IP Library › Granted Patent US 12,517,983
Granted Patent B2
US 12,517,983 · App. 18/452,664 · Granted Jan 6, 2026

SNR detection with few-shot trained models

Inventors: Wei Cheng (Princeton Junction, NJ); Jingchao Ni (Princeton, NJ); Liang Tong (Lawrenceville, NJ); Haifeng Chen (West Windsor, NJ); Yizhou Zhang (Los Angeles, CA)
Assignee: NEC Corporation
G06F18/24133G06F18/2415H04B10/697
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,517,983
App. No.
18/452,664
Granted
Jan 6, 2026
Kind
B2
Abstract

Methods and systems for training a model include determining class prototypes of time series samples from a training dataset. A task corresponding to the time series samples is encoded using the class prototypes and a task-level configuration. A likelihood value is determined based on outputs of a time series density model, a task-class distance from a task embedding model, and a task density model. Parameters of the time series density model, the task embedding model, and the task density model are adjusted responsive to the likelihood value.

Claims (215)

1 . A computer-implemented method for training a model, comprising:

determining class prototypes of time series samples from a training dataset;

encoding a task corresponding to the time series samples, having a measurement of an optical communications signal that includes one or more of data obtained after polarization decomposition and data obtained after digital signal processing, using the class prototypes and a task-level configuration that includes configuration parameters relating to an environment of an optical communications network;

determining a likelihood value based on outputs of a time series density model, a task-class distance from a task embedding model, and a task density model; and

adjusting parameters of the time series density model, the task embedding model, and the task density model responsive to the likelihood value.

2 . The method of claim 1 , further comprising encoding the time series samples into a vector representation before determining the class prototypes.

3 . The method of claim 1 , wherein encoding the task includes encoding task level features that include the task-level configuration and encoding task instance features by aggregating the class prototype.

4 . The method of claim 1 , wherein the task density model determines a task density as a probability of a task given a cluster center and a covariance matrix of a cluster.

5 . The method of claim 1 , wherein the time series density model determines a time series density as a probability of a time series sample based on a class prototype and a covariance matrix of a class.

6 . The method of claim 1 , wherein the task-class distance is determined as a maximum of a distance between a prototype and a set of optimal class prototypes.

7 . The method of claim 1 , wherein the likelihood value has a lower bound:

ℓ

L

(

D

τ

;

θ

,

)

=

1

N

⁢

∑

i

=

1

n

(

log

(

e

i

|

y

i

)

+

𝔼

v

τ

~

⁢

q

ϕ

(

v

τ

|

𝒟

τ

s

)

[

log

(

y

i

|

v

τ

)

+

log

⁢

∑

z

τ

=

1

r

p

⁡

(

v

τ

|

z

τ

)

⁢

p

⁡

(

z

τ

)

]

)

+

H

⁡

(

q

ϕ

(

v

τ

|

𝒟

τ

s

)

)

where H(⋅) is an entropy function, N is a number of classes, p θ,w (e i |y i ) is a probability of an embedding e i given a sample y i with model parameter θ and trainable parameters w, is an expectation function, v τ is an embedded representation of task τ, D t s is a set of training support samples, q ϕ (⋅) is an approximated posterior with inference network ϕ, z τ is a latent task variable, and r is a hyperparameter of a number of components in a mixture distribution.

8 . The method of claim 1 , further comprising determining a new task-level configuration that reflects configuration parameters relating to a new environment, to adapt to the new environment.

9 . A system for training a model, comprising:

a hardware processor; and

a memory that stores a computer program which, when executed by the hardware processor, causes the hardware processor to:

determine class prototypes of time series samples from a training dataset;

encode a task corresponding to the time series samples, having a measurement of an optical communications signal that includes one or more of data obtained after polarization decomposition and data obtained after digital signal processing, using the class prototypes and a task-level configuration that includes configuration parameters relating to an environment of an optical communications network;

determine a likelihood value based on outputs of a time series density model, a task-class distance from a task embedding model, and a task density model; and

adjust parameters of the time series density model, the task embedding model, and the task density model responsive to the likelihood value.

10 . The system of claim 9 , wherein the computer program further causes the hardware processor to encode the time series samples into a vector representation before determining the class prototypes.

11 . The system of claim 9 , wherein the computer program further causes the hardware processor to encode task level features that include the task-level configuration and to encode task instance features by aggregating the class prototype.

12 . The system of claim 9 , wherein the task density model determines a task density as a probability of a task given a cluster center and a covariance matrix of a cluster.

13 . The system of claim 9 , wherein the time series density model determines a time series density as a probability of a time series sample based on a class prototype and a covariance matrix of a class.

