Multi-source transfer learning from pre-trained networks

Patent metadata

Field	Value
Publication number	US-11514318-B2
Application number	US-202016843173-A
Country	US
Kind code	B2
Filing date	Apr 8, 2020
Priority date	Apr 8, 2020
Publication date	Nov 29, 2022
Grant date	Nov 29, 2022

Abstract

Examples described herein provide a computer-implemented method that includes training, by one or more processing devices, a first neural network for classification based on training data in accordance with a first learning objective, the first neural network producing an intermediate feature function and a final feature function as outputs. The computer-implemented method further includes training, by the one or more processing devices, a second neural network for classification based on the intermediate feature function and the final feature function and further based at least in part on target task samples in accordance with a second learning objective. Training the second neural network includes computing maximal correlation functions of each of the intermediate feature function, the final feature function, and the target task samples.

First claim

What is claimed is:

1. A computer-implemented method comprising:
training, by one or more processing devices, a first neural network for classification based on training data in accordance with a first learning objective, the first neural network producing an intermediate feature function and a final feature function as outputs; and
training, by the one or more processing devices, a second neural network for classification based on the intermediate feature function and the final feature function and further based at least in part on target task samples in accordance with a second learning objective, wherein training the second neural network comprises computing maximal correlation functions of each of the intermediate feature function, the final feature function, and the target task samples.

2. The computer-implemented method of claim 1 , further comprising:
making, by the one or more processing devices, a classification decision for new data.

3. The computer-implemented method of claim 1 , wherein making the classification decision for the new data is based at least in part on a maximal weighting modal estimation of distribution.

4. The computer-implemented method of claim 3 , wherein the maximal weighting estimation of the distribution is based on the following equation:

P
^

Y
|
X

(

y
|
x

)

≈

P
^

Y
t

(
y
)

⁢

(

1
+

∑

n
,
i

σ

n
,
i

⁢

f
i

s
n

(
x
)

⁢

g
i

s
n

(
y
)

)

.

5. The computer-implemented method of claim 1 , wherein the training data represents an image.

6. The computer-implemented method of claim 1 , wherein the maximal correlation function for each of the intermediate feature function, the final feature function, and the target task samples is computed based on the following equation:

σ
i

=

max

g
i

(
y
)

𝔼
[

f
i

(
x
)

⁢

g
i

(
y
)

]

.

7. The computer-implemented method of claim 1 , wherein the first neural network is a LeNet architecture network.

8. The computer-implemented method of claim 1 , wherein the first neural network is of an architecture that produces a numerical output.

9. A system comprising:
a memory comprising computer readable instructions; and
one or more processing devices for executing the computer readable instructions, the computer readable instructions controlling the processing device to perform operations comprising:
training, by the one or more processing devices, a first neural network for classification based on training data in accordance with a first learning objective, the first neural network producing an intermediate feature function and a final feature function as outputs; and
training, by the one or more processing devices, a second neural network for classification based on the intermediate feature function and the final feature function and further based at least in part on target task samples in accordance with a second learning objective, wherein training the second neural network comprises computing maximal correlation functions of each of the intermediate feature function, the final feature function, and the target task samples.

10. The system of claim 9 , wherein the operations further comprise:
making, by the one or more processing devices, a classification decision for new data.

11. The system of claim 9 , wherein making the classification decision for the new data is based at least in part on a maximal weighting modal estimation of distribution.

12. The system of claim 11 , wherein the maximal weighting estimation of the distribution is based on the following equation:

P
^

Y
|
X

(

y
|
x

)

≈

P
^

Assignees

• IBM
• Massachusetts Inst Technology

Inventors

• Lee Joshua Ka-Wing
• Sattigeri Prasanna
• Wornell Gregory

Classifications

• G06N3/047

  Probabilistic or stochastic networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06N3/0454

  Physics · mapped topic

• G06N3/0472

  Physics · mapped topic