Interpreting Convolutional Neural Networks Through Compression

Convolutional neural networks (CNNs) achieve state-of-the-art performance in a wide variety of tasks in computer vision. However, interpreting CNNs still remains a challenge. This is mainly due to the large number of parameters in these networks. Here, we investigate the role of compression and particularly pruning filters in the interpretation of CNNs. We exploit our recently-proposed greedy structural compression scheme that prunes filters in a trained CNN. In our compression, the filter importance index is defined as the classification accuracy reduction (CAR) of the network after pruning that filter. The filters are then iteratively pruned based on the CAR index. We demonstrate the interpretability of CAR-compressed CNNs by showing that our algorithm prunes filters with visually redundant pattern selectivity. Specifically, we show the importance of shape-selective filters for object recognition, as opposed to color-selective filters. Out of top 20 CAR-pruned filters in AlexNet, 17 of them in the first layer and 14 of them in the second layer are color-selective filters. Finally, we introduce a variant of our CAR importance index that quantifies the importance of each image class to each CNN filter. We show that the most and the least important class labels present a meaningful interpretation of each filter that is consistent with the visualized pattern selectivity of that filter.

Structural Compression of Convolutional Neural Networks Based on Greedy Filter Pruning

Convolutional neural networks (CNNs) have state-of-the-art performance on many problems in machine vision. However, networks with superior performance often have millions of weights so that it is difficult or impossible to use CNNs on computationally limited devices or to humanly interpret them. A myriad of CNN compression approaches have been proposed and they involve pruning and compressing the weights and filters. In this article, we introduce a greedy structural compression scheme that prunes filters in a trained CNN. We define a filter importance index equal to the classification accuracy reduction (CAR) of the network after pruning that filter (similarly defined as RAR for regression). We then iteratively prune filters based on the CAR index. This algorithm achieves substantially higher classification accuracy in AlexNet compared to other structural compression schemes that prune filters. Pruning half of the filters in the first or second layer of AlexNet, our CAR algorithm achieves 26% and 20% higher classification accuracies respectively, compared to the best benchmark filter pruning scheme. Our CAR algorithm, combined with further weight pruning and compressing, reduces the size of first or second convolutional layer in AlexNet by a factor of 42, while achieving close to original classification accuracy through retraining (or fine-tuning) network. Finally, we demonstrate the interpretability of CAR-compressed CNNs by showing that our algorithm prunes filters with visually redundant functionalities. In fact, out of top 20 CAR-pruned filters in AlexNet, 17 of them in the first layer and 14 of them in the second layer are color-selective filters as opposed to shape-selective filters. To our knowledge, this is the first reported result on the connection between compression and interpretability of CNNs.

Formulas for Counting the Sizes of Markov Equivalence Classes of Directed Acyclic Graphs

The sizes of Markov equivalence classes of directed acyclic graphs play important roles in measuring the uncertainty and complexity in causal learning. A Markov equivalence class can be represented by an essential graph and its undirected subgraphs determine the size of the class. In this paper, we develop a method to derive the formulas for counting the sizes of Markov equivalence classes. We first introduce a new concept of core graph. The size of a Markov equivalence class of interest is a polynomial of the number of vertices given its core graph. Then, we discuss the recursive and explicit formula of the polynomial, and provide an algorithm to derive the size formula via symbolic computation for any given core graph. The proposed size formula derivation sheds light on the relationships between the size of a Markov equivalence class and its representation graph, and makes size counting efficient, even when the essential graphs contain non-sparse undirected subgraphs.

Local identifiability of $l_1$-minimization dictionary learning: a sufficient and almost necessary condition

We study the theoretical properties of learning a dictionary from $N$ signals $\mathbf x_i\in \mathbb R^K$ for $i=1,...,N$ via $l_1$-minimization. We assume that $\mathbf x_i$'s are $i.i.d.$ random linear combinations of the $K$ columns from a complete (i.e., square and invertible) reference dictionary $\mathbf D_0 \in \mathbb R^{K\times K}$. Here, the random linear coefficients are generated from either the $s$-sparse Gaussian model or the Bernoulli-Gaussian model. First, for the population case, we establish a sufficient and almost necessary condition for the reference dictionary $\mathbf D_0$ to be locally identifiable, i.e., a local minimum of the expected $l_1$-norm objective function. Our condition covers both sparse and dense cases of the random linear coefficients and significantly improves the sufficient condition by Gribonval and Schnass (2010). In addition, we show that for a complete $\mu$-coherent reference dictionary, i.e., a dictionary with absolute pairwise column inner-product at most $\mu\in[0,1)$, local identifiability holds even when the random linear coefficient vector has up to $O(\mu^{-2})$ nonzeros on average. Moreover, our local identifiability results also translate to the finite sample case with high probability provided that the number of signals $N$ scales as $O(K\log K)$.

Optimal Subsampling Approaches for Large Sample Linear Regression

A significant hurdle for analyzing large sample data is the lack of effective statistical computing and inference methods. An emerging powerful approach for analyzing large sample data is subsampling, by which one takes a random subsample from the original full sample and uses it as a surrogate for subsequent computation and estimation. In this paper, we study subsampling methods under two scenarios: approximating the full sample ordinary least-square (OLS) estimator and estimating the coefficients in linear regression. We present two algorithms, weighted estimation algorithm and unweighted estimation algorithm, and analyze asymptotic behaviors of their resulting subsample estimators under general conditions. For the weighted estimation algorithm, we propose a criterion for selecting the optimal sampling probability by making use of the asymptotic results. On the basis of the criterion, we provide two novel subsampling methods, the optimal subsampling and the predictor- length subsampling methods. The predictor-length subsampling method is based on the L2 norm of predictors rather than leverage scores. Its computational cost is scalable. For unweighted estimation algorithm, we show that its resulting subsample estimator is not consistent to the full sample OLS estimator. However, it has better performance than the weighted estimation algorithm for estimating the coefficients. Simulation studies and a real data example are used to demonstrate the effectiveness of our proposed subsampling methods.

