Convolutional neural networks (CNNs) require both intensive computation and frequent memory access, which lead to a low processing speed and large power dissipation. Although the characteristics of the different layers in a CNN are frequently quite different, previous hardware designs have employed common optimization schemes for them. This paper proposes a layer-specific design that employs different organizations that are optimized for the different layers. The proposed design employs two layer-specific optimizations: layer-specific mixed data flow and layer-specific mixed precision. The mixed data flow aims to minimize the off-chip access while demanding a minimal on-chip memory (BRAM) resource of an FPGA device. The mixed precision quantization is to achieve both a lossless accuracy and an aggressive model compression, thereby further reducing the off-chip access. A Bayesian optimization approach is used to select the best sparsity for each layer, achieving the best trade-off between the accuracy and compression. This mixing scheme allows the entire network model to be stored in BRAMs of the FPGA to aggressively reduce the off-chip access, and thereby achieves a significant performance enhancement. The model size is reduced by 22.66-28.93 times compared to that in a full-precision network with a negligible degradation of accuracy on VOC, COCO, and ImageNet datasets. Furthermore, the combination of mixed dataflow and mixed precision significantly outperforms the previous works in terms of both throughput, off-chip access, and on-chip memory requirement.
The large model size, high computational operations, and vulnerability against membership inference attack (MIA) have impeded deep learning or deep neural networks (DNNs) popularity, especially on mobile devices. To address the challenge, we envision that the weight pruning technique will help DNNs against MIA while reducing model storage and computational operation. In this work, we propose a pruning algorithm, and we show that the proposed algorithm can find a subnetwork that can prevent privacy leakage from MIA and achieves competitive accuracy with the original DNNs. We also verify our theoretical insights with experiments. Our experimental results illustrate that the attack accuracy using model compression is up to 13.6% and 10% lower than that of the baseline and Min-Max game, accordingly.
Weight quantization for deep ConvNets has shown promising results for applications such as image classification and semantic segmentation and is especially important for applications where memory storage is limited. However, when aiming for quantization without accuracy degradation, different tasks may end up with different bitwidths. This creates complexity for software and hardware support and the complexity accumulates when one considers mixed-precision quantization, in which case each layer's weights use a different bitwidth. Our key insight is that optimizing for the least bitwidth subject to no accuracy degradation is not necessarily an optimal strategy. This is because one cannot decide optimality between two bitwidths if one has a smaller model size while the other has better accuracy. In this work, we take the first step to understand if some weight bitwidth is better than others by aligning all to the same model size using a width-multiplier. Under this setting, somewhat surprisingly, we show that using a single bitwidth for the whole network can achieve better accuracy compared to mixed-precision quantization targeting zero accuracy degradation when both have the same model size. In particular, our results suggest that when the number of channels becomes a target hyperparameter, a single weight bitwidth throughout the network shows superior results for model compression.
Model compression is crucial for deployment of neural networks on devices with limited computational and memory resources. Many different methods show comparable accuracy of the compressed model and similar compression rates. However, the majority of the compression methods are based on heuristics and offer no worst-case guarantees on the trade-off between the compression rate and the approximation error for an arbitrarily new sample. We propose the first efficient structured pruning algorithm with a provable trade-off between its compression rate and the approximation error for any future test sample. Our method is based on the coreset framework and it approximates the output of a layer of neurons/filters by a coreset of neurons/filters in the previous layer and discards the rest. We apply this framework in a layer-by-layer fashion from the bottom to the top. Unlike previous works, our coreset is data independent, meaning that it provably guarantees the accuracy of the function for any input $x\in \mathbb{R}^d$, including an adversarial one.
In this paper, we propose an approach for filter-level pruning with hierarchical knowledge distillation based on the teacher, teaching-assistant, and student framework. Our method makes use of teaching assistants at intermediate pruning levels that share the same architecture and weights as the target student. We propose to prune each model independently using the gradient information from its corresponding teacher. By considering the relative sizes of each student-teacher pair, this formulation provides a natural trade-off between the capacity gap for knowledge distillation and the bias of the filter saliency updates. Our results show improvements in the attainable accuracy and model compression across the CIFAR10 and ImageNet classification tasks using the VGG16and ResNet50 architectures. We provide an extensive evaluation that demonstrates the benefits of using a varying number of teaching assistant models at different sizes.
Compactly representing the visual signals is of fundamental importance in various image/video-centered applications. Although numerous approaches were developed for improving the image and video coding performance by removing the redundancies within visual signals, much less work has been dedicated to the transformation of the visual signals to another well-established modality for better representation capability. In this paper, we propose a new scheme for visual signal representation that leverages the philosophy of transferable modality. In particular, the deep learning model, which characterizes and absorbs the statistics of the input scene with online training, could be efficiently represented in the sense of rate-utility optimization to serve as the enhancement layer in the bitstream. As such, the overall performance can be further guaranteed by optimizing the new modality incorporated. The proposed framework is implemented on the state-of-the-art video coding standard (i.e., versatile video coding), and significantly better representation capability has been observed based on extensive evaluations.
