Why ANN Can’t Be Used In Sequential Data?
Table Of Contents:
- Input Text Having Varying Size.
- Zero Padding – Unnecessary Computation.
- Prediction Problem On Different Input Length.
- Not Considering Sequential Information.
Reason-1:Input Text Having Varying Size.
- In real life, input sentences will have different word counts.
- Suppose you make an ANN having the below structure.
- It has 3 input nodes.
- Our first sentence contains 6 words, hence the weight metrics will be 6 * 3 structure.
- The second sentence contains 3 words hence the weight metrics will be 3 * 3 structure.
- The third sentence contains 4 words hence the weight metrics will be 4 * 3 structure.
- You can see here the structure of the input weight metrics is changing based on the input text.
- Which is not practical for designing.
Reason-2:Zero Padding Unnecessary Computation.
- To solve the first issue of varying length we can use the zero padding technique.
- First, we can count the sentence having maximum words.
- In our case we have the first sentence having a maximum of 6 number of words.
- So we will fix our input text size to a maximum of 6 words.
- In the second sentence, we have a number of words, as we have fixed our input to 6 words, but we have 3 words in 2nd sentence hence we will append 3 more vectors having zero values inside it.
- Hence it is called zero padding.
- The problem with zero padding is that if we have the maximum word of a sentence is 1000 words.
- Then we will fix the input length to 1000 nodes.
- But if we got a sentence having only 5 words then for the rest of the 995 words we have to use zero padding.
- Which will take extra memory and computation power.
- Which will decrease the training speed of the model.
- Which is undesirable.
Reason-3: Prediction Problem On Different Input Length
- In our case, we have set our input length to 6 words while training the model.
- But while predicting suppose we got an input text having the length of 10 words, at that time our model will fail.
- Because we have trained our model with a fixed input size of 6 words, it will not be able to predict for 10 words.
Reason-4: Not Considering Sequential Information
- ANN architecture does not take into account the sequence information of the input text.
- When we pass the input text to the ANN model it will take all the input at a time.
- When you enter vales at a time it will be mixed up inside the network, hence the sequence information is discarded.
- The sequence information is discarded in the ANN model.
- Hence it is not suitable for the sequential data.