# US 2024/0204804 A1 · Data processing system and data processing method for deep neural network model

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2024/0204804 A1
- Document ID: US20240204804A1
- Status: Published application
- Kind: A1
- Application number: 18/522,242
- Filed: 2023-11-29
- Published: 2024-06-20
- Earliest claimed: 2022-12-16
- Assignees: Industrial Technology Research Institute
- Applicants: Industrial Technology Research Institute
- Inventors: Yu-Chih Huang; Li-Yang Tseng
- Web page: https://patentgrep.com/patent/US20240204804A1

## CPC

- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- H03M 13/27: using interleaving techniques
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- H03M 13/6577: {Representation or format of variables, register sizes or word-lengths and quantization}
- H04L 1/0041: {Arrangements at the transmitter end}
- H04L 1/0045: {Arrangements at the receiver end}
- H04L 1/0071: {Use of interleaving (interleaving per se H03M13/27)}
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- H03M: CODING; DECODING; CODE CONVERSION IN GENERAL (using fluidic means F15C4/00; optical analogue/digital converters G02F7/00; coding, decoding or code conversion, specially adapted for particular applications, see the relevant subclasses, e.g. G01D, G01R, G06F, G06T, G09G, G10L, G11B, G11C, H04B, H04L, H04M, H04N; ciphering or deciphering for cryptography or other purposes involving the need for secrecy G09C)
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

A data processing method for a DNN model includes: reading weights of transmission data; quantizing each weight into bits sequentially including first, second, third, and fourth-type bits; sequentially interleaving the first-type bit into a first bit set; sequentially interleaving each second-type bit into second bit sets and reading a second compression rate of each second bit set in response to the compressible second bit sets; interleaving the third-type bit into a third bit set and reading a third compression rate of the third bit set in response to the compressible third bit set; compressing each second bit set with the second compression rate, and compressing the third bit set with the third compression rate; sequentially coding the first bit set, each compressed second bit set, and the compressed third bit set to generate first encoded data corresponding to the transmission data; transmitting the first encoded data to an external device.

## Claims

### Claim 1

A data processing system, comprising:
a memory, configured to store a deep neural network model; and
a processor, coupled to the memory and configured to perform following steps:
reading a plurality of weights from a transmission data of the deep neural network model;
quantizing each of the weights into a plurality of bits, wherein the bits sequentially comprise a first-type bit, a plurality of second-type bits, a third-type bit, and a plurality of fourth-type bits;
interleaving the first-type bit in each of the weights into a first bit set;
sequentially interleaving each of the second-type bits in each of the weights into a plurality of second bit sets, and reading a second compression rate of each of the second bit sets in response to the second bit sets being compressible;
interleaving the third-type bit in each of the weights into a third bit set, and reading a third compression rate of the third bit set in response to the third bit set being compressible;
compressing each of the second bit sets with the second compression rate, and compressing the third bit set with the third compression rate;
sequentially coding the first bit set, each of the compressed second bit sets, and the compressed third bit set to generate a first encoded data corresponding to the transmission data; and
transmitting the first encoded data to an external device via a network.

### Claim 2 (depends on 1)

The data processing system as claimed in claim 1, wherein the processor is further configured to:
receive a second encoded data from the external device;
decode the second encoded data to generate the first bit set, the compressed second bit sets, and the compressed third bit set;
decompress each of the compressed second bit sets and the compressed third bit set;
sequentially de-interleave the first bit set, each of the second bit sets, and the third bit set; and
reconstruct a plurality of de-interleaved bits generated after the first bit set, each of the second bit sets, and the third bit set are de-interleaved as a plurality of reconstructed weights.

### Claim 3 (depends on 1)

The data processing system as claimed in claim 1, wherein the first-type bit is a most significant bit.

### Claim 4 (depends on 1)

The data processing system as claimed in claim 1, wherein the processor is further configured to:
quantize each of the weights into the bits based on an IEEE standard 754 for binary floating-point arithmetic.

### Claim 5 (depends on 4)

The data processing system as claimed in claim 4, wherein the first-type bit is a sign bit.

### Claim 6 (depends on 4)

The data processing system as claimed in claim 4, wherein the second-type bits and the third-type bit are exponent bias bits.

### Claim 7 (depends on 1)

The data processing system as claimed in claim 1, wherein the memory is further configured to store a compression rate data, and the processor is further configured to:
dynamically determine whether the second bit sets and the third bit set are compressible based on the compression rate data and a distribution condition of 0s and 1s in each of the second bit sets and the third bit set, and read the second compression rate and the third compression rate, wherein the compression rate data comprise historical statistics.

### Claim 8

A data processing method for a deep neural network model, comprising:
reading a plurality of weights from a transmission data;
quantizing each of the weights into a plurality of bits, wherein the bits sequentially comprise a first-type bit, a plurality of second-type bits, a third-type bit, and a plurality of fourth-type bits;
interleaving the first-type bit in each of the weights into a first bit set;
sequentially interleaving each of the second-type bits in each of the weights into a plurality of second bit sets, and reading a second compression rate of each of the second bit sets in response to the second bit sets being compressible;
interleaving the third-type bit in each of the weights into a third bit set, and reading a third compression rate of the third bit set in response to the third bit set being compressible;
compressing each of the second bit sets with the second compression rate, and compressing the third bit set with the third compression rate;
sequentially coding the first bit set, each of the compressed second bit sets, and the compressed third bit set to generate a first encoded data corresponding to the transmission data; and
transmitting the first encoded data to an external device.

### Claim 9 (depends on 8)

The data processing method as claimed in claim 8, further comprising:
receiving a second encoded data from the external device;
decoding the second encoded data to generate the first bit set, the compressed second bit sets, and the compressed third bit set;
decompressing each of the compressed second bit sets and the compressed third bit set;
sequentially de-interleaving the first bit set, each of the second bit sets, and the third bit set; and
reconstructing a plurality of de-interleaved bits generated after the first bit set, each of the second bit sets, and the third bit set are de-interleaved as a plurality of reconstructed weights.

### Claim 10 (depends on 8)

The data processing method as claimed in claim 8, wherein the first-type bit is a most significant bit.

### Claim 11 (depends on 8)

The data processing method as claimed in claim 8, further comprising:
quantizing each of the weights into the bits based on an IEEE standard 754 for binary floating-point arithmetic.

### Claim 12 (depends on 11)

The data processing method as claimed in claim 11, wherein the first-type bit is a sign bit.

### Claim 13 (depends on 11)

The data processing method as claimed in claim 11, wherein the second-type bits and the third-type bits are exponent bias bits.

### Claim 14 (depends on 8)

The data processing method as claimed in claim 8, further comprising:
dynamically determining whether the second bit sets and the third bit set are compressible based on a compression rate data and a distribution condition of 0s and 1s in each of the second bit sets and the third bit set, and reading the second compression rate and the third compression rate, wherein the compression rate data comprise historical statistics.
