International Business Machines Corporation

International Business Machines Cororation

Stefano Ambrogio

Pritish Narayanan

HsinYu Tsai

Geoffrey W Burr

Charles Mackin

Wilfried Haensch

Tayfun Gokmen

Pierce I-Jen Chuang

Sanjay Kariyappa

Mathieu Gallot

A system includes a processor, and a resistive processing resistive processing unit coupled to the processor. The resistive processing unit includes an array of cells, wherein the cells respectively include resistive devices, wherein at least a portion of the resistive devices are programmable to store weight values of a given matrix in the array of cells. When the given matrix is stored in the array of cells, the processor is configured to perform a weight extraction process. The weight extraction process applies a set of input vectors to the resistive processing unit to perform analog matrix-vector multiplication operations on the stored matrix, obtains a set of output vectors resulting from the analog matrix-vector multiplication operations, and determines weight values of the given matrix stored in the array of cells utilizing the set of input vectors and the set of output vectors.

Extraction of weight values in resistive processing unit array

A method comprises receiving an input signal for processing in one or more neurons of a neural network, wherein the neural network has zero bias neurons and includes a plurality of resistive processing unit (RPU) weights and each neuron has an activation function. The method also includes applying an arbitrary amplification factor to activation function outputs of the one or more neurons in the neural network, wherein the arbitrary amplification factor is based on a dynamic range of components in the neural network and compensates for conductance drift in values of the RPU weights. The method also includes performing a calculation with the neural network using the amplified activation function outputs of the one or more neurons.

Neural network accelerators resilient to conductance drift

The method includes setting conductances for corresponding non-volatile memory (NVM) devices of a cross-bar array to zero. The method further includes determining a plurality of pulse-widths for the corresponding plurality of NVM devices based on a corresponding plurality of programming errors. Additionally, the method includes programming the NVM devices using the determined pulse-widths. Also, the method includes measuring actual conductances for the corresponding NVM devices. Further, the method includes adjusting scaling factors for the corresponding NVM devices based on the actual conductances and the corresponding programming errors. Additionally, the method includes programming the corresponding NVM devices based on the determined pulse-widths and the scaling factors.

Programming devices and weights in hardware

Methods and systems for performing calculations with a neural network include determining a conductance drift coefficient for resistive processing unit (RPU) weights in a neural network. A correction factor is applied to neuron inputs in the neural network in accordance with the drift coefficient and a time that has elapsed since the RPU weights were programmed. A calculation is performed with the neural network. The correction factor compensates for conductance drift.

Mitigation of conductance drift in neural network resistive processing units

A computer implemented method includes updating weight values associated with a plurality of analog synapses in a cross-bar array that implements an artificial neural network by sending a pulse sequence to the analog synapses. Each analog synapse includes a conductance unit, wherein a weight value of the analog synapse is based on a conductance value of the conductance unit. The pulse sequence changes the conductance value. The method further includes comparing the weight values of the analog synapses with target weight values associated with the analog synapses and selecting a set of analog synapses based on the comparison. The method further includes updating the weight values of the selected analog synapses by sending a set of electric pulses of varying durations.

Growing community of inventors

San Jose, CA, United States of America

Stefano Ambrogio