skip to main content


Title: Structured Sparsity of Convolutional Neural Networks via Nonconvex Sparse Group Regularization
Convolutional neural networks (CNN) have been hugely successful recently with superior accuracy and performance in various imaging applications, such as classification, object detection, and segmentation. However, a highly accurate CNN model requires millions of parameters to be trained and utilized. Even to increase its performance slightly would require significantly more parameters due to adding more layers and/or increasing the number of filters per layer. Apparently, many of these weight parameters turn out to be redundant and extraneous, so the original, dense model can be replaced by its compressed version attained by imposing inter- and intra-group sparsity onto the layer weights during training. In this paper, we propose a nonconvex family of sparse group lasso that blends nonconvex regularization (e.g., transformed L1, L1 - L2, and L0) that induces sparsity onto the individual weights and L2,1 regularization onto the output channels of a layer. We apply variable splitting onto the proposed regularization to develop an algorithm that consists of two steps per iteration: gradient descent and thresholding. Numerical experiments are demonstrated on various CNN architectures showcasing the effectiveness of the nonconvex family of sparse group lasso in network sparsification and test accuracy on par with the current state of the art.  more » « less
Award ID(s):
1632935
NSF-PAR ID:
10206431
Author(s) / Creator(s):
Date Published:
Journal Name:
Frontiers in applied mathematics and statistics
ISSN:
2297-4687
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Tabacu, Lucia (Ed.)
    Convolutional neural networks (CNN) have been hugely successful recently with superior accuracy and performance in various imaging applications, such as classification, object detection, and segmentation. However, a highly accurate CNN model requires millions of parameters to be trained and utilized. Even to increase its performance slightly would require significantly more parameters due to adding more layers and/or increasing the number of filters per layer. Apparently, many of these weight parameters turn out to be redundant and extraneous, so the original, dense model can be replaced by its compressed version attained by imposing inter- and intra-group sparsity onto the layer weights during training. In this paper, we propose a nonconvex family of sparse group lasso that blends nonconvex regularization (e.g., transformed ℓ1, ℓ1 − ℓ2, and ℓ0) that induces sparsity onto the individual weights and ℓ2,1 regularization onto the output channels of a layer. We apply variable splitting onto the proposed regularization to develop an algorithm that consists of two steps per iteration: gradient descent and thresholding. Numerical experiments are demonstrated on various CNN architectures showcasing the effectiveness of the nonconvex family of sparse group lasso in network sparsification and test accuracy on par with the current state of the art. 
    more » « less
  2. Objectively differentiating patient mental states based on electrical activity, as opposed to overt behavior, is a fundamental neuroscience problem with medical applications, such as identifying patients in locked-in state vs. coma. Electroencephalography (EEG), which detects millisecond-level changes in brain activity across a range of frequencies, allows for assessment of external stimulus processing by the brain in a non-invasive manner. We applied machine learning methods to 26-channel EEG data of 24 fluent Deaf signers watching videos of sign language sentences (comprehension condition), and the same videos reversed in time (non-comprehension condition), to objectively separate vision-based high-level cognition states. While spectrotemporal parameters of the stimuli were identical in comprehension vs. non-comprehension conditions, the neural responses of participants varied based on their ability to linguistically decode visual data. We aimed to determine which subset of parameters (specific scalp regions or frequency ranges) would be necessary and sufficient for high classification accuracy of comprehension state. Optical flow, characterizing distribution of velocities of objects in an image, was calculated for each pixel of stimulus videos using MATLAB Vision toolbox. Coherence between optical flow in the stimulus and EEG neural response (per video, per participant) was then computed using canonical component analysis with NoiseTools toolbox. Peak correlations were extracted for each frequency for each electrode, participant, and video. A set of standard ML algorithms were applied to the entire dataset (26 channels, frequencies from .2 Hz to 12.4 Hz, binned in 1 Hz increments), with consistent out-of-sample 100% accuracy for frequencies in .2-1 Hz range for all regions, and above 80% accuracy for frequencies < 4 Hz. Sparse Optimal Scoring (SOS) was then applied to the EEG data to reduce the dimensionality of the features and improve model interpretability. SOS with elastic-net penalty resulted in out-of-sample classification accuracy of 98.89%. The sparsity pattern in the model indicated that frequencies between 0.2–4 Hz were primarily used in the classification, suggesting that underlying data may be group sparse. Further, SOS with group lasso penalty was applied to regional subsets of electrodes (anterior, posterior, left, right). All trials achieved greater than 97% out-of-sample classification accuracy. The sparsity patterns from the trials using 1 Hz bins over individual regions consistently indicated frequencies between 0.2–1 Hz were primarily used in the classification, with anterior and left regions performing the best with 98.89% and 99.17% classification accuracy, respectively. While the sparsity pattern may not be the unique optimal model for a given trial, the high classification accuracy indicates that these models have accurately identified common neural responses to visual linguistic stimuli. Cortical tracking of spectro-temporal change in the visual signal of sign language appears to rely on lower frequencies proportional to the N400/P600 time-domain evoked response potentials, indicating that visual language comprehension is grounded in predictive processing mechanisms. 
