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 Concept Videos

You might also read

Related Articles

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

Sort by
Same author

LRMP: Layer Replication with Mixed Precision for spatial in-memory DNN accelerators.

Frontiers in artificial intelligence·2024
Same author

Fast and robust analog in-memory deep neural network training.

Nature communications·2024
Same author

Compute in-Memory with Non-Volatile Elements for Neural Networks: A Review from a Co-Design Perspective.

Advanced materials (Deerfield Beach, Fla.)·2022
Same author

Neural Network Training With Asymmetric Crosspoint Elements.

Frontiers in artificial intelligence·2022
Same author

Impact of Asymmetric Weight Update on Neural Network Training With Tiki-Taka Algorithm.

Frontiers in neuroscience·2022
Same author

Enabling Training of Neural Networks on Noisy Hardware.

Frontiers in artificial intelligence·2021
Same journal

Synaptic micromechanics and brain softening as a mechanobiological hypothesis for Alzheimer's disease.

Frontiers in neuroscience·2026
Same journal

The relationship between healthy sleep patterns and the risk of scoliosis: a large prospective cohort study.

Frontiers in neuroscience·2026
Same journal

Dynamic functional reorganization in post-stroke aphasia: a state-of-the-art fMRI review from disease evolution to intervention.

Frontiers in neuroscience·2026
Same journal

Correction: Case Report: A possible novel adult-onset, progressive MAO-A hypofunction.

Frontiers in neuroscience·2026
Same journal

Respiratory modulation of neurophysiology and symptoms in athletes with sports-related concussion: a randomized crossover trial.

Frontiers in neuroscience·2026
Same journal

Impact of C-reactive protein-triglyceride-glucose and systemic immune-inflammation indices on obstructive sleep apnea in older adults with depression.

Frontiers in neuroscience·2026
See all related articles

Related Experiment Video

Updated: Dec 26, 2025

Time-dependent Increase in the Network Response to the Stimulation of Neuronal Cell Cultures on Micro-electrode Arrays
10:45

Time-dependent Increase in the Network Response to the Stimulation of Neuronal Cell Cultures on Micro-electrode Arrays

Published on: May 29, 2017

10.2K

Algorithm for Training Neural Networks on Resistive Device Arrays.

Tayfun Gokmen1, Wilfried Haensch1

  • 1IBM Research AI, Yorktown Heights, NY, United States.

Frontiers in Neuroscience
|March 17, 2020
PubMed
Summary
This summary is machine-generated.

A new "Tiki-Taka" algorithm enables accurate deep neural network training on analog hardware, even with non-symmetric device switching. This approach maintains power and speed benefits, relaxing material requirements for resistive crossbar arrays.

Keywords:
analog hardware acceleratorcrossbar arraydeep learningmemristorresistive deviceresistive processing unittraining algorithms

More Related Videos

Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks
11:18

Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks

Published on: March 2, 2015

10.7K
Preparation of Neuronal Co-cultures with Single Cell Precision
09:06

Preparation of Neuronal Co-cultures with Single Cell Precision

Published on: May 20, 2014

14.2K

Related Experiment Videos

Last Updated: Dec 26, 2025

Time-dependent Increase in the Network Response to the Stimulation of Neuronal Cell Cultures on Micro-electrode Arrays
10:45

Time-dependent Increase in the Network Response to the Stimulation of Neuronal Cell Cultures on Micro-electrode Arrays

Published on: May 29, 2017

10.2K
Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks
11:18

Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks

Published on: March 2, 2015

10.7K
Preparation of Neuronal Co-cultures with Single Cell Precision
09:06

Preparation of Neuronal Co-cultures with Single Cell Precision

Published on: May 20, 2014

14.2K

Area of Science:

  • Neuromorphic engineering
  • Materials science
  • Computer science

Background:

  • Resistive cross-point device arrays offer power and speed advantages for deep neural network (DNN) training.
  • Current training methods like stochastic gradient descent (SGD) require symmetrical switching characteristics in resistive devices, a stringent material requirement.
  • Device asymmetry can lead to unintentional cost terms, impacting DNN training accuracy on analog hardware.

Purpose of the Study:

  • To introduce a novel training algorithm, "Tiki-Taka", that overcomes the symmetry requirement for resistive devices in DNN training.
  • To demonstrate that the "Tiki-Taka" algorithm can achieve comparable accuracy to conventional SGD even with non-ideal, asymmetric device switching.
  • To maintain the power and speed benefits of analog hardware while relaxing material specifications for resistive crossbar arrays.

Main Methods:

  • Developed the "Tiki-Taka" algorithm, a coupled dynamical system that simultaneously minimizes the DNN objective function and the cost term arising from device asymmetry.
  • Tested the algorithm on various network architectures, including fully connected, convolutional, and LSTM networks.
  • Compared simulation results of "Tiki-Taka" with asymmetric devices against conventional SGD with symmetric devices.

Main Results:

  • The "Tiki-Taka" algorithm achieved accuracy comparable to conventional SGD using ideal, symmetric devices, but with non-ideal, asymmetric devices.
  • Device asymmetry was shown to introduce an implicit cost term in SGD, which "Tiki-Taka" effectively manages.
  • All operations remain parallel, preserving the power and speed advantages of array architectures.

Conclusions:

  • The "Tiki-Taka" algorithm successfully eliminates the stringent symmetry requirement for resistive devices in DNN training.
  • This algorithmic advancement is crucial for realizing technologically viable resistive crossbar arrays that can outperform digital accelerators.
  • Relaxed material specifications will accelerate the adoption of analog hardware for efficient AI computations.