Back to results
US 2022/0067491 A1Published application
Bayesian network in memory
AssigneeMicron Technology, Inc.
InventorsDmitry Vengertsev, Stewart R. Watson, Jing Gong, Ameya Parab
Life of this patent
Estimating the term…
Abstract
Apparatuses and methods can be related to implementing a Bayesian neural network in a memory. A Bayesian neural network can be implemented utilizing a resistive memory array. The memory array can comprise programmable memory cells that can be programed and used to store weights of the Bayesian neural network and perform operations consistent with the Bayesian neural network.
Claims20
An apparatus, comprising:
- a memory array;
- a controller coupled to the memory array and configured to:
- read data from a first plurality of memory cells of the memory array;
- generate, using a second plurality of memory cells of the memory array, a first plurality of weight values and a first plurality of bias values based on a plurality of deterministic values read from the first plurality of memory cells;
- generate, using a third plurality of memory cells of the memory array, a second plurality of weight values and a second plurality of bias values based on the first plurality of weight values and the second plurality of bias values;
- and
- transmit output data from a fourth plurality of memory cells of the memory array, the output data comprising a result and a confidence in the result that is based at least on:
- an input provided by a host,
- the second plurality of weight values, and
- the second plurality of bias values of a Bayesian neural network.
7 more elements
A method, comprising:
- providing a first data comprising a first partial derivative of a first loss function of a training set and a plurality of parameters to a first bank of a memory device, wherein the first loss function corresponds to a layer of a Bayesian neural network;
- utilizing the first bank to generate a second data, comprising a second partial derivative of a second loss function of a plurality of weights and the plurality of parameter, responsive to receipt of the data comprising the first partial derivative;
- utilizing the second bank to generate a third data comprising a third partial derivative of the second loss function, wherein the third data is generated using the second data;
- utilizing the second bank to generate a fourth data comprising a fourth partial derivative of the second loss function, wherein the fourth data is generated using the second data;
- and
- writing, to another portion of the memory device, updated weight values of the layer of the Bayesian neural network based at least in part on the third data comprising the third partial derivative and the fourth data comprising the fourth partial derivative.
3 more elements
A system, comprising:
- a memory device;
- a plurality of registers;
- a controller coupled to the memory device and the plurality of registers and configured to:
- access a first data representing a first partial derivative of a loss function of a Bayesian neural network from the plurality of registers;
- access a second data representing a second partial derivative of the loss function from the plurality of registers;
- access a third data representing a third partial derivative of the loss function from the plurality of registers;
- and
- update a resistance of the plurality of memory cells, to store an updated plurality of deterministic values, utilizing the first data, the second data, and the third data.
5 more elements
Dependencies are inferred from the claim text.