Extending the error correction capability of a device using a neural network
View Patent ↗A codeword decoder interoperates with a neural network to provide the neural network with an uncorrectable erroneous codeword and to classify the uncorrectable erroneous codeword into a correctable erroneous codeword. The codeword decoder and neural network may be utilized in the error correction unit for a memory array.
1. An apparatus comprising:
a codeword decoder; and
a neural network coupled to the codeword decoder, the neural network to receive an uncorrectable erroneous codeword from the codeword decoder and to classify the uncorrectable erroneous codeword into a correctable erroneous codeword;
wherein the correctable erroneous codeword comprises a reduced bit error count relative to the uncorrectable erroneous codeword.
2. The apparatus of claim 1 , the neural network comprising a plurality of fully-connected hidden layers.
3. The apparatus of claim 1 , all layers of the neural network being fully-connected layers.
4. An apparatus comprising:
a codeword decoder; and
a neural network coupled to the codeword decoder, the neural network to receive an uncorrectable erroneous codeword from the codeword decoder and to classify the uncorrectable erroneous codeword into a correctable erroneous codeword,
the neural network further to provide an error signal to the codeword decoder that the uncorrectable erroneous codeword is uncorrectable.
5. The apparatus of claim 4 , the codeword decoder to output a read error for the uncorrectable erroneous codeword in response to receipt of the error signal.
6. An apparatus comprising:
a codeword decoder; and
a neural network coupled to the codeword decoder, the neural network to receive an uncorrectable erroneous codeword from the codeword decoder and to classify the uncorrectable erroneous codeword into a correctable erroneous codeword,
the neural network further receiving a runtime bit error rate and memory array health parameters for a memory array.
7. The apparatus of claim 6 , the neural network comprising an input layer comprising a first neuron to receive the uncorrectable erroneous codeword, a second neuron to receive the runtime bit error rate, and a third neuron to receive the memory array health parameters.
8. An apparatus comprising:
a codeword decoder; and
a neural network coupled to the codeword decoder, the neural network to receive an uncorrectable erroneous codeword from the codeword decoder and to classify the uncorrectable erroneous codeword into a correctable erroneous codeword,
the neural network configured by applying results of a test on a memory array, the results of the test comprising:
a plurality of noisy codewords read from the memory array during the test;
an average bit error rate for the test; and
memory array health parameters determined by the test.
9. An apparatus comprising:
a codeword decoder; and
a neural network coupled to the codeword decoder, the neural network to receive an uncorrectable erroneous codeword from the codeword decoder and to classify the uncorrectable erroneous codeword into a correctable erroneous codeword, the neural network comprising a plurality of hidden layers, each comprising exactly N neurons where N is a bit length of codewords of a memory array.
10. The apparatus of claim 9 , the neural network comprising a Softmax output layer comprising N neurons.
11. The apparatus of claim 9 , the plurality of hidden layers comprising between 2 and N/2 hidden layers, inclusive.
12. An error correction unit for a memory system, the error correction unit comprising:
a codeword decoder interoperable with a neural network, the codeword decoder responsive to a received codeword determined to be an uncorrectable erroneous codeword by operating the neural network to convert the uncorrectable erroneous codeword into a returned correctable erroneous codeword having a reduced bit error count relative to the uncorrectable erroneous codeword; and
the codeword decoder operating independently of the neural network when the received codeword is determined to be a received correctable erroneous codeword.
13. The error correction unit of claim 12 , the codeword decoder responding to the returned correctable erroneous codeword by applying an error correcting code for the codeword to generate a non-erroneous codeword.
14. The error correction unit of claim 12 , the codeword decoder responding with a read error signal as a result of receiving an error signal from the neural network indicating that the uncorrectable erroneous codeword is uncorrectable.
15. The error correction unit of claim 12 , the neural network comprising a plurality of fully-connected hidden layers.
16. The error correction unit of claim 12 , all layers of the neural network being fully-connected layers.
17. The error correction unit of claim 12 , the neural network comprising an input layer comprising a first neuron to receive the uncorrectable erroneous codeword, a second neuron to receive an average bit error rate for a memory array, and a third neuron to receive memory array health parameters for the memory array.
18. The error correction unit of claim 12 , the neural network comprising:
a plurality of fully-connected hidden layers each comprising at least N neurons where N is a bit length of codewords of a memory array.
19. The error correction unit of claim 18 , the neural network comprising a gradient descent backpropagation learning function.