Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Video

Updated: Oct 5, 2025

Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit
05:30

Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit

Published on: September 8, 2023

706

Hardware-Efficient Stochastic Binary CNN Architectures for Near-Sensor Computing.

Vivek Parmar1, Bogdan Penkovsky2, Damien Querlioz2

  • 1Department of Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, India.

Frontiers in Neuroscience
|January 24, 2022
PubMed
Summary
This summary is machine-generated.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Intrinsic annealing in a hybrid memristor-magnetic tunnel junction Ising machine.

Nature communications·2026
Same author

Memristors for Bayesian in-memory computing.

Nature materials·2025
Same author

Bayesian continual learning and forgetting in neural networks.

Nature communications·2025
Same author

A ferroelectric-memristor memory for both training and inference.

Nature electronics·2025
Same author

Advanced deep learning approach for the fault severity classification of rolling-element bearings.

Scientific reports·2025
Same author

The logarithmic memristor-based Bayesian machine.

Communications engineering·2025
Same journal

CEST MRI reveals nicotine-induced alterations in glutamate-associated molecular connectivity in the mouse brain.

Frontiers in neuroscience·2026
Same journal

Brain protein burden is related to intravoxel incoherent motion: PET-MR imaging study.

Frontiers in neuroscience·2026
Same journal

Screening the optimal rTSMS frequency to orchestrate immune-fibrotic remodeling for adult spinal cord repair.

Frontiers in neuroscience·2026
Same journal

Assessment of tenecteplase target-associated pathogenic mechanisms underlying depression in acute ischemic stroke patients: insights from artificial intelligence-driven multi-omics analysis and <i>in vitro</i> validation.

Frontiers in neuroscience·2026
Same journal

Sex-divergent intrinsic brain function in Parkinson's disease: elevated nigral fluctuations and premotor-visuospatial coupling in female patients.

Frontiers in neuroscience·2026
Same journal

Spatial transcriptomics on an expanded dataset at the brain-electrode interface: exploration of variability and identification of novel biomarkers.

Frontiers in neuroscience·2026
See all related articles

This article introduces a new, energy-efficient way to run artificial intelligence vision tasks directly on small devices. By using a technique called stochastic computing, the researchers converted complex math into simple binary operations. This method allows devices to perform tasks like identifying pathogens in medical images while using significantly less power and memory than standard digital processors.

Area of Science:

  • Artificial intelligence research within Stochastic Binary CNN architectures
  • Embedded systems engineering for edge computing

Background:

Edge devices often struggle to run complex artificial intelligence models due to strict power and memory constraints. Binarized neural networks offer a potential solution by simplifying mathematical operations to reduce computational overhead. However, standard implementations frequently require high-precision first layers that complicate uniform hardware deployment. This discrepancy creates a significant hurdle for efficient on-device processing. Stochastic computing provides a mechanism to transform high-precision arithmetic into sequences of binary operations. Prior research has shown that this approach can maintain model accuracy while simplifying hardware requirements. No prior work had resolved the challenge of creating a fully binarized engine for multi-channel vision inputs. That uncertainty drove the development of the proposed hardware-friendly architecture.

Purpose Of The Study:

The primary aim of this study is to develop a hardware-friendly computation engine for vision applications. The researchers seek to address the challenges associated with deploying binarized neural networks on edge devices. Standard networks often rely on high-precision first layers, which complicate uniform hardware implementation across the entire system. This gap motivated the team to explore stochastic computing as a solution for converting high-precision math into binary operations. By creating a fully binarized engine, the authors intend to maintain model accuracy while significantly reducing power consumption. The study also investigates the feasibility of using analog hardware sources for stochastic sampling. Furthermore, the researchers evaluate the performance of this architecture through both standard datasets and medical imaging diagnostics. This work ultimately aims to provide a scalable framework for efficient on-device artificial intelligence processing.

