Although philosophically analytics without reasoning seems wrong, in machine learning solutions it is commonly found. Recently DataScienceCentral had a related webinar regarding this issue - http://www.datasciencecentral.com/video/the-myth-of-the-machine-learning-black-box
Without disputing the deep learning model, I'd like to mention again the Occam's razor principle (leaner is better) regarding very complex matters.
Edith,
Your questions, raises to me two very different points with regards to reasoning in analytics.
The first is whether the ML model is a black box type (such as a backpropagation NN) which just tends to minimize the error without any explanation on why it chose that particular weight matrix (nor can we easily translate the matrix into something useful for such analysis). In this regard, problems are just an optimization problem which faces serious issue in the same vein as a CEO faces in just slashing costs without analyzing what are the underlying causes of inefficiency.
The second which is more profound is that, if you accept Popperian Philosophy, then there is a fundamentally wrong way of applying ML to scientific problems. Under this view our current inductive analytics is all wrong in arriving at a generalization. Instead what is exposed is a rule to falsify our knowledge by using Modus Tollens. While this does not preclude using inductive analytics to arrive at a possible universal statement. It leaves the objective of Popperian philosophy out by not employing further confirmation through a modus tollens type of reasoning.
Hope this helps
Arturo, Thanks, I agree with the both points.
Having a rational behind the formulas is a must also in my view as it serves for validation as well as getting insights from the analytics.
About the Popperian rejection of induction by extrapolation, thanks for the contribution. of the Modus Tollens perspective. It will take me more time to get into that but I agree with the notion of empiric validation. The statistics induction has too many built in assumptions and I doubt its use in learning big data even for Deduction purposes -- due to the existence of different "states", multi-variant interactions and dynamics that defy "the law of large number", without which there's no ML...
It's not all dark though. I'm optimistic about having a remedy to the above shortcomings.
Mainz, Germany
Dear Ohri,
One way to look at the matter of machine learning and, as you put it, "analysis without reasoning," is that machine learning is mining the results of reasoning already embedded in the big data. If so, then it cannot reasonably be viewed as a substitute for reasoning such as informs the data.
One might entertain the hypothesis that there is a kind of "connection machine" architecture involved in lower cognition, but this ultimately functions to support the reasoning, rational faculties--which regularizes its results. In any case, the lesson would be that machine learning is no substitute for human reasoning, evaluation and formulation of theory, testing, definitions, etc.
H.G.Callaway
Callaway,
I find your notion very intriguing and I would say that current trends in ML do tend to do so, but in my opinion, I would debate the generalization to all ML. Reasoning systems (which I consider also part of ML, but that is also debatable) can derive knowledge that could be considered outside of the given knowledge base through application of logical rules. While derivable from the knowledge base it does a transformation via logical rules deriving new rules from which other rules could further be derived.
Just a thought
Epistemologically (philosophy) is not valid . At least enunciate the principles and/or axioms/beliefs which undergo your research
- Badges
- Science topic