Message Boards Message Boards

0
|
1706 Views
|
2 Replies
|
1 Total Likes
View groups...
Share
Share this post:

NN architecture to solve Problem like Target = Input1 * Input 2 + Input3

Don Baechtel
Posted 10 months ago

NN with neurons that are designed as a Sum of Weighted Inputs is no good at solving problems where the Output Targets are, in part, a Product or a Function of multiple inputs. Is there a NN architecture that can solve problems like Target = Input1 * Input 2 + Input3? What do these architectures look like? Do these Input Products of multiple Inputs need to be represented as additional Inputs like Input6 = Input1 * Input2? Or should permutations of Input Products be treated as additional Hidden Layers as a Sum of Weighted Products of Inputs? Then what about Target = Input1 * Input2 * Input3 ? What is your advice?
POSTED BY: Don Baechtel
2 Replies
Sort By:
Don Baechtel
Don Baechtel
Posted 10 months ago

THANKS.
POSTED BY: Don Baechtel

It really depends on what you want your trainable parameters to be. For a model such as

a * Input1 * Input 2 + b * Input3 + c

you can use this network

FunctionLayer[
 Function[
  NetArray[]*#Input1*#Input2 + NetArray[]*#Input3 + NetArray[]]]

If, let's say, the input are 5-d arrays, the fully specified net will be

d = 5;
net = NetInitialize@FunctionLayer[Function[
    NetArray["Dimensions" -> d]*#Input1*#Input2 +
     NetArray["Dimensions" -> d]*#Input3 +
     NetArray["Dimensions" -> d]],
   "Input1" -> d, "Input2" -> d, "Input3" -> d]
POSTED BY: Giulio Alessandrini
Reply to this discussion
Community posts can be styled and formatted using the Markdown syntax.
Tag limit exceeded
Note: Only the first five people you tag will receive an email notification; the other tagged names will appear as links to their profiles.
Reply Preview
Attachments
Remove
or Discard

Group Abstract Group Abstract