Tensor Networks (TN) offer a powerful framework to efficiently represent very high-dimensional objects. TN have recently shown their potential for machine learning applications and offer a unifying view of common tensor decomposition models such as Tucker, tensor train (TT) and tensor ring (TR). However, identifying the best tensor network structure from data for a given task is challenging. In this work, we leverage the TN formalism to develop a generic and efficient adaptive algorithm to jointly learn the structure and the parameters of a TN from data. Our method is based on a simple greedy approach starting from a rank one tensor and successively identifying the most promising tensor network edges for small rank increments. Our algorithm can adaptively identify TN structures with small number of parameters that effectively optimize any differentiable objective function. Experiments on tensor decomposition, tensor completion and model compression tasks demonstrate the effectiveness of the proposed algorithm. In particular, our method outperforms the state-of-the-art evolutionary topology search [Li and Sun, 2020] for tensor decomposition of images (while being orders of magnitude faster) and finds efficient tensor network structures to compress neural networks outperforming popular TT based approaches [Novikov et al., 2015].
Increasing demand for on-device Automatic Speech Recognition (ASR) systems has resulted in renewed interests in developing automatic model compression techniques. Past research have shown that AutoML-based Low Rank Factorization (LRF) technique, when applied to an end-to-end Encoder-Attention-Decoder style ASR model, can achieve a speedup of up to 3.7x, outperforming laborious manual rank-selection approaches. However, we show that current AutoML-based search techniques only work up to a certain compression level, beyond which they fail to produce compressed models with acceptable word error rates (WER). In this work, we propose an iterative AutoML-based LRF approach that achieves over 5x compression without degrading the WER, thereby advancing the state-of-the-art in ASR compression.
In this work, we tackle model efficiency by exploiting redundancy in the \textit{implicit structure} of the building blocks of convolutional neural networks. We start our analysis by introducing a general definition of Composite Kernel structures that enable the execution of convolution operations in the form of efficient, scaled, sum-pooling components. As its special case, we propose \textit{Structured Convolutions} and show that these allow decomposition of the convolution operation into a sum-pooling operation followed by a convolution with significantly lower complexity and fewer weights. We show how this decomposition can be applied to 2D and 3D kernels as well as the fully-connected layers. Furthermore, we present a Structural Regularization loss that promotes neural network layers to leverage on this desired structure in a way that, after training, they can be decomposed with negligible performance loss. By applying our method to a wide range of CNN architectures, we demonstrate "structured" versions of the ResNets that are up to 2$\times$ smaller and a new Structured-MobileNetV2 that is more efficient while staying within an accuracy loss of 1% on ImageNet and CIFAR-10 datasets. We also show similar structured versions of EfficientNet on ImageNet and HRNet architecture for semantic segmentation on the Cityscapes dataset. Our method performs equally well or superior in terms of the complexity reduction in comparison to the existing tensor decomposition and channel pruning methods.
We approach the problem of implicit regularization in deep learning from a geometrical viewpoint. We highlight a regularization effect induced by a dynamical alignment of the neural tangent features introduced by Jacot et al, along a small number of task-relevant directions. This can be interpreted as a combined mechanism of feature selection and compression. By extrapolating a new analysis of Rademacher complexity bounds for linear models, we motivate and study a heuristic complexity measure that captures this phenomenon, in terms of sequences of tangent kernel classes along optimization paths.
Knowledge distillation has become increasingly important in model compression. It boosts the performance of a miniaturized student network with the supervision of the output distribution and feature maps from a sophisticated teacher network. Some recent works introduce multi-teacher distillation to provide more supervision to the student network. However, the effectiveness of multi-teacher distillation methods are accompanied by costly computation resources. To tackle with both the efficiency and the effectiveness of knowledge distillation, we introduce the feature aggregation to imitate the multi-teacher distillation in the single-teacher distillation framework by extracting informative supervision from multiple teacher feature maps. Specifically, we introduce DFA, a two-stage Differentiable Feature Aggregation search method that motivated by DARTS in neural architecture search, to efficiently find the aggregations. In the first stage, DFA formulates the searching problem as a bi-level optimization and leverages a novel bridge loss, which consists of a student-to-teacher path and a teacher-to-student path, to find appropriate feature aggregations. The two paths act as two players against each other, trying to optimize the unified architecture parameters to the opposite directions while guaranteeing both expressivity and learnability of the feature aggregation simultaneously. In the second stage, DFA performs knowledge distillation with the derived feature aggregation. Experimental results show that DFA outperforms existing methods on CIFAR-100 and CINIC-10 datasets under various teacher-student settings, verifying the effectiveness and robustness of the design.
