RBM, both inference and learning use a Markov Chain Monte Carlo (MCMC) procedure called Gibbs sampling [1]
. To implement Gibbs sampling using binary RBMs each neuron is sampled based on a sigmoidal activation function which is a function of the total input from other connected neurons. A DBN can be thought of as multiple RBMs ‘stacked’ onto each other, forming an
partitioned Markov Random Field with one visible layer , and multiple hidden layers with .RBMs and DBNs are commonly implemented on highperformance platforms such as CPUs and GPUs deployed on cloud computing infrastructures. However recent progress in the field of neuromorphic VLSI using both digital and analog circuit elements ([2], [3]) has demonstrated the potential of performing inference on RBMs/DBNs implemented on such substrates for ultra lowpower, realtime applications. In such implementations, which consist of massively parallel arrays of spiking neurons, classification accuracy is an experimentally quantifiable metric for discriminative RBMs/DBNs as one can experimentally compute classification precision and accuracy using standard labeled datasets such as MNIST. On the other hand, implementing metrics for generative applications on such hardware remains a difficult task because of the highdimensional nature of realworld data.^{1}^{1}1
Essentially one is attempting to verify the correctness of an RBMbased probability distribution model for describing a highdimensional data vector.
In this paper, we demonstrate the applicability of using a highly scaleable nonparametric framework to quantify the performance of RBMs and DBNs for generative tasks on realworld highdimensional data sets based on goodnessoffit testing.Ii Generative Inference and Performance Metrics
An RBM represents the probabilistic model of a data set using the Boltzmann distribution shown below:
(1) 
is often called the energy function and depends on the state of both the visible and hidden units of the RBM. is a matrix of weights between and , and , are the biases of the visible units and hidden units respectively. For binary RBMs, the Gibbs sampling procedure guarantees that a stationary Boltzmann distribution is achieved if the state of each RBM neuron is sampled according to the sigmoidal probability rule given below [1]:
(2) 
Here , the th element of the matrix , is the weight between neurons and , and denotes the bias of neuron .
We would like to compare the Boltzmann distribution generated by the (neuromorphic hardware) Gibbs sampler implemented on the neuromorphic substrate with that of the distribution generated by the benchmark ideal Gibbs sampler (in our case, implemented in software). This calls for a measure of dissimilarity between two distributions together with a way to estimate this dissimilarity based on samples from these distributions. The KullbackLeibler (KL) divergence
[4] is a common choice. However it is difficult to compute because doing so for RBMs requires the computation of the normalizing constant in the denominator of (1), known as the partition function, whose explicit calculation is prohibitive for high dimensional real world data.Annealed Importance sampling (AIS) [5] is another algorithm that has been used to quantify performance of RBMs/DBNs during generative tasks. Unlike KLdivergence, AIS does not suffer from computational complexity limitations. However, AIS only takes into consideration the parameters of the RBM in (1) trained with software and not the parameters of the hardware Gibbs sampler per se.
A practical framework is to perform offline training of the RBM/DBN and then perform inference using Gibbs sampling by utilizing the dynamics of the underlying hardware sampler implemented on the substrate. In this paper, we propose using goodnessoffit testing to compare the probability distributions generated by the hardware and benchmark software Gibbs samplers. To illustrate the methodology, we picked the Crossmatch test [6] among the class of graphbased tests nicely reviewed in [7].
Iii The Crossmatch test
The Crossmatch test [6] is an example of nonparametric goodnessoffit test in arbitrary dimensions based on a graph construction. Given two set of samples, and , respectively from two probability distributions and on some space equipped with a dissimilarity measure . (The test below generalizes to the case of samples of different sizes.) Let if and if . The Crossmatch procedure starts by computing a optimal nonbipartite matching of the ’s, meaning, it finds a permutation of that minimizes
We note that other matchings (such as greedy matching) can be used [8]. The procedure then computes the crossmatch statistic, denoted which is defined as the number of matched pairs including an and a . The test rejects for small values of . The null distribution of is known in closed form [6]:
Having computed , the resulting pvalue is therefore equal to , where
In our case we are dealing with distributions on , where is the dimension of the visible layer. The most common measure of dissimilarity is the Hamming distance, meaning
This is the dissimilarity measure that we use.
Our process for comparing the output from two Gibbs samplers is as follows. We generate samples from each Gibbs sampler and compute the pvalue from the Crossmatch test based on the Hamming distance. We repeat the process many times, resulting in that many pvalues. If the Gibbs samplers were to generate the same distribution, then the pvalues would be approximately uniform in . We can then look at statistics of these pvalues. We chose to look at the mean pvalue, which we use below as a similarity measure between two Gibbs sampler distributions.

