Results 41 to 50 of about 27,097 (268)

Research on three-step accelerated gradient algorithm in deep learning

Statistical Theory and Related Fields, 2022
Gradient descent (GD) algorithm is the widely used optimisation method in training machine learning and deep learning models. In this paper, based on GD, Polyak's momentum (PM), and Nesterov accelerated gradient (NAG), we give the convergence of the ...
Yongqiang Lian, Yincai Tang, Shirong Zhou   +2 more
doaj   +1 more source

All‐in‐One Analog AI Hardware: On‐Chip Training and Inference with Conductive‐Metal‐Oxide/HfOx ReRAM Devices

Advanced Functional Materials, EarlyView.
An all‐in‐one analog AI accelerator is presented, enabling on‐chip training, weight retention, and long‐term inference acceleration. It leverages a BEOL‐integrated CMO/HfOx ReRAM array with low‐voltage operation (<1.5 V), multi‐bit capability over 32 states, low programming noise (10 nS), and near‐ideal weight transfer.
Donato Francesco Falcone, Victoria Clerico, Wooseok Choi, Tommaso Stecconi, Folkert Horst, Laura Bégon‐Lours, Matteo Galetta, Antonio La Porta, Nikhil Garg, Fabien Alibart, Bert Jan Offrein, Valeria Bragaglia   +11 more
wiley   +1 more source

Backward Signal Propagation: A Symmetry-Based Training Method for Neural Networks

Algorithms
While backpropagation (BP) has long served as the cornerstone of training deep neural networks, it relies heavily on strict differentiation logic and global gradient information, lacking biological plausibility. In this paper, we systematically present a
Kun Jiang, Zhihong Fu
doaj   +1 more source

Monolithic Oxidation Enables Ultrathin Vertically Graded Tantalum Oxide for Low‐Voltage, Low‐Variability Memristive Switching

Advanced Functional Materials, EarlyView.
Monolithic UV‐ozone oxidation of Ta forms an ultrathin Ta2O5/TaOx bilayer enabling resistive switching with a vertical defect gradient. A stoichiometric surface layer over an oxygen‐deficient sublayer promotes localized filament nucleation near the top interface, enabling low‐voltage operation, and reduced cycle‐to‐cycle variability.
Seunghoon Yang, Dong Hyun Lee, Conan Weiland, Won‐Il Lee, Kim Kisslinger, Mingzhao Liu, Aswin kumar Anbalagan, Xiao Tong, Soumyajit Mandal, Nikhil Tiwale, Houk Jang, Chang‐Yong Nam   +11 more
wiley   +1 more source

Towards New Generation, Biologically Plausible Deep Neural Network Learning

Sci, 2022
Artificial neural networks in their various different forms convincingly dominate machine learning of the present day. Nevertheless, the manner in which these networks are trained, in particular by using end-to-end backpropagation, presents a major ...
Anirudh Apparaju, Ognjen Arandjelović
doaj   +1 more source

Neuromorphic Electronics for Intelligence Everywhere: Emerging Devices, Flexible Platforms, and Scalable System Architectures

Advanced Materials, EarlyView.
The perspective presents an integrated view of neuromorphic technologies, from device physics to real‐time applicability, while highlighting the necessity of full‐stack co‐optimization. By outlining practical hardware‐level strategies to exploit device behavior and mitigate non‐idealities, it shows pathways for building efficient, scalable, and ...
Kapil Bhardwaj, Roshni Satheesh Babu, Yuxin Xia, Eva Bestelink, Radu Sporea, Nikos Hastas, Ioannis Zeimpekis, Dimitra G. Georgiadou, Sayani Majumdar   +8 more
wiley   +1 more source

Memristive Motion‐Streak Neuron for Spatiotemporal Multiple Object Detection

Advanced Materials, EarlyView.
This work proposes motion‐streak neuron hardware using amorphous In–Ga–Zn–O optoelectronic memristors and dynamic memristors for spatiotemporal multiple‐object detection. The motion streak neuron array processes optical and electrical stimuli, allowing speed‐dependent filament relaxation to encode motion trajectories.
Hyungjun Park, Jin Hong Kim, Hyun Wook Kim, Néstor Ghenzi, Sung Keun Shim, Joon‐Kyu Han, Min Jung Chung, Dong Hoon Shin, Kyung Seok Woo, Cheol Seong Hwang   +9 more
wiley   +1 more source

Backpropagation-Friendly Eigendecomposition

CoRR, 2019
Eigendecomposition (ED) is widely used in deep networks. However, the backpropagation of its results tends to be numerically unstable, whether using ED directly or approximating it with the Power Iteration method, particularly when dealing with large matrices.
Wang, Wei, Dang, Zheng, Hu, Yinlin, Fua, Pascal, Salzmann, Mathieu   +4 more
openaire   +4 more sources

Inverse Design of Nanoparticulate Materials

Advanced Materials, EarlyView.
Inverse design shifts nanomaterial development from empirical trial‐and‐error to predictive model‐driven strategies. It can rely on knowledge‐based, data‐based, or hybrid process and property functions. This perspective article provides a practical framework for applying inverse design based on instructive examples. It discusses which modeling approach
Nabi Etienne Traoré, Annika Mauch, Michelle Berthold, Lukas Pflug, Wolfgang Peukert, Nicolas Vogel   +5 more
wiley   +1 more source

Recunoașterea unei cifre scrise de mână folosind o rețea neuronală convoluțională și biblioteca TensorFlow unei cifre scrise de mână folosind o rețea neuronală convoluțională și biblioteca TensorFlow [PDF]

Revista Română de Informatică și Automatică, 2019
In this paper it is proposed to solve a visual problem of recognizing a handwritten figure. A machine learning technique will be used in which a result is produced based on previous experience.
Paul TEODORESCU
doaj   +1 more source