So, I began my blogs after a loooong break with the classic philosophy of Tao Te Ching (The Way), as you can see. Although I like reading philosophy as a hobby, it’s not that I am taking a philosophy class or something as of now, but I feel its highly related to what I learned in the last class of “Learning and Memory.”
We already know what learning means, so what is generalization? It’s the payoff for all your hard work to learn something! When you learn something new and gain experience, generalization helps you to extend that experience to something which you might never have experienced or explicitly learned before. It’s like, say you learned to drive a car; then you can drive any other 4 wheeler without much effort, right? That’s obvious for us, but how and why it happens is still blurry in neuroscience and AI as well.
Discrimination, in a way, is the exact opposite of generalization. It’s more about learning to differentiate between the learned object depending on various properties and context. It helps in responding differently to different things. It reduces your generalization scope. Taking the earlier example, say after learning your car and boasting your skills about driving any vehicle like a pro. One day your friend buys the latest model of an automatic transmission sedan and asks you to check out the hot wheels. As usual, you jump in, and guess what, you find there is neither gearstick nor clutch there in the hot seat! Still, you roar the engines and spin the wheel. Suddenly a dog comes in front, and your usual instincts drop in. You try to find the gearstick to control and press the brake pad like a clutch. And I guess everyone can imagine the rest of the story. You embarrassingly look into your friend’s eyes and asks him how do you even drive this thing.
Lesson learned you learned to discriminate between a manual and an automatic car. Overgeneralization is a bad policy, and discrimination helps to reduce the scope of erroneous generalization. But that’s true for the other way round as well. If you discriminate between every car, then you won’t be able to drive any other one then your car.
Enough driving! Focussing on the topic at hand, to understand these concepts, scientists often use something called the generalization gradient. It’s like plotting a graph of various stimulant response pairs, where the shape is mostly like a bell curve or Gaussian. The wider the curve, the higher the generalization scope of the study. The narrower it is, the more discriminative it will be when compared with other stimuli. There is also the case of peak shifts post learning. It may occur due to the supposed summation of underlying excitatory and inhibitory responses.
I feel the concepts of Generalization and Discrimination in the context of the neuroscience of learning and memory…, are more like Yin and Yang. Both are necessary, and neither of them is good or bad. It’s their balance that achieves harmony.