Method and apparatus for learning dependency of feature data

1 . A neural network device for learning dependence of feature data, comprising:

a memory in which at least one program is stored; and

a processor that performs a calculation by executing the at least one program,

wherein the processor is configured to:

acquire graph information including a data node for a human body;

extract feature data corresponding to a plurality of joints constituting the human body from the graph information;

acquire a self-attention output corresponding to the feature data based on a self-attention mechanism, wherein, to acquire the self-attention output, the processor is further configured to:

identify a plurality of point positions corresponding to the plurality of joints from the feature data;

perform the positional encoding based on the plurality of point positions, wherein the positional encoding is performed according to Equation 2 below:

PE ( p i )= P ( p i /10000 2q/d )  (Equation 2),

where p i denotes any one of the plurality of point positions, q denotes a dimension, and d denotes a total embedding dimension value;

identify geodesic positions corresponding to the plurality of joints from the feature data; and

perform the geodesic encoding based on the geodesic position; and

generate result data for a motion of the human body based on the self-attention output, and

the self-attention output includes position information acquired based on positional encoding of the feature data and structural information acquired based on geodesic encoding of the feature data.

2 . The neural network device of claim 1 , wherein, to identify the geodesic positions corresponding to the plurality of joints from the feature data, the processor is further configured to:

identify geodesic distances to the plurality of point positions;

generate the plurality of groups based on the geodesic distances; and

determine group positions corresponding to the plurality of groups.

3 . The neural network device of claim 2 , wherein, to identify the geodesic distances to the plurality of point positions, the processor is further configured to: identify a predetermined one of the plurality of point positions as a reference position, and identify geodesic distances between the plurality of point positions and the reference position.

4 . The neural network device of claim 3 , wherein the plurality of groups include point positions having the same geodesic distance among the plurality of point positions.

5 . The neural network device of claim 3 , wherein the geodesic encoding is performed according to Equation 3 below:

GE

⁡

(

p

i

)

=

𝒫

⁡

(

g

i

/

10000

2

⁢

q

/

d

)

,

g

i

=

𝒢

⁡

(

p

i

,

p

r

)

,

(

Equation

⁢

3

)

where p r denotes the reference position, q denotes a dimension, and d denotes a total embedding dimension value.

6 . A method of operating a neural network device for learning dependence of feature data, comprising:

acquiring graph information including a data node for a human body;

extracting feature data corresponding to a plurality of joints constituting the human body from the graph information;

acquiring a self-attention output corresponding to the feature data based on a self-attention mechanism, wherein the acquiring of the self-attention output includes:

identifying a plurality of point positions corresponding to the plurality of joints from the feature data;

performing the positional encoding based on the plurality of point positions, wherein the positional encoding is performed according to Equation 2 below:

PE ( p i )= P ( p i /10000 2q/d )  (Equation 2),

where p i denotes any one of the plurality of point positions, q denotes a dimension, and d denotes a total embedding dimension value;

identifying geodesic positions corresponding to the plurality of joints from the feature data; and

performing the geodesic encoding based on the geodesic position; and

generating result data for a motion of the human body based on the self-attention output,

wherein the self-attention output includes position information acquired based on positional encoding of the feature data and structural information acquired based on geodesic encoding of the feature data.

7 . The method of claim 6 , wherein the identifying of the geodesic positions corresponding to the plurality of joints from the feature data includes:

identifying geodesic distances to the plurality of point positions;

generating the plurality of groups based on the geodesic distances; and

determining group positions corresponding to the plurality of groups.

8 . The method of claim 7 , wherein the identifying of the geodesic distances to the plurality of point positions includes:

identifying a predetermined one of the plurality of point positions as a reference position; and

identifying geodesic distances between the plurality of point positions and the reference position.

9 . The method of claim 8 , wherein the plurality of groups include point positions having the same geodesic distance among the plurality of point positions.

10 . The method of claim 8 , wherein the geodesic encoding is performed according to Equation 3 below:

GE

⁡

(

p

i

)

=

𝒫

⁡

(

g

i

/

10000

2

⁢

q

/

d

)

,

g

i

=

𝒢

⁡

(

p

i

,

p

r

)

,

(

Equation

⁢

3

)

where p r denotes the reference position, q denotes a dimension, and d denotes a total embedding dimension value.