We propose a dedicated winner-take-all circuit to efficiently implement the intra-column competition between cells in Hierarchical Temporal Memory which is a crucial part of the algorithm. All inputs and outputs are charge-based for compatibility wit
h standard CMOS. The circuit incorporates memristors for competitive advantage to emulate a column with a cell in a predictive state. The circuit can also detect columns bursting by passive averaging and comparison of the cell outputs. The proposed spintronic devices and circuit are thoroughly described and a series of simulations are used to predict the performance. The simulations indicate that the circuit can complete a nine-cell, nine-input competition operation in under 15 ns at a cost of about 25 pJ.
We report the performance characteristics of a notional Convolutional Neural Network based on the previously-proposed Multiply-Accumulate-Activate-Pool set, an MTJ-based spintronic circuit made to compute multiple neural functionalities in parallel.
A study of image classification with the MNIST handwritten digits dataset using this network is provided via simulation. The effect of changing the weight representation precision, the severity of device process variation within the MAAP sets and the computational redundancy are provided. The emulated network achieves between 90 and 95% image classification accuracy at a cost of ~100 nJ per image.
We propose a new network architecture for standard spin-Hall magnetic tunnel junction-based spintronic neurons that allows them to compute multiple critical convolutional neural network functionalities simultaneously and in parallel, saving space and
time. An approximation to the Rectified Linear Unit transfer function and the local pooling function are computed simultaneously with the convolution operation itself. A proof-of-concept simulation is performed on the MNIST dataset, achieving up to 98% accuracy at a cost of less than 1 nJ for all convolution, activation and pooling operations combined. The simulations are remarkably robust to thermal noise, performing well even with very small magnetic layers.
An in situ measurement of spin transport in a graphene nonlocal spin valve is used to quantify the spin current absorbed by a small (250 nm $times$ 750 nm) metallic island. The experiment allows for successive depositions of either Fe or Cu without b
reaking vacuum, so that the thickness of the island is the only parameter that is varied. Furthermore, by measuring the effect of the island using separate contacts for injection and detection, we isolate the effect of spin absorption from any change in the spin injection and detection mechanisms. As inferred from the thickness dependence, the effective spin current $j_e = frac{2e}{hbar} j_s$ absorbed by Fe is as large as $10^8$ A/m$^2$. The maximum value of $j_e$ is limited by the resistance-area product of the graphene/Fe interface, which is as small as 3 $Omegamu$m$^2$. The spin current absorbed by the same thickness of Cu is smaller than for Fe, as expected given the longer spin diffusion length and larger spin resistance of Cu compared to Fe. These results allow for a quantitative assessment of the prospects for achieving spin transfer torque switching of a nanomagnet using a graphene-based nonlocal spin valve.