Keywords:
RRAM (resistive RAM)binarized neural network (BNN)in-memory computing (IMC)near-sensor computingstochastic computing (SC)Artificial IntelligenceNeural NetworksMicroscopy DiagnosticsOxRAM Circuits

Frequently Asked Questions

The researchers propose a fully binarized computation engine utilizing stochastic sampling. This mechanism converts high-precision arithmetic into binary operations, allowing for efficient execution on edge hardware compared to traditional floating-point digital systems.

The authors utilize OxRAM-based circuits to perform stochastic sampling. These circuits are paired with in-memory computing modules to facilitate binarized multiplication, which is necessary for processing multi-channel inputs in vision applications.

A uniform hardware mapping is necessary because standard networks often require high-precision first layers. This requirement creates a bottleneck that prevents efficient deployment across the entire neural network architecture.

More Related Videos

Microfluidic Platform with Multiplexed Electronic Detection for Spatial Tracking of Particles
11:54

Microfluidic Platform with Multiplexed Electronic Detection for Spatial Tracking of Particles

Published on: March 13, 2017

9.4K

Related Experiment Videos

Last Updated: Oct 5, 2025

Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit
05:30

Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit

Published on: September 8, 2023

706
Microfluidic Platform with Multiplexed Electronic Detection for Spatial Tracking of Particles
11:54

Microfluidic Platform with Multiplexed Electronic Detection for Spatial Tracking of Particles

Published on: March 13, 2017

9.4K

Main Methods:

The research team designed a fully binarized computation engine tailored for vision tasks. They utilized stochastic sampling derived from analog hardware sources to approximate normal distributions. The review approach involved validating this pipeline using the CIFAR-10 dataset for initial performance benchmarks. To assess real-world utility, the authors conducted a case study focused on microscopy image diagnostics. They integrated OxRAM-based circuits to manage the sampling requirements of the system. In-memory computing modules were deployed to execute the necessary binarized multiplication operations. The experimental framework compared these results against conventional floating-precision digital implementations. This comprehensive evaluation strategy ensured the assessment of both energy efficiency and memory footprint reduction.

Main Results:

The proposed architecture achieves a 1,000-fold improvement in energy efficiency compared to standard floating-precision digital systems. Memory usage is reduced by a factor of 45 through the implementation of this binarized engine. The researchers successfully validated the pipeline using the CIFAR-10 dataset to confirm accuracy retention. Microscopy image diagnostics for pathogen detection served as a successful case study for real-world application. The system effectively utilizes OxRAM-based circuits to handle stochastic sampling tasks. In-memory computing components facilitate the required binarized multiplication without sacrificing performance. These results demonstrate that stochastic computing can replace high-precision arithmetic in edge-based vision models. The findings provide evidence that uniform hardware mapping is achievable through this specialized computation approach.

Conclusions:

The authors demonstrate that their stochastic computation engine achieves substantial energy savings compared to traditional floating-point digital systems. This architecture provides a viable path for deploying complex vision models on resource-constrained hardware. The researchers show that their approach maintains performance parity while significantly reducing memory footprints. By utilizing OxRAM-based circuits, the system effectively handles stochastic sampling for binary operations. The study confirms the utility of this method for both standard datasets and specialized medical diagnostics. These findings suggest that stochastic techniques offer a robust alternative to conventional high-precision arithmetic in edge environments. The implementation achieves a one-thousand-fold increase in energy efficiency over standard digital counterparts. This work highlights the potential for in-memory computing to support scalable and efficient artificial intelligence deployments.

The researchers employ analog hardware sources to generate non-uniform, normal distribution samples. This data type is essential for the stochastic conversion process, enabling the system to maintain accuracy while using binary logic.

The team measured energy efficiency and memory usage. They found the proposed system is 1,000 times more energy-efficient than floating-precision digital implementations and achieves a 45-fold reduction in memory requirements.

The authors propose that their engine is suitable for real-world microscopy image diagnostics. They suggest this approach enables pathogen detection on edge devices, which is otherwise difficult with conventional, power-hungry digital processors.