This paper proposes \textit{layer fusion} - a model compression technique that discovers which weights to combine and then fuses weights of similar fully-connected, convolutional and attention layers. Layer fusion can significantly reduce the number of layers of the original network with little additional computation overhead, while maintaining competitive performance. From experiments on CIFAR-10, we find that various deep convolution neural networks can remain within 2\% accuracy points of the original networks up to a compression ratio of 3.33 when iteratively retrained with layer fusion. For experiments on the WikiText-2 language modelling dataset where pretrained transformer models are used, we achieve compression that leads to a network that is 20\% of its original size while being within 5 perplexity points of the original network. We also find that other well-established compression techniques can achieve competitive performance when compared to their original networks given a sufficient number of retraining steps. Generally, we observe a clear inflection point in performance as the amount of compression increases, suggesting a bound on the amount of compression that can be achieved before an exponential degradation in performance.
Closing the gap between the hardware requirements of state-of-the-art convolutional neural networks and the limited resources constraining embedded applications is the next big challenge in deep learning research. The computational complexity and memory footprint of such neural networks are typically daunting for deployment in resource constrained environments. Model compression techniques, such as pruning, are emphasized among other optimization methods for solving this problem. Most existing techniques require domain expertise or result in irregular sparse representations, which increase the burden of deploying deep learning applications on embedded hardware accelerators. In this paper, we propose the autoencoder-based low-rank filter-sharing technique technique (ALF). When applied to various networks, ALF is compared to state-of-the-art pruning methods, demonstrating its efficient compression capabilities on theoretical metrics as well as on an accurate, deterministic hardware-model. In our experiments, ALF showed a reduction of 70\% in network parameters, 61\% in operations and 41\% in execution time, with minimal loss in accuracy.
Mobile devices are becoming an important carrier for deep learning tasks, as they are being equipped with powerful, high-end mobile CPUs and GPUs. However, it is still a challenging task to execute 3D Convolutional Neural Networks (CNNs) targeting for real-time performance, besides high inference accuracy. The reason is more complex model structure and higher model dimensionality overwhelm the available computation/storage resources on mobile devices. A natural way may be turning to deep learning weight pruning techniques. However, the direct generalization of existing 2D CNN weight pruning methods to 3D CNNs is not ideal for fully exploiting mobile parallelism while achieving high inference accuracy. This paper proposes RT3D, a model compression and mobile acceleration framework for 3D CNNs, seamlessly integrating neural network weight pruning and compiler code generation techniques. We propose and investigate two structured sparsity schemes i.e., the vanilla structured sparsity and kernel group structured (KGS) sparsity that are mobile acceleration friendly. The vanilla sparsity removes whole kernel groups, while KGS sparsity is a more fine-grained structured sparsity that enjoys higher flexibility while exploiting full on-device parallelism. We propose a reweighted regularization pruning algorithm to achieve the proposed sparsity schemes. The inference time speedup due to sparsity is approaching the pruning rate of the whole model FLOPs (floating point operations). RT3D demonstrates up to 29.1$\times$ speedup in end-to-end inference time comparing with current mobile frameworks supporting 3D CNNs, with moderate 1%-1.5% accuracy loss. The end-to-end inference time for 16 video frames could be within 150 ms, when executing representative C3D and R(2+1)D models on a cellphone. For the first time, real-time execution of 3D CNNs is achieved on off-the-shelf mobiles.
In natural language processing (NLP), the "Transformer" architecture was proposed as the first transduction model replying entirely on self-attention mechanisms without using sequence-aligned recurrent neural networks (RNNs) or convolution, and it achieved significant improvements for sequence to sequence tasks. The introduced intensive computation and storage of these pre-trained language representations has impeded their popularity into computation and memory-constrained devices. The field-programmable gate array (FPGA) is widely used to accelerate deep learning algorithms for its high parallelism and low latency. However, the trained models are still too large to accommodate to an FPGA fabric. In this paper, we propose an efficient acceleration framework, Ftrans, for transformer-based large scale language representations. Our framework includes enhanced block-circulant matrix (BCM)-based weight representation to enable model compression on large-scale language representations at the algorithm level with few accuracy degradation, and an acceleration design at the architecture level. Experimental results show that our proposed framework significantly reduces the model size of NLP models by up to 16 times. Our FPGA design achieves 27.07x and 81x improvement in performance and energy efficiency compared to CPU, and up to 8.80x improvement in energy efficiency compared to GPU.