14 . The system of claim 9 , wherein the task-class distance is determined as a maximum of a distance between a prototype and a set of optimal class prototypes.

15 . The system of claim 9 , wherein the likelihood value has a lower bound:

ℓ

L

(

𝒟

τ

;

θ

,

)

=

1

N

⁢

∑

i

=

1

n

(

log

(

e

i

|

y

i

)

+

𝔼

v

τ

~

q

ϕ

(

v

τ

|

𝒟

τ

s

)

[

log

(

y

i

|

v

τ

)

+

log

⁢

∑

z

τ

=

1

r

p

⁡

(

v

τ

|

z

τ

)

⁢

p

⁢

(

z

τ

)

]

)

+

H

⁡

(

q

ϕ

(

v

τ

|

𝒟

τ

s

)

)

where H(⋅) is an entropy function, N is a number of classes, p θ,w (e i |y i ) is a probability of an embedding e i given a sample y i with model parameter θ and trainable parameters w, is an expectation function, v τ is an embedded representation of task τ, D τ s is a set of training support samples, q ϕ (⋅) is an approximated posterior with inference network ϕ, z τ is a latent task variable, and r is a hyperparameter of a number of components in a mixture distribution.

16 . The system of claim 9 , wherein the computer program further causes the hardware processor to determine a new task-level configuration that reflects configuration parameters relating to a new environment, to adapt to the new environment.

Assignments (2)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Nov 18, 2025
From: NEC LABORATORIES AMERICA, INC.
To: NEC CORPORATION
Reel/Frame 072938/0913 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Aug 21, 2023
From: CHENG, WEI; NI, JINGCHAO; TONG, LIANG; CHEN, HAIFENG; ZHANG, YIZHOU
To: NEC LABORATORIES AMERICA, INC.
Reel/Frame 064646/0688 →
Continuity (3)
Provisional Application 63408550 · Sep 21, 2022
Provisional Application 63399747 · Aug 22, 2022
Related Publication 20240070232A1 · Feb 29, 2024
References Cited (23)
US 10963754B1 · Ravichandran · 2021 [cited by examiner]
US 11551004B2 · Wohlwend · 2023 [cited by examiner]
US 20200259700A1 · Bhalla · 2020 [cited by examiner]
US 20210201116A1 · Rabinowitz · 2021 [cited by examiner]
US 20220129758A1 · Sallee · 2022 [cited by examiner]
US 20230076127A1 · Yu · 2023 [cited by examiner]
US 20230236927A1 · Tang · 2023 [cited by examiner]
US 20230244917A1 · Loaiza Ganem · 2023 [cited by examiner]
US 20230252302A1 · Tong · 2023 [cited by examiner]
US 20240064161A1 · Liu · 2024 [cited by examiner]
US 20240104344A1 · Tang · 2024 [cited by examiner]
US 20240303149A1 · Chen · 2024 [cited by examiner]
EP 3754548A1 · 2020 [cited by examiner]
WO WO2021175422A1 · 2021 [cited by examiner]
WO WO2022076915A1 · 2022 [cited by examiner]
Snell et al; Prototypical Networks for Few-shot Learning ;2017; University of Toronto, pp. 1-13. (Year: 2017). [cited by examiner]
Zhang et al; Constellation Diagram Analyzer Based on Few Shot Learning ; 2021; Optical Society of America; pp. 1-3. (Year: 2021). [cited by examiner]
Zhang et al; (Fast adaptation of multi-task meta-learning for optical performance monitoring; Jul. 2023; Optics Express vol. 31, No. 14, pp. 1-15. (Year: 2023). [cited by examiner]
Finn et al., “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks”, arXiv:1703.03400v3 [cs.LG] Jul. 18, 2017, pp. 1-13. [cited by applicant]
Snell et al., “Prototypical Networks for Few-shot Learning”, arXiv:1703.05175v2 [cs.LG] Jun. 19, 2017, pp. 1-13. [cited by applicant]
Yao et al., “Hierarchically Structured Meta-learning”, arXiv:1905.05301v2 [cs.LG] Nov. 15, 2019, pp. 1-19. [cited by applicant]
Ba et al., “Predicting Deep Zero-Shot Convolutional Neural Networks using Textual Descriptions”, arXiv:1506.00511v2 [cs.LG] Sep. 25, 2015, pp. 1-15. [cited by applicant]
Reed et al., “Learning Deep Representations of Fine-Grained Visual Descriptions”, arXiv:1605.05395v1 [cs.CV] May 17, 2016, pp. 1-10. [cited by applicant]