We present Polynomial Attention Drop-in Replacement (PADRe), a novel and unifying framework designed to replace the conventional self-attention mechanism in transformer models. Notably, several recent alternative attention mechanisms, including Hyena, Mamba, SimA, Conv2Former, and Castling-ViT, can be viewed as specific instances of our PADRe framework. PADRe leverages polynomial functions and draws upon established results from approximation theory, enhancing computational efficiency without compromising accuracy. PADRe's key components include multiplicative nonlinearities, which we implement using straightforward, hardware-friendly operations such as Hadamard products, incurring only linear computational and memory costs. PADRe further avoids the need for using complex functions such as Softmax, yet it maintains comparable or superior accuracy compared to traditional self-attention. We assess the effectiveness of PADRe as a drop-in replacement for self-attention across diverse computer vision tasks. These tasks include image classification, image-based 2D object detection, and 3D point cloud object detection. Empirical results demonstrate that PADRe runs significantly faster than the conventional self-attention (11x~43x faster on server GPU and mobile NPU) while maintaining similar accuracy when substituting self-attention in the transformer models.

Neural Architecture Search (NAS) automates and prospers the design of neural networks. Estimator-based NAS has been proposed recently to model the relationship between architectures and their performance to enable scalable and flexible search. However, existing estimator-based methods encode the architecture into a latent space without considering graph similarity. Ignoring graph similarity in node-based search space may induce a large inconsistency between similar graphs and their distance in the continuous encoding space, leading to inaccurate encoding representation and/or reduced representation capacity that can yield sub-optimal search results. To preserve graph correlation information in encoding, we propose NASGEM which stands for Neural Architecture Search via Graph Embedding Method. NASGEM is driven by a novel graph embedding method equipped with similarity measures to capture the graph topology information. By precisely estimating the graph distance and using an auxiliary Weisfeiler-Lehman kernel to guide the encoding, NASGEM can utilize additional structural information to get more accurate graph representation to improve the search efficiency. GEMNet, a set of networks discovered by NASGEM, consistently outperforms networks crafted by existing search methods in classification tasks, i.e., with 0.4%-3.6% higher accuracy while having 11%- 21% fewer Multiply-Accumulates. We further transfer GEMNet for COCO object detection. In both one-stage and twostage detectors, our GEMNet surpasses its manually-crafted and automatically-searched counterparts.

Resource is an important constraint when deploying Deep Neural Networks (DNNs) on mobile and edge devices. Existing works commonly adopt the cell-based search approach, which limits the flexibility of network patterns in learned cell structures. Moreover, due to the topology-agnostic nature of existing works, including both cell-based and node-based approaches, the search process is time consuming and the performance of found architecture may be sub-optimal. To address these problems, we propose AutoShrink, a topology-aware Neural Architecture Search (NAS) for searching efficient building blocks of neural architectures. Our method is node-based and thus can learn flexible network patterns in cell structures within a topological search space. Directed Acyclic Graphs (DAGs) are used to abstract DNN architectures and progressively optimize the cell structure through edge shrinking. As the search space intrinsically reduces as the edges are progressively shrunk, AutoShrink explores more flexible search space with even less search time. We evaluate AutoShrink on image classification and language tasks by crafting ShrinkCNN and ShrinkRNN models. ShrinkCNN is able to achieve up to 48% parameter reduction and save 34% Multiply-Accumulates (MACs) on ImageNet-1K with comparable accuracy of state-of-the-art (SOTA) models. Specifically, both ShrinkCNN and ShrinkRNN are crafted within 1.5 GPU hours, which is 7.2× and 6.7× faster than the crafting time of SOTA CNN and RNN models, respectively.

In recent years, neuromorphic computing systems (NCS) have gained popularity in accelerating neural network computation because of their high energy efficiency. The known vulnerability of neural networks to adversarial attack, however, raises a severe security concern of NCS. In addition, there are certain application scenarios in which users have limited access to the NCS. In such scenarios, defense technologies that require changing the training methods of the NCS, e.g., adversarial training become impracticable. In this work, we propose AdverQuil - an efficient adversarial detection and alleviation technique for black-box NCS. AdverQuil can identify the adversarial strength of input examples and select the best strategy for NCS to respond to the attack, without changing structure/parameter of the original neural network or its training method. Experimental results show that on MNIST and CIFAR-10 datasets, AdverQuil achieves a high efficiency of 79.5 - 167K image/sec/watt. AdverQuil introduces less than 25% of hardware overhead, and can be combined with various adversarial alleviation techniques to provide a flexible trade-off between hardware cost, energy efficiency and classification accuracy.

Recurrent neural networks (RNNs), especially long short-term memories (LSTMs) have been widely used in speech recognition and natural language processing. As the sizes of RNN models grow for better performance, the computation cost and therefore the required hardware resource increase rapidly. We propose an efficient structural sparsity (ESS) learning method for acoustic modeling in speech recognition. ESS aims to generate a model that offers higher execution efficiency while maintaining the accuracy. A three-step training pipeline is developed in our work. First, we apply the group Lasso regularization method during training process and learn a structural sparse model from scratch. Then the learned sparse structures will be fixed and cannot be changed. Finally, we retrain the model and update the nonzero parameters in the model. We applied our ESS method on classic HMM+LSTM model on Kaldi toolkit. The experimental results show that ESS can remove 72.5% weight groups in the weight matrices when slightly increasing the word error rate (WER) 1.1%.

As a large-scale commercial spiking-based neuromorphic computing platform, IBM TrueNorth processor received tremendous attentions in society. However, one of the known issues in TrueNorth design is the limited precision of synaptic weights. The current workaround is running multiple neural network copies in which the average value of each synaptic weight is close to that in the original network. We theoretically analyze the impacts of low data precision in the TrueNorth chip on inference accuracy, core occupation, and performance, and present a probability-biased learning method to enhance the inference accuracy through reducing the random variance of each computation copy. Our experimental results proved that the proposed techniques considerably improve the computation accuracy of TrueNorth platform and reduce the incurred hardware and performance overheads. Among all the tested methods, L1TEA regularization achieved the best result, say, up to 2.74% accuracy enhancement when deploying MNIST application onto TrueNorth platform. In May 2016, IBM TrueNorth team implemented convolutional neural networks (CNN) on TrueNorth processor and coincidently use a similar method, say, trinary weights, {-1, 0, 1}. It achieves near state-of-the-art accuracy on 8 standard datasets. In addition, to further evaluate TrueNorth performance on CNN, we test similar deep convolutional networks on True-North, GPU and FPGA. Among all, GPU has the highest through-put. But if we consider energy consumption, TrueNorth processor is the most energy-efficient one, say, > 6000 frames/sec/Watt.