    more » « less
  3. Sparsification of neural networks is one of the effective complexity reduction methods to improve efficiency and generalizability. Binarized activation offers an additional computational saving for inference. Due to vanishing gradient issue in training networks with binarized activation, coarse gradient (a.k.a. straight through estimator) is adopted in practice. In this paper, we study the problem of coarse gradient descent (CGD) learning of a one hidden layer convolutional neural network (CNN) with binarized activation function and sparse weights. It is known that when the input data is Gaussian distributed, no-overlap one hidden layer CNN with ReLU activation and general weight can be learned by GD in polynomial time at high probability in regression problems with ground truth. We propose a relaxed variable splitting method integrating thresholding and coarse gradient descent. The sparsity in network weight is realized through thresholding during the CGD training process. We prove that under thresholding of L1, L0, and transformed-L1 penalties, no-overlap binary activation CNN can be learned with high probability, and the iterative weights converge to a global limit which is a transformation of the true weight under a novel sparsifying operation. We found explicit error estimates of sparse weights from the true weights. 
    more » « less
  4. Model compression is an important technique to facilitate efficient embedded and hardware implementations of deep neural networks (DNNs), a number of prior works are dedicated to model compression techniques. The target is to simultaneously reduce the model storage size and accelerate the computation, with minor effect on accuracy. Two important categories of DNN model compression techniques are weight pruning and weight quantization. The former leverages the redundancy in the number of weights, whereas the latter leverages the redundancy in bit representation of weights. These two sources of redundancy can be combined, thereby leading to a higher degree of DNN model compression. However, a systematic framework of joint weight pruning and quantization of DNNs is lacking, thereby limiting the available model compression ratio. Moreover, the computation reduction, energy efficiency improvement, and hardware performance overhead need to be accounted besides simply model size reduction, and the hardware performance overhead resulted from weight pruning method needs to be taken into consideration. To address these limitations, we present ADMM-NN, the first algorithm-hardware co-optimization framework of DNNs using Alternating Direction Method of Multipliers (ADMM), a powerful technique to solve non-convex optimization problems with possibly combinatorial constraints. The first part of ADMM-NN is a systematic, joint framework of DNN weight pruning and quantization using ADMM. It can be understood as a smart regularization technique with regularization target dynamically updated in each ADMM iteration, thereby resulting in higher performance in model compression than the state-of-the-art. The second part is hardware-aware DNN optimizations to facilitate hardware-level implementations. We perform ADMM-based weight pruning and quantization considering (i) the computation reduction and energy efficiency improvement, and (ii) the hardware performance overhead due to irregular sparsity. The first requirement prioritizes the convolutional layer compression over fully-connected layers, while the latter requires a concept of the break-even pruning ratio, defined as the minimum pruning ratio of a specific layer that results in no hardware performance degradation. Without accuracy loss, ADMM-NN achieves 85× and 24× pruning on LeNet-5 and AlexNet models, respectively, --- significantly higher than the state-of-the-art. The improvements become more significant when focusing on computation reduction. Combining weight pruning and quantization, we achieve 1,910× and 231× reductions in overall model size on these two benchmarks, when focusing on data storage. Highly promising results are also observed on other representative DNNs such as VGGNet and ResNet-50. We release codes and models at https://github.com/yeshaokai/admm-nn. 
    more » « less
  5. Edge machine learning can deliver low-latency and private artificial intelligent (AI) services for mobile devices by leveraging computation and storage resources at the network edge. This paper presents an energy-efficient edge processing framework to execute deep learning inference tasks at the edge computing nodes whose wireless connections to mobile devices are prone to channel uncertainties. Aimed at minimizing the sum of computation and transmission power consumption with probabilistic quality-of-service (QoS) constraints, we formulate a joint inference tasking and downlink beamforming problem that is characterized by a group sparse objective function. We provide a statistical learning based robust optimization approach to approximate the highly intractable probabilistic-QoS constraints by nonconvex quadratic constraints, which are further reformulated as matrix inequalities with a rank-one constraint via matrix lifting. We design a reweighted power minimization approach by iteratively reweighted ℓ1 minimization with difference-of-convex-functions (DC) regularization and updating weights, where the reweighted approach is adopted for enhancing group sparsity whereas the DC regularization is designed for inducing rank-one solutions. Numerical results demonstrate that the proposed approach outperforms other state-of-the-art approaches. 
    more » « less