Academia.eduAcademia.edu

Figure 3 – uploaded by Alex Serb

See full PDFdownloadDownload figure
Figure 3. Image recognition task with TiOx memristor-based neural network simulation in NeuroPack (A) Binary handwritten digits from MNIST dataset cropped to 22 x 22. Dark blue and yellow represent weights 0 and | respectively. (B) - (D) Conductance sets before training (B), after 2000 epochs (C) and after 10000 epochs (D). Conductance spans from 0.0000869S to 0.0004416S, indicating final memristor RSs ranging from around 2264 ohms to 11507 ohms. (E) Concept diagram of spiking neural network used in this image recognition task. In practice the network is fully connected. 22 x 22 -pixel images are unrolled to 484-bit input sending to 484 input neurons correspondingly as input spikes. In this task, a two-layer network with 484 input neurons and 10 output neurons is applied. (F) Accuracy curves over the training process for both memristor version and non-memristor version. Minibatch size of 100 are used to calculate the accuracy changes during the training. Provided 2000 images from a separated test set memristor version and non-memristor version achieved accuracy of 82.00% and 83.55% respectively.

Figure 3 Image recognition task with TiOx memristor-based neural network simulation in NeuroPack (A) Binary handwritten digits from MNIST dataset cropped to 22 x 22. Dark blue and yellow represent weights 0 and | respectively. (B) - (D) Conductance sets before training (B), after 2000 epochs (C) and after 10000 epochs (D). Conductance spans from 0.0000869S to 0.0004416S, indicating final memristor RSs ranging from around 2264 ohms to 11507 ohms. (E) Concept diagram of spiking neural network used in this image recognition task. In practice the network is fully connected. 22 x 22 -pixel images are unrolled to 484-bit input sending to 484 input neurons correspondingly as input spikes. In this task, a two-layer network with 484 input neurons and 10 output neurons is applied. (F) Accuracy curves over the training process for both memristor version and non-memristor version. Minibatch size of 100 are used to calculate the accuracy changes during the training. Provided 2000 images from a separated test set memristor version and non-memristor version achieved accuracy of 82.00% and 83.55% respectively.

Related Figures (5)

Figure 1. NeuroPack workflow. The system reads three configuration files: connectivity matrix, (neural) stimuli and config. The memristor model is embedded within the virtual memristor array, which initialises a number of memristor devices according to instructions in the configuration and connectivity matrix files. Neuron models and learning rules are placed in the core file. In ’updating fire history” stage, there are three separate steps: calculating fire states assuming neurons fire ’freely’; adding network-level constraints; and updating the fire history matrix. *Calculating AW” is supported by most learning rules except STDP. By replacing the core file, users can apply different neuron models and learning rules.
Figure 1. NeuroPack workflow. The system reads three configuration files: connectivity matrix, (neural) stimuli and config. The memristor model is embedded within the virtual memristor array, which initialises a number of memristor devices according to instructions in the configuration and connectivity matrix files. Neuron models and learning rules are placed in the core file. In ’updating fire history” stage, there are three separate steps: calculating fire states assuming neurons fire ’freely’; adding network-level constraints; and updating the fire history matrix. *Calculating AW” is supported by most learning rules except STDP. By replacing the core file, users can apply different neuron models and learning rules.
in the pulse parameter selection module. After attaining the set of parameters that will lead to closest resistance a pulse is sent and the new resistance is read. This predict-write-verify process is repeated until Rnew—Rex ecte: . : : the calculated “S$ —“*veted ig smaller than the R-tolerance or the maximum step number is reached. Rexpected
in the pulse parameter selection module. After attaining the set of parameters that will lead to closest resistance a pulse is sent and the new resistance is read. This predict-write-verify process is repeated until Rnew—Rex ecte: . : : the calculated “S$ —“*veted ig smaller than the R-tolerance or the maximum step number is reached. Rexpected
Figure 4. Classification accuracy affected by R tolerance. (A) and (B) Accuracy in training and test phases with different R tolerance respectively. (C) Resistance changes in a memristor representing a stimulated synapse (between input neuron 250 and output neuron 6) in training process with different R tolerance. Notice that we use virtual resistance values for the baseline. (D)-(H) Memristor resistive states before training (D), after 1000 epochs for R tolerance of 2%(E) or 0.1%(F), and after 10000 epochs for R tolerance of 2%(G) or 0.1%(H). The total range of memristor RSs is from 2.26kQ to 11.5kQ, which gives the range of weights as 0.98 to 0.086 correspondingly.
Figure 4. Classification accuracy affected by R tolerance. (A) and (B) Accuracy in training and test phases with different R tolerance respectively. (C) Resistance changes in a memristor representing a stimulated synapse (between input neuron 250 and output neuron 6) in training process with different R tolerance. Notice that we use virtual resistance values for the baseline. (D)-(H) Memristor resistive states before training (D), after 1000 epochs for R tolerance of 2%(E) or 0.1%(F), and after 10000 epochs for R tolerance of 2%(G) or 0.1%(H). The total range of memristor RSs is from 2.26kQ to 11.5kQ, which gives the range of weights as 0.98 to 0.086 correspondingly.
Figure 5. Classification with versions using selector-based and selectorless memristor arrays. (A) Accuracy in selector-based and selectorless arrays. (B) Resistance changes in memristors representing a stimulated (between input neuron 250 and output 6) and a non-stimulated (between input neuron 10 and output neuron 6) synapses. (C) - (E) give memristor resistive states before training (C), after training for both selector-based (D) and selectorless (E) versions. The color bar shows the memristor states from 2.26k-25kOmega, which give weights in range 0.98 to 0.01 correspondingly.
Figure 5. Classification with versions using selector-based and selectorless memristor arrays. (A) Accuracy in selector-based and selectorless arrays. (B) Resistance changes in memristors representing a stimulated (between input neuron 250 and output 6) and a non-stimulated (between input neuron 10 and output neuron 6) synapses. (C) - (E) give memristor resistive states before training (C), after training for both selector-based (D) and selectorless (E) versions. The color bar shows the memristor states from 2.26k-25kOmega, which give weights in range 0.98 to 0.01 correspondingly.
Table 1. Parameters used in NeuroPack for MNIST handwritten digit classification task. oo _ arrays it may be possible to decode useful information from the state of the system after training. These investigations illustrate the role NeuroPack can play in assisting users to design and validate neuro- inspired concepts and improve system performance involving emerging nanoscale memory technologies. We envisage that by varying datasets, biasing and other experimental parameters, device technology and connectivity patterns, users from across the community will be able to use this tool to generate the results that suit their needs quickly and efficiently.
Table 1. Parameters used in NeuroPack for MNIST handwritten digit classification task. oo _ arrays it may be possible to decode useful information from the state of the system after training. These investigations illustrate the role NeuroPack can play in assisting users to design and validate neuro- inspired concepts and improve system performance involving emerging nanoscale memory technologies. We envisage that by varying datasets, biasing and other experimental parameters, device technology and connectivity patterns, users from across the community will be able to use this tool to generate the results that suit their needs quickly and efficiently.
Related topics:
EngineeringComputer ScienceComputer ArchitectureArtificial IntelligenceNeuromorphic Engineeringpython programming language
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved