# US 2026/0037830 A1 · Dynamic supernet learning apparatus and method for neural architecture search

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

- Number: US 2026/0037830 A1
- Document ID: US20260037830A1
- Status: Published application
- Kind: A1
- Application number: 18/799,660
- Filed: 2024-08-09
- Published: 2026-02-05
- Earliest claimed: 2024-07-30
- Assignees: UIF (University Industry Foundation), Yonsei University
- Applicants: UIF (University Industry Foundation), Yonsei University
- Inventors: Bumsub HAM; Jeimin JEON; Youngmin OH; Junghyup LEE; Donghyeon BAEK; Dohyung KIM; Chanho EOM
- Web page: https://patentgrep.com/patent/US20260037830A1

## CPC

- G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn
- G06N 3/045: Combinations of networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

The disclosed embodiment provides a supernet learning apparatus and method that can accurately compare each subnet based on performance when searching for a neural architecture by performing learning by adjusting the learning rate according to the complexity of each subnet extracted from the supernet, by performing the steps of analyzing the complexity of a subnet repeatedly extracted from the supernet, and setting a learning rate that is dynamically variable according to the number of learning repetitions based on the complexity analyzed in the extracted subnet, learning the extracted subnet using the learning rate that is set to be variable according to the number of learning repetitions, and merging the trained subnet into the supernet to obtain a trained supernet.

## Claims

### Claim 1

A supernet learning apparatus comprising:
a memory; and
a processor that executes at least a part of operations according to a program stored in the memory,
wherein the processor performs the steps of:
analyzing a complexity of a subnet repeatedly extracted from a supernet, and setting a learning rate that is dynamically variable according to a number of learning repetitions based on the complexity analyzed in the subnet,
learning the subnet using the learning rate that is set to be variable according to the number of learning repetitions, and
merging the subnet into the supernet to obtain a trained supernet.

### Claim 2 (depends on 1)

The supernet learning apparatus according to claim 1,
wherein the processor analyzes the complexity based on a number of weights included in a plurality of operation layers constituting the subnet.

### Claim 3 (depends on 1)

The supernet learning apparatus according to claim 1,
wherein the processor is configured to adjust the learning rate to gradually decrease as the number of learning repetitions increases, but a size of the decrease is adjusted differently depending on the complexity of the subnet.

### Claim 4 (depends on 1)

The supernet learning apparatus according to claim 1,
wherein the processor
sets the learning rate to decrease slowly with an increase in the number of learning repetitions when the complexity of the subnet is relatively high based on maximum complexity and minimum complexity, and
sets the learning rate to decrease quickly with an increase in the number of learning repetitions when the complexity of the subnet is relatively low.

### Claim 5 (depends on 1)

The supernet learning apparatus according to claim 1,
wherein the processor
calculates a learning reduction coefficient that controls speed of decrease of the learning rate for the subnet according to the complexity of the subnet, and
dynamically adjusts and sets the learning rate according to the number of learning repetitions using the learning reduction coefficient.

### Claim 6 (depends on 5)

The supernet learning apparatus according to claim 5,
wherein the processor sets the learning rate ηt according to Equation
η
t
=
η
0
·
(
1
-
t
T
)
γ
⁡
(
α
)
where, T represents a total number of learning repetitions, t represents the number of learning repetitions, n0 represents an initial learning rate, and γ(α) represents the learning reduction coefficient.

### Claim 7 (depends on 5)

The supernet learning apparatus according to claim 5,
wherein the processor calculates the learning reduction coefficient γ(α) according to the complexity C(α) of the subnet α according to Equation
γ
⁡
(
α
)
=
ωlog
⁡
(
𝒞
⁡
(
α
)
)
+
τ
where, ω and τ represent normalization weight and normalization bias.

### Claim 8 (depends on 7)

The supernet learning apparatus according to claim 7,
wherein the processor calculates the normalization weight ω according to Equation
ω
=
-
γ
max
-
γ
min
log
⁡
(
𝒞
max
)
-
log
⁡
(
𝒞
min
)
where, Cmax and Cmin represent the maximum complexity and minimum complexity of subnets being extracted, and γmax and γmin represent the maximum learning reduction coefficient and minimum learning reduction coefficient set for the learning reduction coefficient.

