Back to results
US 11,886,999 B2Granted
Artificial neural network training in memory
AssigneeMicron Technology, Inc.
InventorsSaideep Tiku, Poorna Kale
Life of this patent
Estimating the term…
Abstract
Apparatuses and methods can be related to implementing age-based network training. An artificial neural network (ANN) can be trained by introducing errors into the ANN. The errors and the quantity of errors introduced into the ANN can be based on age-based characteristics of the memory device.
Claims19
A method, comprising:
- accessing, from a memory array of a memory device, a bit string of data representing elements of an artificial neural network model, the elements comprising a plurality of weights and a plurality of biases;
- identifying a type of the memory device that implements the artificial neural network model;
- responsive to identifying the type of the memory device, introducing logical values into the bit string to modify the plurality of weights and the plurality of biases, the logical values representative of a number of expected errors that correspond to the memory device;
- and implementing the artificial neural network model in the memory device based at least in part on writing the introduced logic values of the bit string to the memory array.
1 more element
An apparatus, comprising:
- a memory array;
- a controller coupled to the memory array and configured to:
- access a bit string of data representing elements of an artificial neural network model, the elements comprising a plurality of nodes and a plurality of layers of the artificial neural network model stored in the memory array;
- modify the plurality of nodes and the plurality of layers of the artificial neural network model by introducing logical values in the bit string, the logical values representing a number of expected errors of the memory array and by adding a number of nodes to the plurality of nodes of the artificial neural network model;
- and train the artificial neural network model including the logical values.
2 more elements
A non-transitory computer-readable medium having computer-readable instructions, which when executed by a computer, cause the computer to:
- determine whether an error correction code (ECC) is utilized;
- responsive to determining that the ECC is utilized, modify a bit string of data representing elements of an artificial neural network model implemented at an edge of a memory array by adding logical values to the bit string based on characteristics of the ECC;
- and
- implement the artificial neural network model at the edge of the memory array.
1 more element
A method, comprising:
- accessing, from a memory array of a memory device, a bit string of data representing elements of an artificial neural network model, the elements comprising a plurality of weights and a plurality of biases;
- introducing logical values into the bit string to modify the plurality of weights and the plurality of biases, the logical values representative of a number of expected errors that correspond to the memory device and wherein introducing the logical values includes introducing a fixed quantity of the number of errors regardless of an expected bit-error rate of the memory device or a type of the memory device;
- and
- implementing the artificial neural network model in the memory device based at least in part on writing the introduced logic values of the bit string to the memory array.
1 more element
An apparatus, comprising:
- a memory array;
- a controller coupled to the memory array and configured to:
- access a bit string of data representing elements of an artificial neural network model, the elements comprising a plurality of nodes and a plurality of layers of the artificial neural network model stored in the memory array;
- determine a simulated age of the memory array;
- modify the plurality of nodes and the plurality of layers of the artificial neural network model by introducing logical values in the bit string based on the simulated age of the memory array, the logical values representing a number of expected errors of the memory array;
- and train the artificial neural network model including the logical values.
3 more elements
Dependencies are inferred from the claim text.