Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression

Large Language Models (LLMs) have achieved remarkable breakthroughs. However, the huge number of parameters in LLMs require significant amount of memory storage in inference, which prevents their practical deployment in many applications. To reduce memory storage of LLMs, singular value decomposition (SVD) provides a promising solution to approximate weight matrices for compressing LLMs. In this paper, we take a step further to explore parameter sharing across different layers with SVD to achieve more effective compression for LLMs. Specifically, weight matrices in different layers are decomposed and represented as a linear combination of a set of shared basis vectors and unique coefficients. The types of weight matrices and the layer selection for basis sharing are examined when compressing LLMs to maintain the performance. Comprehensive experiments demonstrate that Basis Sharing outperforms state-of-the-art SVD-based compression approaches and parameter sharing techniques, especially under large compression ratios. Code is available at: https://github.com/TUDa-HWAI/Basis_Sharing

BasisN: Reprogramming-Free RRAM-Based In-Memory-Computing by Basis Combination for Deep Neural Networks

Deep neural networks (DNNs) have made breakthroughs in various fields including image recognition and language processing. DNNs execute hundreds of millions of multiply-and-accumulate (MAC) operations. To efficiently accelerate such computations, analog in-memory-computing platforms have emerged leveraging emerging devices such as resistive RAM (RRAM). However, such accelerators face the hurdle of being required to have sufficient on-chip crossbars to hold all the weights of a DNN. Otherwise, RRAM cells in the crossbars need to be reprogramed to process further layers, which causes huge time/energy overhead due to the extremely slow writing and verification of the RRAM cells. As a result, it is still not possible to deploy such accelerators to process large-scale DNNs in industry. To address this problem, we propose the BasisN framework to accelerate DNNs on any number of available crossbars without reprogramming. BasisN introduces a novel representation of the kernels in DNN layers as combinations of global basis vectors shared between all layers with quantized coefficients. These basis vectors are written to crossbars only once and used for the computations of all layers with marginal hardware modification. BasisN also provides a novel training approach to enhance computation parallelization with the global basis vectors and optimize the coefficients to construct the kernels. Experimental results demonstrate that cycles per inference and energy-delay product were reduced to below 1% compared with applying reprogramming on crossbars in processing large-scale DNNs such as DenseNet and ResNet on ImageNet and CIFAR100 datasets, while the training and hardware costs are negligible.

LiveMind: Low-latency Large Language Models with Simultaneous Inference

In this paper, we introduce a novel low-latency inference framework for large language models (LLMs) inference which enables LLMs to perform inferences with incomplete prompts. By reallocating computational processes to prompt input phase, we achieve a substantial reduction in latency, thereby significantly enhancing the interactive experience for users of LLMs. The framework adeptly manages the visibility of the streaming prompt to the model, allowing it to infer from incomplete prompts or await additional prompts. Compared with traditional inference methods that utilize complete prompts, our approach demonstrates an average reduction of 59% in response latency on the MMLU-Pro dataset, while maintaining comparable accuracy. Additionally, our framework facilitates collaborative inference and output across different models. By employing an LLM for inference and a small language model (SLM) for output, we achieve an average 68% reduction in response latency, alongside a 5.5% improvement in accuracy on the MMLU-Pro dataset compared with the SLM baseline. For long prompts exceeding 20 sentences, the response latency can be reduced by up to 93%.

EncodingNet: A Novel Encoding-based MAC Design for Efficient Neural Network Acceleration

Deep neural networks (DNNs) have achieved great breakthroughs in many fields such as image classification and natural language processing. However, the execution of DNNs needs to conduct massive numbers of multiply-accumulate (MAC) operations on hardware and thus incurs a large power consumption. To address this challenge, we propose a novel digital MAC design based on encoding. In this new design, the multipliers are replaced by simple logic gates to project the results onto a wide bit representation. These bits carry individual position weights, which can be trained for specific neural networks to enhance inference accuracy. The outputs of the new multipliers are added by bit-wise weighted accumulation and the accumulation results are compatible with existing computing platforms accelerating neural networks with either uniform or non-uniform quantization. Since the multiplication function is replaced by simple logic projection, the critical paths in the resulting circuits become much shorter. Correspondingly, pipelining stages in the MAC array can be reduced, leading to a significantly smaller area as well as a better power efficiency. The proposed design has been synthesized and verified by ResNet18-Cifar10, ResNet20-Cifar100 and ResNet50-ImageNet. The experimental results confirmed the reduction of circuit area by up to 79.63% and the reduction of power consumption of executing DNNs by up to 70.18%, while the accuracy of the neural networks can still be well maintained.

Class-Aware Pruning for Efficient Neural Networks

Deep neural networks (DNNs) have demonstrated remarkable success in various fields. However, the large number of floating-point operations (FLOPs) in DNNs poses challenges for their deployment in resource-constrained applications, e.g., edge devices. To address the problem, pruning has been introduced to reduce the computational cost in executing DNNs. Previous pruning strategies are based on weight values, gradient values and activation outputs. Different from previous pruning solutions, in this paper, we propose a class-aware pruning technique to compress DNNs, which provides a novel perspective to reduce the computational cost of DNNs. In each iteration, the neural network training is modified to facilitate the class-aware pruning. Afterwards, the importance of filters with respect to the number of classes is evaluated. The filters that are only important for a few number of classes are removed. The neural network is then retrained to compensate for the incurred accuracy loss. The pruning iterations end until no filter can be removed anymore, indicating that the remaining filters are very important for many classes. This pruning technique outperforms previous pruning solutions in terms of accuracy, pruning ratio and the reduction of FLOPs. Experimental results confirm that this class-aware pruning technique can significantly reduce the number of weights and FLOPs, while maintaining a high inference accuracy.