Estimation Stability with Cross Validation (ESCV)

Cross-validation (CV) is often used to select the regularization parameter in high dimensional problems. However, when applied to the sparse modeling method Lasso, CV leads to models that are unstable in high-dimensions, and consequently not suited for reliable interpretation. In this paper, we propose a model-free criterion ESCV based on a new estimation stability (ES) metric and CV. Our proposed ESCV finds a locally ES-optimal model smaller than the CV choice so that the it fits the data and also enjoys estimation stability property. We demonstrate that ESCV is an effective alternative to CV at a similar easily parallelizable computational cost. In particular, we compare the two approaches with respect to several performance measures when applied to the Lasso on both simulated and real data sets. For dependent predictors common in practice, our main finding is that, ESCV cuts down false positive rates often by a large margin, while sacrificing little of true positive rates. ESCV usually outperforms CV in terms of parameter estimation while giving similar performance as CV in terms of prediction. For the two real data sets from neuroscience and cell biology, the models found by ESCV are less than half of the model sizes by CV. Judged based on subject knowledge, they are more plausible than those by CV as well. We also discuss some regularization parameter alignment issues that come up in both approaches.

The geometry of kernelized spectral clustering

Clustering of data sets is a standard problem in many areas of science and engineering. The method of spectral clustering is based on embedding the data set using a kernel function, and using the top eigenvectors of the normalized Laplacian to recover the connected components. We study the performance of spectral clustering in recovering the latent labels of i.i.d. samples from a finite mixture of nonparametric distributions. The difficulty of this label recovery problem depends on the overlap between mixture components and how easily a mixture component is divided into two nonoverlapping components. When the overlap is small compared to the indivisibility of the mixture components, the principal eigenspace of the population-level normalized Laplacian operator is approximately spanned by the square-root kernelized component densities. In the finite sample setting, and under the same assumption, embedded samples from different components are approximately orthogonal with high probability when the sample size is large. As a corollary we control the fraction of samples mislabeled by spectral clustering under finite mixtures with nonparametric components.

Co-clustering for directed graphs: the Stochastic co-Blockmodel and spectral algorithm Di-Sim

Directed graphs have asymmetric connections, yet the current graph clustering methodologies cannot identify the potentially global structure of these asymmetries. We give a spectral algorithm called di-sim that builds on a dual measure of similarity that correspond to how a node (i) sends and (ii) receives edges. Using di-sim, we analyze the global asymmetries in the networks of Enron emails, political blogs, and the c elegans neural connectome. In each example, a small subset of nodes have persistent asymmetries; these nodes send edges with one cluster, but receive edges with another cluster. Previous approaches would have assigned these asymmetric nodes to only one cluster, failing to identify their sending/receiving asymmetries. Regularization and "projection" are two steps of di-sim that are essential for spectral clustering algorithms to work in practice. The theoretical results show that these steps make the algorithm weakly consistent under the degree corrected Stochastic co-Blockmodel, a model that generalizes the Stochastic Blockmodel to allow for both (i) degree heterogeneity and (ii) the global asymmetries that we intend to detect. The theoretical results make no assumptions on the smallest degree nodes. Instead, the theorem requires that the average degree grows sufficiently fast and that the weak consistency only applies to the subset of the nodes with sufficiently large leverage scores. The results results also apply to bipartite graphs.

Error Rate Bounds and Iterative Weighted Majority Voting for Crowdsourcing

Crowdsourcing has become an effective and popular tool for human-powered computation to label large datasets. Since the workers can be unreliable, it is common in crowdsourcing to assign multiple workers to one task, and to aggregate the labels in order to obtain results of high quality. In this paper, we provide finite-sample exponential bounds on the error rate (in probability and in expectation) of general aggregation rules under the Dawid-Skene crowdsourcing model. The bounds are derived for multi-class labeling, and can be used to analyze many aggregation methods, including majority voting, weighted majority voting and the oracle Maximum A Posteriori (MAP) rule. We show that the oracle MAP rule approximately optimizes our upper bound on the mean error rate of weighted majority voting in certain setting. We propose an iterative weighted majority voting (IWMV) method that optimizes the error rate bound and approximates the oracle MAP rule. Its one step version has a provable theoretical guarantee on the error rate. The IWMV method is intuitive and computationally simple. Experimental results on simulated and real data show that IWMV performs at least on par with the state-of-the-art methods, and it has a much lower computational cost (around one hundred times faster) than the state-of-the-art methods.

Statistical guarantees for the EM algorithm: From population to sample-based analysis

We develop a general framework for proving rigorous guarantees on the performance of the EM algorithm and a variant known as gradient EM. Our analysis is divided into two parts: a treatment of these algorithms at the population level (in the limit of infinite data), followed by results that apply to updates based on a finite set of samples. First, we characterize the domain of attraction of any global maximizer of the population likelihood. This characterization is based on a novel view of the EM updates as a perturbed form of likelihood ascent, or in parallel, of the gradient EM updates as a perturbed form of standard gradient ascent. Leveraging this characterization, we then provide non-asymptotic guarantees on the EM and gradient EM algorithms when applied to a finite set of samples. We develop consequences of our general theory for three canonical examples of incomplete-data problems: mixture of Gaussians, mixture of regressions, and linear regression with covariates missing completely at random. In each case, our theory guarantees that with a suitable initialization, a relatively small number of EM (or gradient EM) steps will yield (with high probability) an estimate that is within statistical error of the MLE. We provide simulations to confirm this theoretically predicted behavior.