According to the bureau of labor statistics Computer and Information Research Scientists jobs are projected to grow 16%. While I believe that one of the “hot words” for now is Machine Learning, the applications and use cases can only speak for themselves. This field is growing exponentially fast and I feel that its foundation is somewhat shaky. This opinion is based off the answers I have received from the following three questions:

  • How does one appropriately develop a neural network and determine how good it is?
  • Is there any statistical/mathematical approach that has more foundation that would be better to use than a Neural Network?
  • If this product goes into production how can we guarantee that the model can be updated, built upon and changed based on the changes of the user?

Most of the common answers I have received is as follows:

  • There is not any hard science (at the moment) that can help us determine how to build neural networks
  • We think that using a Neural Network will be a “Novel” Approach to this problem/ We think it will produce interesting results/ Based off of previous results from Neural Networks we believe can get better results by implementing them.
  • No answer has been provided yet other than training a new network.

While again, I do believe Neural Network in general can have a huge impact in progressing science and helping with other general problems I feel as if though we as scientist should be careful as how we move forward.  Without knowing we can implement our own biases into these Networks and this in turn can cause damage. In order to better move forward and build upon the shoulder of giants we should first make sure that our foundation is solid. In order to do this I think that it is imperative to look into answering these questions.

 

Share this Post

2 Comments

  1. Hi, if my understanding is correct, in case of image classification and NLP, if I have a pre-trained model, to train on new data, I can reshape the data according to the pre-trained model. I am trying to use transfer learning for a regression problem. Consider I train a base model with 15 parameters and 1 million rows. I train a model. Now if I want to use this model for a similar problem where I have only 14 parameters, one parameter is missing. Will the pre-trained model be of any use. Is there a way I can use transfer learning in such cases?

    1. TLDR: No the input parameters need to be exactly the same

      Sadly Neural networks arent that advanced yet; if you imagine the basic “neural network” it has a set of inputs and connections throughout the layers that end up being weighted to some output. “transfer learning” from my understanding takes the weights and biases from one neural network and maps it completely to the same neural network. I like to think of this as an electrical circuit; if two circuits are the same they power will travel the same path. However, if you have two circuits that are not the same, the power will not take the same path.

      While this is my understanding, I am not sure of your particular problem. It might be possible to “wrap” one neural network onto another, but that doesn’t mean the accuracy will be the same. The best thing to do in my opinion is to build a new model and train it.

Leave a Comment

Your email address will not be published. Required fields are marked *

*
*