Early Classification for Dynamic Inference of Neural Networks

Deep neural networks (DNNs) have been successfully applied in various fields. In DNNs, a large number of multiply-accumulate (MAC) operations is required to be performed, posing critical challenges in applying them in resource-constrained platforms, e.g., edge devices. Dynamic neural networks have been introduced to allow a structural adaption, e.g., early-exit, according to different inputs to reduce the computational cost of DNNs. Existing early-exit techniques deploy classifiers at intermediate layers of DNNs to push them to make a classification decision as early as possible. However, the learned features at early layers might not be sufficient to exclude all the irrelevant classes and decide the correct class, leading to suboptimal results. To address this challenge, in this paper, we propose a class-based early-exit for dynamic inference. Instead of pushing DNNs to make a dynamic decision at intermediate layers, we take advantages of the learned features in these layers to exclude as many irrelevant classes as possible, so that later layers only have to determine the target class among the remaining classes. Until at a layer only one class remains, this class is the corresponding classification result. To realize this class-based exclusion, we assign each class with a classifier at intermediate layers and train the networks together with these classifiers. Afterwards, an exclusion strategy is developed to exclude irrelevant classes at early layers. Experimental results demonstrate the computational cost of DNNs in inference can be reduced significantly.

Logic Design of Neural Networks for High-Throughput and Low-Power Applications

Neural networks (NNs) have been successfully deployed in various fields. In NNs, a large number of multiplyaccumulate (MAC) operations need to be performed. Most existing digital hardware platforms rely on parallel MAC units to accelerate these MAC operations. However, under a given area constraint, the number of MAC units in such platforms is limited, so MAC units have to be reused to perform MAC operations in a neural network. Accordingly, the throughput in generating classification results is not high, which prevents the application of traditional hardware platforms in extreme-throughput scenarios. Besides, the power consumption of such platforms is also high, mainly due to data movement. To overcome this challenge, in this paper, we propose to flatten and implement all the operations at neurons, e.g., MAC and ReLU, in a neural network with their corresponding logic circuits. To improve the throughput and reduce the power consumption of such logic designs, the weight values are embedded into the MAC units to simplify the logic, which can reduce the delay of the MAC units and the power consumption incurred by weight movement. The retiming technique is further used to improve the throughput of the logic circuits for neural networks. In addition, we propose a hardware-aware training method to reduce the area of logic designs of neural networks. Experimental results demonstrate that the proposed logic designs can achieve high throughput and low power consumption for several high-throughput applications.

Expressivity Enhancement with Efficient Quadratic Neurons for Convolutional Neural Networks

Convolutional neural networks (CNNs) have been successfully applied in a range of fields such as image classification and object segmentation. To improve their expressivity, various techniques, such as novel CNN architectures, have been explored. However, the performance gain from such techniques tends to diminish. To address this challenge, many researchers have shifted their focus to increasing the non-linearity of neurons, the fundamental building blocks of neural networks, to enhance the network expressivity. Nevertheless, most of these approaches incur a large number of parameters and thus formidable computation cost inevitably, impairing their efficiency to be deployed in practice. In this work, an efficient quadratic neuron structure is proposed to preserve the non-linearity with only negligible parameter and computation cost overhead. The proposed quadratic neuron can maximize the utilization of second-order computation information to improve the network performance. The experimental results have demonstrated that the proposed quadratic neuron can achieve a higher accuracy and a better computation efficiency in classification tasks compared with both linear neurons and non-linear neurons from previous works.

PowerPruning: Selecting Weights and Activations for Power-Efficient Neural Network Acceleration

Deep neural networks (DNNs) have been successfully applied in various fields. A major challenge of deploying DNNs, especially on edge devices, is power consumption, due to the large number of multiply-and-accumulate (MAC) operations. To address this challenge, we propose PowerPruning, a novel method to reduce power consumption in digital neural network accelerators by selecting weights that lead to less power consumption in MAC operations. In addition, the timing characteristics of the selected weights together with all activation transitions are evaluated. The weights and activations that lead to small delays are further selected. Consequently, the maximum delay of the sensitized circuit paths in the MAC units is reduced even without modifying MAC units, which thus allows a flexible scaling of supply voltage to reduce power consumption further. Together with retraining, the proposed method can reduce power consumption of DNNs on hardware by up to 78.3% with only a slight accuracy loss.

CorrectNet: Robustness Enhancement of Analog In-Memory Computing for Neural Networks by Error Suppression and Compensation

The last decade has witnessed the breakthrough of deep neural networks (DNNs) in many fields. With the increasing depth of DNNs, hundreds of millions of multiply-and-accumulate (MAC) operations need to be executed. To accelerate such operations efficiently, analog in-memory computing platforms based on emerging devices, e.g., resistive RAM (RRAM), have been introduced. These acceleration platforms rely on analog properties of the devices and thus suffer from process variations and noise. Consequently, weights in neural networks configured into these platforms can deviate from the expected values, which may lead to feature errors and a significant degradation of inference accuracy. To address this issue, in this paper, we propose a framework to enhance the robustness of neural networks under variations and noise. First, a modified Lipschitz constant regularization is proposed during neural network training to suppress the amplification of errors propagated through network layers. Afterwards, error compensation is introduced at necessary locations determined by reinforcement learning to rescue the feature maps with remaining errors. Experimental results demonstrate that inference accuracy of neural networks can be recovered from as low as 1.69% under variations and noise back to more than 95% of their original accuracy, while the training and hardware cost are negligible.