A comprehensive review of Binary Neural Network

1.1. Background

• Growing CNN compute and storage demands limit deployment on smartphones and tiny IoT devices, motivating compact, lightweight networks that retain acceptable accuracy.Model compression approaches include parameter quantization, pruning, low-rank factorization, compact convolutional filters, and knowledge distillation.Source: §1.1 · pp. 1–2

1.2. Motivation

• BNNs use 1-bit activations and weights, theoretically enabling 32 times lower memory storage and 58 times faster inference speed than traditional 32-bit CNNs.The introduction notes applications including classification, pattern recognition, computer vision, and natural language processing, and use on resource-limited devices.Source: §1.2 · pp. 2–3

1.3. Related Work

• This review addresses gaps and errors in earlier surveys by focusing on convolutional BNNs with 1-bit activations and weights across optimization, deployment, and applications.It groups optimization techniques into five categories and lists works by contribution, since a work may use multiple techniques.Source: §1.3 · p. 3

1.4. Organization

• The paper proceeds from BNN principles and early challenges through optimization, modeling frameworks, hardware, applications, and a concluding summary.The applications section includes tables tracking performance progress.Source: §1.4 · p. 3

2.1. What is BNN?

• BNNs use 1-bit activations and weights in hidden layers, replacing floating-point matrix operations with XNOR and popcount to reduce storage and computation.Rastegari et al. (2016) reported 32 times lower memory saving and 58 times faster convolution operations than 32-bit CNNs.Source: §2.1 · p. 4

2.2. Forward Propagation

• BNN forward propagation binarizes activations and weights before convolution, enabling bitwise XNOR and popcount to replace matrix multiplication.The neural cell adds binarization steps absent from the 32-bit CNN computation path.Source: §2.2 · p. 5

2.3. Backward Propagation

• Because the sign function has zero derivative, Binarized-Neural-Networks use STE to learn binary weights, while Bitwise-Neural-Networks use a two-step training strategy.In Binarized-Neural-Networks, real weights are updated during training and discarded after binarized weights are saved; Bitwise-Neural-Networks first trains compressed real-valued parameters, then initializes and trains the target network with a strategy similar to STE (Courbariaux et al., 2016; Kim and Smaragdis, 2016).Source: §2.3 · p. 6

2.4. Summary

• Naive BNNs offer faster inference and smaller weight sizes, but parameter binarization causes information loss and lower accuracy than full-precision CNNs.The review introduces later optimization methods as responses to this accuracy limitation.Source: §2.4 · p. 6

3.1.1. Scaling Factor

• Quantization-error methods reduce information loss through scaling, alternative binary representations, and reshaping activation and weight distributions before binarization.XNOR-Net adds channel-wise scaling factors α and β, while XNOR-Net++ (Bulat and Tzimiropoulos, 2019) merges them; SI-BNN (Wang et al., 2020) uses [0, +1] activations and [-1, +1] weights, and AdaBin selects binary sets per layer. Other approaches reshape distributions before binarization, including IR-Net and Bi-half.Source: §3.1.1 · p. 7, §3.1.2 · p. 8, §3.1.3 · p. 9

3.2. Loss Function Improvement

• Distribution losses or regularization supplement cross-entropy, while LNS, Real-to-Bin (Martinez et al., 2020), and ReActNet (Liu et al., 2020) propose losses for transfer learning.The total loss combines cross-entropy with an added distribution loss or regularization term, weighted by a balancing hyperparameter λ.Source: §3.2 · pp. 9–10

3.3. Gradient Approximation

• Because the sign function has zero derivative, BNNs use gradient approximations, including STE variants, polynomial or sigmoid functions, dynamic estimators, and smoothed-loss gradients.IR-Net and RBNN adjust gradient approximations during training; BinaryDuo (Kim et al., 2020) estimates gradients from a smoothed loss function.Source: §3.3 · p. 10

