Some people in Academia criticize those who use neural networks for using a tool without knowing what is happening inside. While I agree (with academics like Cynthia Rudin) that we should benchmark these models against white box models and statistical methods, I think this opinion against black boxes has its origin in conflating data science with pure sciences.
In my opinion data science is not a science (for example in Popperian or Kuhnian sense).
Although data science is highly mathematical it shares methologies with humanities and social sciences where declaration of limitations is usually applied over attempts to reach "objective" results as in pure sciences.
A neural network is an instrument like a radio to capture a pattern of particular complexity from a sample of data. Main engineering task at the application level is to engineer the sample (dataset) where model is a mere tool.
I think a well-formulated philosophy of deep learning is in need.
I agree with your statement that "data science is not a science (for example in Popperian or Kuhnian sense)."
Indeed, data science in its current state seeks a lot of empirical statistical associations, data correlations, and patterns in hope that they can be helpful for discovering some hidden insights. But it is the reason why data science cannot be called real science at all.
In general, science can be broadly defined as a logical system of getting and organizing insights in some area of knowledge as a set of some principles, explanations that can be used for making some previously unknown but testable predictions. In order to qualify as a principle, insight must be both highly general (applicable to many settings) and stable (relevant now and in the future developments).
What are specific data science principles that make it real science? Data science is just a loose collection of various computational methods applied to fit empirical data sets. All generated or collected data represent the past, i.e. what was already happen. However, it is well known that the future is never defined solely by its past regardless of the amount of past data.
A discovered empirical association can, of course, be a starting point in research but such research is never becoming real science until it includes a causal relationship between the variables (factors or features), i.e. understanding. A neural network totally lacks that important feature. That's why its power is inherently limited.
Yes. But I think there is another approach we can consider. Instead of looking at data science as an incomplete would-be science, we can keep it in the domain of parasciences to allow constructivist methodology just like social sciences and humanities.
Dear Thamalu Maliththa Piyadigama
Generally, The approximation given by the neural network will not give you any insight on the form of f. There is no simple link between the weights and the function being approximated. Even the analysis of which input characteristic is irrelevant is a open problem
This discussion might be useful, have a look
https://stats.stackexchange.com/questions/93705/why-are-neural-networks-described-as-black-box-models
Kind Regards
Qamar Ul Islam
- Badges
- Science topic