Clo-HDnn is an on-device learning (ODL) accelerator designed for emerging continual learning (CL) tasks. Clo-HDnn integrates hyperdimensional computing (HDC) along with low-cost Kronecker HD Encoder and weight clustering feature extraction (WCFE) to optimize accuracy and efficiency. Clo-HDnn adopts gradient-free CL to efficiently update and store the learned knowledge in the form of class hypervectors. Its dual-mode operation enables bypassing costly feature ex- traction for simpler datasets, while progressive search reduces complexity by up to $61 \%$ by encoding and comparing only partial query hypervectors. Achieving 4.66 TFLOPS/W (FE) and 3.78 TOPS/W (classifier), Clo-HDnn delivers $7.77 \times$ and $4.85 \times$ higher energy efficiency compared to SOTA ODL accelerators.

Hyperdimensional Computing (HDC), a technique inspired by cognitive models of computation, has been proposed as an efficient and robust alternative basis for machine learning.HDC programs are often manually written in low-level and target specific languages targeting CPUs, GPUs, and FPGAs-these codes cannot be easily retargeted onto HDC-specific accelerators.No previous programming system enables productive development of HDC programs and generates efficient code for several hardware targets.We propose a heterogeneous programming system for HDC: a novel programming language, HDC++, for writing applications using a unified programming model, including HDC-specific primitives to improve programmability, and a heterogeneous compiler, HPVM-HDC, that provides an intermediate representation for compiling HDC programs to many hardware targets.We implement two tuning optimizations, automatic binarization and reduction perforation, that exploit the error resilient nature of HDC.Our evaluation shows that HPVM-HDC generates performance-competitive code for CPUs and GPUs, achieving a geomean speed-up of 1.17x over optimized baseline CUDA implementations with a geomean * Equally contributing authors.

Weight clustering is an effective technique for compressing deep neural networks (DNNs) memory by using a limited number of unique weights and low-bit weight indexes to store clustering information. In this paper, we propose PatterNet, which enforces shared clustering topologies on filters. Cluster sharing leads to a greater extent of memory reduction by reusing the index information. PatterNet effectively factorizes input activations and post-processes the unique weights, which saves multiplications by several orders of magnitude. Furthermore, PatterNet reduces the add operations by harnessing the fact that filters sharing a clustering pattern have the same factorized terms. We introduce techniques for determining and assigning clustering patterns and training a network to fulfill the target patterns. We also propose and implement an efficient accelerator that builds upon the patterned filters. Experimental results show that PatterNet shrinks the memory and operation count up to 80.2% and 73.1%, respectively, with similar accuracy to the baseline models. PatterNet accelerator improves the energy efficiency by 107x over Nvidia 1080 1080 GTX and 2.2x over state of the art.

The efficiency of convolutional neural networks (CNNs) is dictated by the computationally intensive convolution operations. Thereby, extensive research has sought to alleviate the cost of convolution. In this work, we opt to skipping the ineffectual outputs, which can be used along with the other approaches such as quantization and pruning. Contrary to the previous works that skip ineffectual inputs, i.e., zero activations and weights, we predict and skip the ineffectual outputs by exploiting two key attributes of the state-of-the-art CNN architectures: ReLU (which filters out the negative outputs) is the widely-used activation function, and max-pooling (which discards the non-max outputs) is the commonly used down-sampling layer. To realize that, we propose TermiNETor, which breaks the convolution to bitwise shift-and-add operations. During the inference, weights are processed bit by bit, whereby TermiNETor uses a simple heuristic to predict whether the final output will be ineffectual. Upon prediction, TermiNETor terminates the convolution for ineffectual outputs. We propose a calibration flow that takes the prediction-based forward propagation into account and realizes more accurate predictions. In addition, we propose a novel hardware architecture to take advantage of early termination by feeding new convolution to the released processing elements. We demonstrate the effectiveness of TermiNETor on various networks and datasets. Experimental results show that our framework achieves up to 1.7× reduction of operation count compared to non-skipping baseline without accuracy degradation. TermiNETor accelerator, augmented with the skipping flow, improves the average energy efficiency by 3.84× over the zero weight/input skipping accelerator SCNN, and by 1.98× over FuseKNA which skips zero inputs as well as repetitive computations in a bit-serial manner. The proposed architecture exhibits good throughput and energy consumption scalability by increasing the processing elements, meaning that it can also target high-performance applications while preserving the energy efficiency.