# US 12,602,587 B2 · Multi-task deep learning network and generation method thereof

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

- Number: US 12,602,587 B2
- Document ID: US12602587B2
- Status: Granted
- Kind: B2
- Application number: 17/955,567
- Filed: 2022-09-29
- Published: 2023-04-13
- Granted: 2026-04-14
- Earliest claimed: 2021-10-08
- Assignees: ALi Corporation
- Applicants: ALi Corporation
- Inventors: Jou-Yun Pan; Keng-Chih Chen
- Web page: https://patentgrep.com/patent/US12602587B2

## CPC

- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/045: Combinations of networks
- G06N 3/063: using electronic means
- G06N 3/096: Transfer learning
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- Y02D: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE

## Abstract

A multi-task deep learning network and a generation method thereof are provided. The generation method of a multi-task deep learning network includes: building at least one shared layer, where the shared layer is configured to receive a plurality of pieces of input information and generate a plurality of pieces of processed feature information; building a plurality of groups of task-specific layers, wherein the groups of task-specific layers compute and generate a plurality of groups of output information corresponding to a plurality of different tasks according to the pieces of processed feature information; and activating at least one of the groups of task-specific layers in stages according to a power supply state of an electronic apparatus.

## Claims

### Claim 1

A generation method of a multi-task deep learning network, comprising:
building at least one shared layer, wherein the at least one shared layer is configured to receive a plurality of pieces of input information and generate a plurality of pieces of processed feature information;
building a plurality of groups of task-specific layers, wherein the groups of task-specific layers compute and generate a plurality of groups of output information corresponding to a plurality of different tasks according to the pieces of processed feature information; and
activating at least one of the groups of task-specific layers in stages according to a power supply state of an electronic apparatus and a power consumption requirement of each of the task-specific layers,
wherein the step of building the groups of task-specific layers further comprises:
creating a plurality of pieces of power consumption information corresponding to a plurality of arithmetic units;
selecting plural ones among the arithmetic units to build a selected model based on minimum total power consumption according to a corresponding accuracy requirement and the pieces of power consumption information for each one among the groups of task-specific layers;
performing a network weight sparsity analysis and a utilization evaluation to optimize network pruning to generate a hardware-aware model according to the selected model; and
generating each one of the groups of task-specific layers according to the hardware-aware model.

### Claim 2 (depends on 1)

The generation method according to claim 1, wherein a plurality of requirement conditions are required for generating the groups of task-specific layers, each of the requirement conditions comprises at least one of a power consumption requirement, a computing power requirement, and a hardware resource requirement, and the requirement conditions corresponds to a plurality of priorities.

### Claim 3 (depends on 2)

The generation method according to claim 2, wherein the one among the plurality of groups of task-specific layers corresponding to a highest priority has at least one of a minimum power consumption requirement, a minimum computing power requirement, and a minimum hardware requirement.

### Claim 4 (depends on 2)

The generation method according to claim 2, wherein the step of activating at least one of the groups of task-specific layers in stages according to the power supply state of the electronic apparatus and the power consumption requirement of each of the task-specific layers further comprises:
turning on at least one of the groups of task-specific layers according to the priorities and the requirement conditions based on the power supply state of the electronic apparatus.

### Claim 5 (depends on 1)

The generation method according to claim 1, wherein the step before performing the network weight sparsity analysis and the utilization evaluation to optimize network pruning to generate the hardware-aware model according to the selected model further comprises:
performing a computational optimization analysis for the selected model through a compiler.

### Claim 6 (depends on 1)

The generation method according to claim 1, wherein the step of selecting plural ones among the arithmetic units to build the selected model based on the minimum total power consumption according to the corresponding accuracy requirement and the pieces of power consumption information further comprises:
selecting a plurality of selected arithmetic units with highest computing power among the arithmetic units to establish an initial model;
cropping the initial model based on the corresponding accuracy requirement; and
obtaining the selected model according to the minimum total power consumption.

### Claim 7

A multi-task deep learning network, suitable for an electronic apparatus, comprising:
at least one shared layer, wherein the at least one shared layer is configured to receive a plurality of pieces of input information and generate a plurality of pieces of processed feature information; and
a plurality of groups of task-specific layers, wherein the groups of task-specific layers compute and generate a plurality of groups of output information corresponding to a plurality of different tasks according to the pieces of processed feature information,
wherein each one of the groups of task-specific layers is activated in stages according to a power supply state of the electronic apparatus and a power consumption requirement of each of the task-specific layers,
wherein the groups of task-specific layers are generated by a processor with computing power, and the processor is configured to:
creating a plurality of pieces of power consumption information corresponding to a plurality of arithmetic units;
selecting plural ones among the arithmetic units to build a selected model based on minimum total power consumption according to a corresponding accuracy requirement and the pieces of power consumption information for each one among the groups of task-specific layers;
performing a network weight sparsity analysis and a utilization evaluation to optimize network pruning to generate a hardware-aware model according to the selected model; and
generating each one of the groups of task-specific layers according to the hardware-aware model.

### Claim 8 (depends on 7)

The multi-task deep learning network according to claim 7, wherein a plurality of requirement conditions are required for generating the groups of task-specific layers, each of the requirement conditions comprises at least one of a power consumption requirement, a computing power requirement, and a hardware requirement, and the requirement conditions corresponds to a plurality of priorities, and the one among the plurality of groups of task-specific layers corresponding to a highest priority has at least one of a minimum power consumption requirement, a minimum computing power requirement, and a minimum hardware requirement.

### Claim 9 (depends on 8)

The multi-task deep learning network according to claim 8, wherein the electronic apparatus turns on at least one of the groups of task-specific layers according to the priorities and the requirement conditions based on the power supply state.