3
$\begingroup$

I know this is an ongoing and hard question to answer, but if anyone has experience in this then please share your knowledge.

Suppose I have made a neural network with the task of predicting an event x. Therefore, the network can be structured as a 2-class binary classification problem with n inputs and a single output.

Aside from training the network to predict such events, I'm having a hard time trying to visualise any patterns/inferences which have been made by the neural network.

As of now, I can think of the following:

  • Variable importance based on weight connection strength. (Garson, Olden)
  • Partial dependent plots (observe the networks output whilst sampling an input of interest.

Although these plots are nice and give us a descent amount of information, I'm looking to produce a more graphicy plot.

After researching, I've discovered Self Organizing Maps (SOM) which can be used as a visual tool of discovering patterns which may or may not exist in the giving the inputs. Essentially vector quantisation.

So my current thought process is to:

  1. Build a SOM map for all training data and analyse the patterns
  2. Classify the data using my neural network model
  3. Build new SOM maps for the classified cohorts and see if there's anything interesting.

So I guess I'm essentially looking for an intuitive plot which can show the inferences made by the neural network (or any machine learning approach for that matter).

Following advice in the comment section, my current network structure is as follows: 

25-5-1 # 25 inputs, 5 units in a single hidden layer, single output

I started with an equal number of units in the hidden layer but reduced to 5 after optimising via the ROC metric.

Improve this question
$\endgroup$
7
  • $\begingroup$ Do you have a particular DNN architecture in mind ? I would think that can get the conversation more structure... $\endgroup$
    – user180775
    Commented Nov 20, 2017 at 14:48
  • $\begingroup$ @ZakC Hi, at the moment, I'm just using a feed-forward network with a single hidden layer. The current topology is 25-5-1. Thanks for your reply $\endgroup$
    – Sam
    Commented Nov 20, 2017 at 14:56
  • $\begingroup$ Do you know about the saliency maps? For instance: Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps by Simonyan et al. These accomplish two things: i) "visualising the notion of the class", ii) a saliency map, that shows regions which contribute most to the decision/prediction. $\endgroup$
    – user166243
    Commented Nov 20, 2017 at 15:13
  • $\begingroup$ @user2137591 Thanks for your reply. Are saliency maps unique to convnet (image problems) or transferable to numeric data also? $\endgroup$
    – Sam
    Commented Nov 20, 2017 at 15:15
  • $\begingroup$ you can use them for numeric data as well (images are also numeric data). The difference with images is that in images the pixels in a neighborhood are usually correlated, hence you get more coherent visualizations. But the method is generic. $\endgroup$
    – user166243
    Commented Nov 20, 2017 at 15:21

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.