### Claim 9 (depends on 7)

The supernet learning apparatus according to claim 7,
wherein the processor calculates the normalization bias τ according to Equation
τ
=
γ
min
-
ωlog
⁡
(
𝒞
max
)
where, Cmax represents the maximum complexity among the complexities of extracted multiple subnets, and γmin represents the minimum learning reduction coefficient set for the learning reduction coefficient.

### Claim 10 (depends on 1)

The supernet learning apparatus according to claim 1,
wherein the processor
divides the supernet into multiple sub-supernets, extract subnets from each of the divided multiple sub-supernets, and
when the extracted subnets are trained, merges the trained subnets to obtain multiple trained sub-supernets, and merges the multiple trained sub-supernets again to obtain the trained supernet.

### Claim 11

A supernet learning method performed by a processor, the method including the steps of:
analyzing a complexity of a subnet repeatedly extracted from a supernet, and setting a learning rate that is dynamically variable according to a number of learning repetitions based on the complexity analyzed in the subnet;
learning the subnet using the learning rate that is set to be variable according to the number of learning repetitions; and
merging the subnet with the supernet to obtain a trained supernet.

### Claim 12 (depends on 11)

The supernet learning method according to claim 11,
wherein the step of setting a learning rate includes analyzing the complexity based on a number of weights included in a plurality of operation layers constituting the subnet.

### Claim 13 (depends on 11)

The supernet learning method according to claim 11,
wherein the step of setting a learning rate includes adjusting the learning rate to gradually decrease as the number of learning repetitions increases, but a size of the decrease is adjusted differently depending on the complexity of the subnet.

### Claim 14 (depends on 11)

The supernet learning method according to claim 11,
wherein the step of setting a learning rate includes
setting the learning rate to decrease slowly with an increase in the number of learning repetitions when the complexity of the subnet is relatively high based on the specified maximum complexity and minimum complexity, and
setting the learning rate to decrease quickly with an increase in the number of learning repetitions when the complexity of the subnet is relatively low.

### Claim 15 (depends on 11)

The supernet learning method according to claim 11,
wherein the step of setting a learning rate includes
calculating a learning reduction coefficient that controls speed of decrease of the learning rate for the subnet according to the complexity of the subnet, and
dynamically adjusting and setting the learning rate according to the number of learning repetitions using the learning reduction coefficient.

### Claim 16 (depends on 15)

The supernet learning method according to claim 15,
wherein the step of setting a learning rate includes
setting the learning rate ηt according to Equation
η
t
=
η
0
·
(
1
-
t
T
)
γ
⁡
(
α
)
where, T represents a total number of learning repetitions, t represents the number of learning repetitions, η0 represents an initial learning rate, and γ(α) represents the learning reduction coefficient.

### Claim 17 (depends on 15)

The supernet learning method according to claim 15,
wherein the step of setting a learning rate includes
calculating the learning reduction coefficient γ(α) according to the complexity C(α) of the subnet α according to Equation
γ
⁡
(
α
)
=
ωlog
⁡
(
𝒞
⁡
(
α
)
)
+
τ
where, ω and η represent normalization weight and normalization bias.

### Claim 18 (depends on 17)

The supernet learning method according to claim 17,
wherein the step of setting a learning rate includes
calculating the normalization weight ω according to Equation
ω
=
-
γ
max
-
γ
min
log
⁡
(
𝒞
max
)
-
log
⁡
(
𝒞
min
)
where, Cmax and Cmin represent the maximum complexity and minimum complexity of subnets being extracted, and γmax and γmin represent the maximum learning reduction coefficient and minimum learning reduction coefficient set for the learning reduction coefficient.

### Claim 19 (depends on 17)

The supernet learning method according to claim 17,
wherein the step of setting a learning rate includes
calculating the normalization bias τ according to Equation
τ
=
γ
min
-
ωlog
⁡
(
𝒞
max
)
where, Cmax represents the maximum complexity among the complexities of extracted multiple subnets, and γmin represents the minimum learning reduction coefficient set for the learning reduction coefficient.

### Claim 20 (depends on 11)

The supernet learning method according to claim 11,
wherein the step of obtaining the trained supernet includes,
when the supernet is divided into multiple sub-supernets, and subnets are extracted from each of the divided multiple sub-supernets,
merging the subnets to obtain multiple trained sub-supernets, and merging the multiple trained sub-supernets again to obtain the trained supernet.