3.4. Network Topology Structure

• Topology methods modify classical networks, search or customize architectures, and develop lightweight BNNs; MeliusNet (Bethge et al., 2020) and ReActNet (Liu et al., 2020) are reported to surpass full-precision lightweight MobileNet accuracy with fewer operations.Other approaches include ensemble BNNs, channel-search methods, and MobileNet-inspired designs; FracBNN (Zhang et al., 2021) uses two 1-bit activation schemes and has competitive ImageNet top-1 prediction results versus full-precision MobileNet-v2.Source: §3.4 · p. 13

3.5. Training Strategy and Tricks

• Training strategies include progressive quantization, optimizer and activation changes, pruning, transfer from pretrained real-valued networks, and direct on-chip or edge training.Real-to-Bin (Martinez et al., 2020) uses two-step transfer teaching, and BNN-Adam reports improving Real-to-Bin and ReActNet performance; BNN-EP explores on-chip training with compact circuitry.Source: §3.5 · pp. 14–15

3.6. Summary

• Open questions include layer-wise information-loss optimization, balancing accuracy and operations, reducing BNN training time, and whether training tricks generalize across datasets and architectures.The review says training still commonly requires costly, time-consuming conventional hardware and identifies no breakthrough that significantly reduces training time; it also calls for comparisons of strategies across datasets and BNN variants.Source: §3.6 · p. 16

4.1. Open Source Frameworks of BNN Introductions

• Python frameworks cannot store binary data or perform bit-type operations, while BMXNet stores binarized weights and uses XNOR/popcount for matrix multiplication.BMXNet supports quantization and binarization through custom MXNet operators.Source: §4.1 · p. 16
• daBNN (Zhang et al., 2019) targets ARM devices with bit-packing, direct convolution, and memory-layout optimizations, reporting 7–23 times faster single binary convolutions than BMXNet.Its upgraded bit-packing scheme reports about 4 times the speed of the naive sequential method.Source: §4.1 · pp. 16–17
• BMXNet2 (Bethge et al., 2020) separates training and inference code and reuses MXNet operators to improve maintainability, debugging, and unit testing.It adds three C++ backend functions: sign, round with STE, and gradcancel.Source: §4.1 · p. 17
• Riptide (Fromm et al., 2020) uses TensorFlow and TVM for optimized BNN kernels and removes floating-point arithmetic from intermediate glue layers, reporting 4–12 times speed-up over floating-point implementation.Source: §4.1 · p. 17
• FINN (Blott et al., 2018) supports model development, training, conversion, and deployment on Xilinx FPGA boards through Brevitas, the FINN compiler, and PYNQ.It supports both low-bit networks and 1-bit BNNs.Source: §4.1 · p. 17
• Larq (Bannink et al., 2021) supports TensorFlow and TensorFlow Lite BNN training, conversion, and deployment through a quantization library and Larq Compute Engine.The paper reports Larq as the state-of-the-art fastest BNN framework among existing inference frameworks.Source: §4.1 · p. 17

4.2. BNN Library for CPUs and GPUs

• Other CPU/GPU efforts include Espresso (Pedersoli et al., 2017), BitFlow (Hu et al., 2018), BSTC (Li et al., 2019), O3BNN, and Phonebit.Espresso provides C/CUDA bit-type and bitwise operations; Phonebit accelerates BNN inference on Android mobile GPUs.Source: §4.2 · p. 17

4.3. Summary

• Only FINN and Larq remain actively maintained; their Xilinx-board and TensorFlow-model restrictions leave cross-platform imports and GPU execution via C++/CUDA as open needs.The paper identifies a cross-platform framework that efficiently imports BNN models from different libraries as a research opportunity; open-source platforms can also reduce low-level reimplementation for deployment on tiny devices.Source: §4.3 · p. 18

• FPGA suits BNN because its dominant computations are bitwise logic operations and its memory requirements are greatly reduced.Source: §5 · p. 18
• The section introduces resource-constrained FPGA and ARM-based SBC platforms as hardware for efficient trained-model deployment.Source: §5 · pp. 18–19
• Efficiency evaluation uses inference memory and speed, such as FPS or latency, plus FPGA DSP, LUT, BRAM, and power measurements.The section notes that BNN efficiency results on these platforms are listed in tables in the next section.Source: §5 · p. 19