A pvalue near 0 indicates that the distributions are far enough for the Crossmatch test to notice with confidence that samples generated by the two distributions do not come from the same distribution.

A pvalue near 0.5 indicates that the distributions are close enough that the Crossmatch test would not notice a (statistically) significant difference between how the two samples are distributed.
Iv Digital Sampler Parameter Selection
TrueNorth is a neuromorphic processor composed of digital integrateandfire (I&F) neurons with the capability to implement both stochastic and deterministic leak and threshold values [3]. The following algorithm for realizing the sigmoidal sampling rule (2) using these dynamical properties to perform MCMC sampling in RBMs was proposed in [9] and is as follows:
The membrane potential of an I&F neuron which implements an RBM neuron is denoted by . sampling in this algorithm is dependent on four parameters: , the number of time steps used for sampling; , the deterministic threshold; , the number of bits allocated for the stochastic threshold variation; and the value of the leak. In order to obtain a useful dynamic range, a multiplicative factor is applied to the weights and biases obtained from the offline RBM training [9]. After integration, the sampled value of an RBM neuron is set to 1 if the corresponding I&F neuron spikes in any of the allowed number of sampling intervals .
Index  (Tw, Vt, TM, leak)  scale 

G1  (1, 130, 8, 0)  50 
G2  (1, 80, 8, 102)  50 
G3  (2, 0, 8, 100)  50 
G4  (8, 79, 9, 49)  50 
G5  (16, 50, 9, 15)  30 
G6  (16, 100, 10, 30)  50 
G7  (16, 633, 8, 90)  100 
In this paper we select 7 different sets of parameters for the digital sampler as shown in Table I. The procedure for testing the 7 samplers was as follows. First, we trained an RBM in software with 784 visible units and 500 hidden units on the MNIST dataset consisting of grayscale images of 5000 handwritten digits (the training data) to generate the weight matrix as well as the bias values. Following this, we initialized the visible units of the RBM with an initial set of values each of which corresponded to one of the 1000 MNIST digits in the set of test data. We generated samples using the benchmark ideal Gibbs sampler in software which we refer to as the ideal data (see Fig. 2a). Following this, we then use the dynamics of the 7 digital samplers to perform MCMC sampling and thereby create seven sets of samples. Figure 2 illustrates the MNIST digits obtained by sampling with the ideal Gibbs sampler as well as with 2 hardware samplers from Table I (G2 and G5).
For comparing each digital sampler against the ideal benchmark, 5,000 trials of Crossmatch were run, resulting in 5,000 pvalues (for each digital sampler). The distribution of the pvalues from the simulations are shown in Figure 3. From this figure and the description of each digital sampler in Table I, it can be seen that the sampler latency and are both critical to the generative performance of the RBM. Figure 4 shows a histogram of the distribution of pvalues for digital samplers G2 and G5 (see Table I). To choose the appropriate sampler out of the samplers from Table I, we took the ratio between the mean pvalue for each sampler, and its estimated energy consumption. The ratio of pvalue to energy, the Energy Performance Efficiency (EPEff), is shown in Figure 5. The figure clearly indicates that out of all samplers from Table I sampler G2 has the highest EPEff and is thus the appropriate sampler to choose.
V Network Resource Optimization
In order to implement the sampling algorithm described in Section IV, with arbitrary values, each RBM neuron, which has accumulated its input from other connected neurons, utilizes two TrueNorth neurons. Specifically, one I&F neuron with a stochastic threshold receiving inputs (data neuron) and a second neuron to produce the stochastic leak of 0 or 1 which is multiplied by the weight factor on the connected axon as per the value used in the algorithm (leak neuron). Figure 6 shows this implementation where the green neurons are the data neurons and the red neurons are the leak neurons. Since a single TrueNorth core has only 256 neurons, a direct implementation of this mapping scheme will result in only 50 percent of the neurons on a crossbar utilized for sampling which can result in a large number of cores for RBM/DBN implementations.
One way to improve core utilization efficiency is to limit the number of leak neurons by relaying a single leak value to multiple data neurons. Such a modification would however reduce the quality of the generated samples since the original algorithm proposed in [9] requires the used for data neurons to be identically and independently distributed (i.i.d.). In such a scenario, Crossmatch can be used to study the loss of generative performance and determine an optimal density of data neurons to leak neurons which is henceforth referred to as the leak density ().
We chose sampler G2 to test the loss of performance and the optimal value of . Using this sampler, and 7 different values of leak density , we studied the variation of generative performance during sampling from the MNIST RBM implemented on TrueNorth. The results of the experiment are shown in Figure 7 below.
Vi Energyperformance tradeoff for Network Resource Optimization
We studied the variation of the EPEff (defined in Section IV) with respect to for the generative RBM model of MNIST implemented on TrueNorth. Here the energy estimated value consumed by TrueNorth during sampling. The results are shown in Figure 8. For low values, lower EPEff values were produced on account of the large number of cores required for sampling. For large values, low EPEff values were observed on account of the significant loss in accuracy as reflected by the Crossmatch pvalues. It is clear from Figure 8 that maximizes the EPEff, and is thus the optimal choice for implementing sampling from the MNIST RBM on TrueNorth.
Vii Applicability to design in analog neuromorphic systems
Leaky integrateandfire neurons which can be implemented in analog neuromorphic systems also have the capability to perform Gibbs sampling from the generative model of an RBM using their own neuron dynamics [10]. The behavior of such neurons are governed by the following equation:
(3) 
where C is the membrane capacitance, is the reset potential, is the neuron’s membrane potential, is the leak conductance, is the neuron’s synaptic current, is the neuron’s threshold, and is a noise term in which
is the noise variance. Similar to the discussion in Section IV, Crossmatch can be used to determine optimal design parameters for the Gibbs sampler on the analog neuromorphic substrate. Additionally in such a system each neuron has its own i.i.d. noise term
as shown in (3). Such continuoustime noise injection utilizes significant system bandwidth hence it is highly desirable to reduce this by connecting the same noise source to multiple neurons. Therefore, by applying the decisiondirected strategy presented in Sections V and VI, the same noise source could be reutilized among multiple neurons, improving system throughput, latency, power and hardware usage.Viii Conclusions
We have demonstrated how a nonparametric goodnessoffit test like the Crossmatch can be used to evaluate the performance of a generative RBM model implemented on neuromorphic VLSI substrates. Such a test, and the pvalues it provides when applied to many realized samples, can provide a useful tool for quantifying the accuracy of a Gibbs sampler, whether digital or analog. In particular, for the problem at hand, such a tool can facilitate the choice of hardware parameters and optimization of network resources.
Ix Acknowledgments
The authors would like to thank the team members of the BrainInspired Computing group at IBM Almaden for supporting this project. The authors would also like to thank the Calit2/QI Pattern Recognition Laboratory at UCSD for providing funding support for attending the conference as well as Bruno Pedroni at UCSD.
References
 [1] S. Haykin, Neural Networks and Learning Machines (3rd Edition). Prentice Hall, 2008.
 [2] G. Indiveri, B. LinaresBarranco, T. J. Hamilton, A. Van Schaik, R. EtienneCummings, T. Delbruck, S.C. Liu, P. Dudek, P. Häfliger, S. Renaud et al., “Neuromorphic silicon neuron circuits,” Frontiers in neuroscience, vol. 5, 2011.
 [3] P. A. Merolla, J. V. Arthur, R. AlvarezIcaza, A. S. Cassidy, J. Sawada, F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y. Nakamura et al., “A million spikingneuron integrated circuit with a scalable communication network and interface,” Science, vol. 345, no. 6197, pp. 668–673, 2014.
 [4] T. M. Cover and J. A. Thomas, Elements of information theory. John Wiley & Sons, 2012.

