K. Saláma

Wenzhe Guo, Mohammed E Fouda, Ahmed M. Eltawil et al.

Empowered by the backpropagation (BP) algorithm, deep neural networks have dominated the race in solving various cognitive tasks. The restricted training pattern in the standard BP requires end-to-end error propagation, causing large memory cost and prohibiting model parallelization. Existing local training methods aim to resolve the training obstacle by completely cutting off the backward path between modules and isolating their gradients to reduce memory cost and accelerate the training process. These methods prevent errors from flowing between modules and hence information exchange, resulting in inferior performance. This work proposes a novel local training algorithm, BackLink, which introduces inter-module backward dependency and allows errors to flow between modules. The algorithm facilitates information to flow backward along with the network. To preserve the computational advantage of local training, BackLink restricts the error propagation length within the module. Extensive experiments performed in various deep convolutional neural networks demonstrate that our method consistently improves the classification performance of local training algorithms over other methods. For example, in ResNet32 with 16 local modules, our method surpasses the conventional greedy local training method by 4.00\% and a recent work by 1.83\% in accuracy on CIFAR10, respectively. Analysis of computational costs reveals that small overheads are incurred in GPU memory costs and runtime on multiple GPUs. Our method can lead up to a 79\% reduction in memory cost and 52\% in simulation runtime in ResNet110 compared to the standard BP. Therefore, our method could create new opportunities for improving training algorithms towards better efficiency and biological plausibility.

Olga Krestinskaya, Li Zhang, Khaled Nabil Salama

The amount of data processed in the cloud, the development of Internet-of-Things (IoT) applications, and growing data privacy concerns force the transition from cloud-based to edge-based processing. Limited energy and computational resources on edge push the transition from traditional von Neumann architectures to In-memory Computing (IMC), especially for machine learning and neural network applications. Network compression techniques are applied to implement a neural network on limited hardware resources. Quantization is one of the most efficient network compression techniques allowing to reduce the memory footprint, latency, and energy consumption. This paper provides a comprehensive review of IMC-based Quantized Neural Networks (QNN) and links software-based quantization approaches to IMC hardware implementation. Moreover, open challenges, QNN design requirements, recommendations, and perspectives along with an IMC-based QNN hardware roadmap are provided.

Olga Krestinskaya, Mohammed E. Fouda, Ahmed Eltawil et al.

Designing generalized in-memory computing (IMC) hardware that efficiently supports a variety of workloads requires extensive design space exploration, which is infeasible to perform manually. Optimizing hardware individually for each workload or solely for the largest workload often fails to yield the most efficient generalized solutions. To address this, we propose a joint hardware-workload optimization framework that identifies optimised IMC chip architecture parameters, enabling more efficient, workload-flexible hardware. We show that joint optimization achieves 36%, 36%, 20%, and 69% better energy-latency-area scores for VGG16, ResNet18, AlexNet, and MobileNetV3, respectively, compared to the separate architecture parameters search optimizing for a single largest workload. Additionally, we quantify the performance trade-offs and losses of the resulting generalized IMC hardware compared to workload-specific IMC designs.

Mariam Rakka, Marios Fournarakis, Olga Krestinskaya et al.

The rapid scaling of language models (LMs) has resulted in unprecedented computational, memory, and energy requirements, making their training and deployment increasingly unsustainable. Quantization has emerged as an essential compression technique to reduce model size, alleviate memory bottlenecks, and accelerate inference. However, while uniform low-bit quantization (e.g., INT8, INT4) provides significant efficiency gains, it can degrade accuracy in sensitive components of transformer-based LMs. Mixed-precision quantization offers a promising alternative by selectively allocating precision across layers or within tensors to balance efficiency and accuracy. This survey provides a comprehensive overview of Mixed-Precision quantization frameworks for LMs (MXPLMs). We first review quantization fundamentals, including uniform and non-uniform quantizers, quantization granularity, and methods widely used in post-training quantization. We then categorize and compare recent MXPLM frameworks according to their bit allocation strategies and precision configurations across weights, activations, and key-value caches. A comparative analysis highlights differences in perplexity, zero-shot task performance, and deployment trade-offs. Furthermore, we contrast MXPLMs with earlier mixed-precision quantization methods for deep neural networks, identifying strategies that transfer and those that face challenges in the LM setting. Finally, we summarize open issues and future directions, including hardware-aware design, activation quantization, and scalable optimization methods for billion-parameter models. By consolidating recent advances, this work serves as a reference for understanding the current landscape and research prospects of mixed-precision quantization for large-scale language models.