6.1. Image Classification

• The review surveys BNN performance on CIFAR-10 and ILSVRC2012 ImageNet, reporting model, topology, year, and accuracy, with efficiency measures also included for ImageNet.The stated efficiency measures are BOPs, FLOPs, and OPs, where OPs = FLOPs + 1/64*BOPs.Source: §6.1 · p. 20
• ImageNet efficiency comparisons include BOPs for bitwise operations, FLOPs for floating-point operations outside BOPs, and OPs; for full-precision networks, OPs equals FLOPs.Source: §6.1 · p. 20

6.1.1. CIFAR-10 Dataset

• CIFAR-10 contains 60,000 color images across 10 classes, with 50,000 training images and 10,000 test images.Each image is 32x32 pixels; each class has 6,000 images.Source: §6.1.1 · p. 20
• The CIFAR-10 tables present BNN accuracy comparisons and FPGA-based efficiency comparisons.Source: Table 11 · p. 20, Table 12 · p. 20, Table 13 · p. 21

6.1.2. ImageNet Dataset

• ILSVRC2012 ImageNet includes more than 1.2 million training images across about 1,000 classes, plus 50,000 validation and 100,000 test images.Source: §6.1.2 · p. 22
• The ImageNet comparison tables list BNN performance results across models and topologies, alongside efficiency comparisons for FPGA implementations.The supplied results passage contains model entries, but its metric column labels are not provided here.Source: Table 15 · p. 22, Table 16 · p. 23, §6.1.2 · p. 23

6.2. Point Cloud Classification

• Binarization amplifies information loss during point-wise feature pooling and causes substantial distortion during point-wise feature extraction in 3D tasks.These challenges make 3D BNN tasks more difficult than 2D image classification.Source: §6.2 · p. 27
• BiPointNet (Qin et al., 2020) proposed the first binarization method for learning on 3D point clouds.Source: §6.2 · p. 27
• The reported evaluation used ModelNet40, with 12,311 pre-aligned shapes across 40 categories, and included accuracy and Arm-based efficiency comparisons.Table 21 presents efficiency comparisons on an Arm-based platform; Table 22 presents point-cloud classification accuracy comparisons.Source: §6.2 · p. 27, Table 21 · p. 27, Table 22 · p. 28

6.3. Object Detection

• The review reports few published BNN object-detection works, reflecting the task’s greater complexity than 2D image classification.It identifies PASCAL-VOC and MS-COCO as popular evaluation datasets.Source: §6.3 · p. 28
• LWS-Det uses student-teacher training to minimize layer-wise angular loss through differentiable binarization search and amplitude error through learned scale factors.Source: §6.3 · p. 28
• BSF-XNOR improved accuracy over XNOR-Net on VisDrone2019, and the review compiles BNN benchmark results for VOC2007 and MS-COCO minival.The review notes that FQN results cited as BNN results were actually from a 4-bit quantization network.Source: §6.3 · pp. 28–29, Table 23 · p. 29, Table 24 · p. 30
• ASDA-FRCNN reconstructs full-precision kernels from BNN-based amplitude and direction, while BiDet uses the information bottleneck to remove redundancy and focus on informative detection predictions.ASDA-FRCNN also introduces a loss function for kernel reconstruction.Source: §6.3 · pp. 28–29

6.4. Semantic Segmentation

• Group-Net applies BNNs to semantic segmentation and proposes BPAC to improve mIOU, the averaged pixel intersection-over-union measure.The evaluation uses PASCAL VOC 2012, which has 20 foreground classes and one background class.Source: §6.4 · p. 31
• The PASCAL VOC 2012 test results for semantic segmentation are presented in Table 27.The dataset contains 1,464 training, 1,449 validation, and 1,456 test images.Source: §6.4 · p. 31, Table 27 · p. 31