[5]
R. Salakhutdinov and I. Murray, “On the quantitative analysis of deep belief
networks,” in
Proceedings of the 25th international conference on Machine learning
. ACM, 2008, pp. 872–879.  [6] P. R. Rosenbaum, “An exact distributionfree test comparing two multivariate distributions based on adjacency,” Journal of the Royal Statistical Society: Series B (Statistical Methodology), vol. 67, no. 4, pp. 515–530, 2005.
 [7] B. B. Bhattacharya, “Power of graphbased twosample tests,” arXiv preprint arXiv:1508.07530, 2015.
 [8] E. AriasCastro and B. Pelletier, “On the consistency of the crossmatch test,” arXiv preprint arXiv:1509.05790, 2015.
 [9] S. Das, B. Pedroni, P. Merolla, J. Arthur, A. Cassidy, D. Modha, G. Cauwenberghs, and K. KreutzDelgado, ““Gibbs Sampling with LowPower Spiking Digital Neurons”,” in IEEE Int. Symp. Circuits and Systems, 2015.

[10]
E. Neftci, S. Das, B. Pedroni, K. KreutzDelgado, and G. Cauwenberghs, “Eventdriven contrastive divergence for spiking neuromorphic systems,”
Frontiers in Neuroscience, vol. 7, p. 272, 2013.
Comments
There are no comments yet.