Wenzhe Guo, Mohammed E. Fouda, Ahmed M. Eltawil et al.

Directly training spiking neural networks (SNNs) has remained challenging due to complex neural dynamics and intrinsic non-differentiability in firing functions. The well-known backpropagation through time (BPTT) algorithm proposed to train SNNs suffers from large memory footprint and prohibits backward and update unlocking, making it impossible to exploit the potential of locally-supervised training methods. This work proposes an efficient and direct training algorithm for SNNs that integrates a locally-supervised training method with a temporally-truncated BPTT algorithm. The proposed algorithm explores both temporal and spatial locality in BPTT and contributes to significant reduction in computational cost including GPU memory utilization, main memory access and arithmetic operations. We thoroughly explore the design space concerning temporal truncation length and local training block size and benchmark their impact on classification accuracy of different networks running different types of tasks. The results reveal that temporal truncation has a negative effect on the accuracy of classifying frame-based datasets, but leads to improvement in accuracy on dynamic-vision-sensor (DVS) recorded datasets. In spite of resulting information loss, local training is capable of alleviating overfitting. The combined effect of temporal truncation and local training can lead to the slowdown of accuracy drop and even improvement in accuracy. In addition, training deep SNNs models such as AlexNet classifying CIFAR10-DVS dataset leads to 7.26% increase in accuracy, 89.94% reduction in GPU memory, 10.79% reduction in memory access, and 99.64% reduction in MAC operations compared to the standard end-to-end BPTT.

Kamilya Smagulova, Mohammed E. Fouda, Fadi Kurdahi et al.

Deep Neural Networks (DNNs), as a subset of Machine Learning (ML) techniques, entail that real-world data can be learned and that decisions can be made in real-time. However, their wide adoption is hindered by a number of software and hardware limitations. The existing general-purpose hardware platforms used to accelerate DNNs are facing new challenges associated with the growing amount of data and are exponentially increasing the complexity of computations. An emerging non-volatile memory (NVM) devices and processing-in-memory (PIM) paradigm is creating a new hardware architecture generation with increased computing and storage capabilities. In particular, the shift towards ReRAM-based in-memory computing has great potential in the implementation of area and power efficient inference and in training large-scale neural network architectures. These can accelerate the process of the IoT-enabled AI technologies entering our daily life. In this survey, we review the state-of-the-art ReRAM-based DNN many-core accelerators, and their superiority compared to CMOS counterparts was shown. The review covers different aspects of hardware and software realization of DNN accelerators, their present limitations, and future prospectives. In particular, comparison of the accelerators shows the need for the introduction of new performance metrics and benchmarking standards. In addition, the major concerns regarding the efficient design of accelerators include a lack of accuracy in simulation tools for software and hardware co-design.

Olga Krestinskaya, Khaled Nabil Salama, Alex Pappachen James

The on-chip implementation of learning algorithms would speed-up the training of neural networks in crossbar arrays. The circuit level design and implementation of backpropagation algorithm using gradient descent operation for neural network architectures is an open problem. In this paper, we proposed the analog backpropagation learning circuits for various memristive learning architectures, such as Deep Neural Network (DNN), Binary Neural Network (BNN), Multiple Neural Network (MNN), Hierarchical Temporal Memory (HTM) and Long-Short Term Memory (LSTM). The circuit design and verification is done using TSMC 180nm CMOS process models, and TiO2 based memristor models. The application level validations of the system are done using XOR problem, MNIST character and Yale face image databases

Irina Dolzhikova, Khaled Salama, Vipin Kizheppatt et al.

Synaptic Sampling Machine (SSM) is a type of neural network model that considers biological unreliability of the synapses. We propose the circuit design of the SSM neural network which is realized through the memristive-CMOS crossbar structure with the synaptic sampling cell (SSC) being used as a basic stochastic unit. The increase in the edge computing devices in the Internet of things era, drives the need for hardware acceleration for data processing and computing. The computational considerations of the processing speed and possibility for the real-time realization pushes the synaptic sampling algorithm that demonstrated promising results on software for hardware implementation.