Search This Blog

Monday, October 26, 2015

What studying Machine Learning has really taught me

I find myself surprised, even a little disappointed, by how emotions drive many of my decisions. Emotions feel like a necessary evil; many times they prevent me from making morally wrong decisions, much like religion does, but at other times they cloud my judgement and make it hard for me to think logically. I think we all have something to learn from Machine Learning here.

Whenever I am presented with a tough choice, more often than not I am able to reason about the merits and demerits of each option. This makes sense - years or training data provided to the system in my brain, along with positive and negative feedback, are helping it draw sharper boundaries based on the different attribute values presented to it. But here's where things get tricky. While I am able to reason about the positive and negative feedback that I received from previous choices that I have made, there are elements of the classification system that have formed as a result of emotions, possessing weights that I have no idea how to interpret. Let me explain that a bit better. Say I watch a violent movie by telling myself that I won't be affected by it (I do this often, because I believe that I understand better than to let some movie affect me). But this is analogous to saying, "Hey I'll just pass these (partially) misclassified data points through the training module of this Machine Learning system, but it's all good! I have added conditional statements to prevent these points from affecting these few weight values, so the integrity of the whole system will be preserved". While I can be pretty sure that doing this won't significantly affect the weight values that I have kept fixed, I can make no such assumptions about weights that I am unaware of. So the repercussions of passing such points through my classification system could be catastrophic to the accuracy and general coherency of the system. While it is true that the brain is a much more complex system than I have outlined and this might not be how it works, it is highly likely that watching that violent movie does affect weight values somewhere in a way that I cannot hope to understand. Over time a combination of these modified weight values could mutate the decision boundaries produced by the system to such an extent that I would be shocked the next time that I attempt to understand the decision making process.
Coming back to how I started this post, I am now able to relate the involvement of emotions in the decisions that I make and the inscrutable weight values in my classification system. It is pretty clear now that the reason that I am shocked when emotions affect my decisions is that I did not know the decision boundaries had been bent so. However, there is one aspect of this that doesn't fit well into this whole picture. Whenever I am presented with a choice, I see not one decision boundary, but two or more of them. One boundary that is the result of logical thinking, formed by weighing the pros and cons of each option, and another one, that is the result of emotions. Of course, I am oversimplifying; the two decision boundaries are not completely independent of each other, and sometimes there are way more than two sets of boundaries. But let's start with the easy case of two independent boundaries and solve that first.
The fact that there I am presented with two boundaries that I am able to perceive leads me to believe that the classification system in my brain uses an ensemble of classifiers to arrive at a decision. This raises many pertinent issues; are all the classifiers independent of each other? Were they trained independently, i.e. emotion related data points in a separate classifier and other points elsewhere (if this is even possible)? What is the function used to combine their output decisions? Why do I feel that the output produced by the emotion-only classifier is weighted much higher than the others in the combination function?
I've made a promise to myself to better analyze these data points and decision boundaries to try to understand how the classification system works. Furthermore, until I've done so to a satisfactory level, I'm going to assume that even mundane things that I do everyday are affecting weight values in ways that I do not understand, and so I need to keep a close watch on my emotions whenever I receive any training data (which is just about all the time!). Also, if anyone has any data files that can help me study this phenomenon, please email the labelled examples (preferably in CSV format) to me for analysis.
Lol jk :)