Learning generic representations with deep networks requires massive training samples and significant computer resources. To learn a new specific task, an important issue is to transfer the generic teacher's representation to a student network. In this paper, we propose to use a metric between representations that is based on a functional view of neurons. We use optimal transport to quantify the match between two representations, yielding a distance that embeds some invariances inherent to the representation of deep networks. This distance defines a regularizer promoting the similarity of the student's representation with that of the teacher. Our approach can be used in any learning context where representation transfer is applicable. We experiment here on two standard settings: inductive transfer learning, where the teacher's representation is transferred to a student network of same architecture for a new related task, and knowledge distillation, where the teacher's representation is transferred to a student of simpler architecture for the same task (model compression). Our approach also lends itself to solving new learning problems; we demonstrate this by showing how to directly transfer the teacher's representation to a simpler architecture student for a new related task.
This paper proposes PuRL - a deep reinforcement learning (RL) based algorithm for pruning neural networks. Unlike current RL based model compression approaches where feedback is given only at the end of each episode to the agent, PuRL provides rewards at every pruning step. This enables PuRL to achieve sparsity and accuracy comparable to current state-of-the-art methods, while having a much shorter training cycle. PuRL achieves more than 80% sparsity on the ResNet-50 model while retaining a Top-1 accuracy of 75.37% on the ImageNet dataset. Through our experiments we show that PuRL is also able to sparsify already efficient architectures like MobileNet-V2. In addition to performance characterisation experiments, we also provide a discussion and analysis of the various RL design choices that went into the tuning of the Markov Decision Process underlying PuRL. Lastly, we point out that PuRL is simple to use and can be easily adapted for various architectures.
Soft labeling becomes a common output regularization for generalization and model compression of deep neural networks. However, the effect of soft labeling on out-of-distribution (OOD) detection, which is an important topic of machine learning safety, is not explored. In this study, we show that soft labeling can determine OOD detection performance. Specifically, how to regularize outputs of incorrect classes by soft labeling can deteriorate or improve OOD detection. Based on the empirical results, we postulate a future work for OOD-robust DNNs: a proper output regularization by soft labeling can construct OOD-robust DNNs without additional training of OOD samples or modifying the models, while improving classification accuracy.
Knowledge distillation (KD) is commonly deemed as an effective model compression technique in which a compact model (student) is trained under the supervision of a larger pretrained model or an ensemble of models (teacher). Various techniques have been proposed since the original formulation, which mimic different aspects of the teacher such as the representation space, decision boundary, or intra-data relationship. Some methods replace the one-way knowledge distillation from a static teacher with collaborative learning between a cohort of students. Despite the recent advances, a clear understanding of where knowledge resides in a deep neural network and an optimal method for capturing knowledge from teacher and transferring it to student remains an open question. In this study, we provide an extensive study on nine different KD methods which covers a broad spectrum of approaches to capture and transfer knowledge. We demonstrate the versatility of the KD framework on different datasets and network architectures under varying capacity gaps between the teacher and student. The study provides intuition for the effects of mimicking different aspects of the teacher and derives insights from the performance of the different distillation approaches to guide the design of more effective KD methods. Furthermore, our study shows the effectiveness of the KD framework in learning efficiently under varying severity levels of label noise and class imbalance, consistently providing generalization gains over standard training. We emphasize that the efficacy of KD goes much beyond a model compression technique and it should be considered as a general-purpose training paradigm which offers more robustness to common challenges in the real-world datasets compared to the standard training procedure.
Model compression and acceleration are attracting increasing attentions due to the demand for embedded devices and mobile applications. Research on efficient convolutional neural networks (CNNs) aims at removing feature redundancy by decomposing or optimizing the convolutional calculation. In this work, feature redundancy is assumed to exist among channels in CNN architectures, which provides some leeway to boost calculation efficiency. Aiming at channel compression, a novel convolutional construction named compact convolution is proposed to embrace the progress in spatial convolution, channel grouping and pooling operation. Specifically, the depth-wise separable convolution and the point-wise interchannel operation are utilized to efficiently extract features. Different from the existing channel compression method which usually introduces considerable learnable weights, the proposed compact convolution can reduce feature redundancy with no extra parameters. With the point-wise interchannel operation, compact convolutions implicitly squeeze the channel dimension of feature maps. To explore the rules on reducing channel redundancy in neural networks, the comparison is made among different point-wise interchannel operations. Moreover, compact convolutions are extended to tackle with multiple tasks, such as acoustic scene classification, sound event detection and image classification. The extensive experiments demonstrate that our compact convolution not only exhibits high effectiveness in several multimedia tasks, but also can be efficiently implemented by benefiting from parallel computation.