6.5. Natural Language Processing

• Jain et al. (2020) explored BNN for text classification, and Bai et al. (2020) first applied BNN methods to BERT for NLP tasks.Source: §6.5 · p. 31
• Qin et al. (2022) proposed BiBERT, a fully binarized BERT for NLP, while Liu et al. (2022) developed a 1–2-bit transformer quantization pipeline to reduce the performance gap from full precision.Source: §6.5 · p. 31
• GLUE provides diverse NLP tasks; Tables 28–29 cover its task corpora and the latest BNN benchmark results, with tasks including acceptability, sentiment, paraphrase, similarity, inference, and entailment.Source: §6.5 · pp. 31–32, Table 28 · p. 32, Table 29 · p. 32

6.6. Unsupervised and Semi-supervised Learning

• Unsupervised learning finds patterns in unlabeled data; self-supervised learning predicts information from inputs before later supervised tasks, while semi-supervised learning mixes labeled and unlabeled data.Source: §6.6 · p. 32
• Shen et al. (2021) proposed s2-BNN, a guided learning paradigm that distills target binary-network training from a real-valued network.Source: §6.6 · p. 32
• Kim and Choi (2022) designed BURN to jointly train full-precision and binary networks for unsupervised representation learning, with feature-similarity loss, dynamic loss balancing, and modified multi-stage training.Source: §6.6 · p. 32
• Table 30 reports ImageNet linear evaluation and semi-supervised fine-tuning after pretraining, using top-1 accuracy and 1% or 10% labels.Source: Table 30 · p. 33, §6.6 · p. 33
• Table 31 covers transfer-learning top-1 results after pretraining on object-centric datasets CIFAR10, CIFAR100, CUB-200-2011, and Birdsnap, and scene-centric Places205.Source: Table 31 · p. 33, §6.6 · p. 33

6.7. Other Tasks

• BNNs support human pose estimation and face alignment (Bulat and Tzimiropoulos, 2017), facial-mask compliance on edge devices, and low-storage image retrieval through BNNH (Zhang et al., 2021).Source: §6.7 · p. 33
• 89.2% accuracy and 384 frames per second on KTH were achieved by a 3D BNN for human-action recognition.Source: §6.7 · p. 33
• BNNs have been proposed for ECG and EEG applications toward smart autonomous healthcare devices, as well as speech recognition, (Qian and Xiang, 2019), and (Gao et al., 2021).Source: §6.7 · p. 33
• BiFSMN (2022) and its improved version BiFSMN2 (2022) extend BNN applications to keyword spotting.Source: §6.7 · p. 33
• BNN research also includes graph neural network binarization strategies and Binary DAD-Net, the first BNN-based semantic-segmentation network for driveable-area detection.Source: §6.7 · p. 33

6.8. Summary

• The review identifies architecture selection and search, accumulated and layer-varying information loss, and the scarcity of 3D-task research as unresolved challenges.It asks how to automatically find 2D or 3D BNN architectures with higher accuracy and fewer OPs. Binarized CNN layers except the input and output layers lose information; loss accumulates in deeper layers, differs across layers, and can be more severe in 3D tasks.Source: §6.8 · p. 33
• Developing transformer-based BNNs for vision is an open opportunity, motivated by ViT's patch-based encoder design (Dosovitskiy et al., 2020) and reports that transformer/convolution variants beat CNNs at equal weight size.Source: §6.8 · p. 34

7. Conclusions

• BNNs can reduce storage, network complexity, and energy consumption for resource-limited devices, but binarization causes a significant performance drop.The review notes growing research interest in BNNs since 2016 because they can enable deployment on resource-limited devices.Source: §7 · p. 34
• The review focuses mainly on networks with 1-bit activations and weights and classifies techniques across quantization, distributions, losses, gradients, topology, and training.It presents a comparative classification of techniques across these network components.Source: §7 · p. 34
• The review surveys efficient BNN platforms and application progress, identifies research gaps and available methods, and offers recommendations and future research directions.The authors anticipate that BNN will become an essential topic in computer vision and deep learning for students and professionals.Source: §7 · p. 34