Comparing Neural Networks to Real World
So, we think that Rectified Linear Unit are better at learning than sigmoid units, don’t you think that people with a linear approach to life who think that if this happen then this, if they have given 50% best the output will be 50% of best. Those who do not relate one thing complexly with others but follow simplicity are the people doing wonders here.
Don’t you think that our history on earth can be thought of as some neural network. As we provide more of our knowledge to our children, better they would become in utilising the resources. I still think that if we add links in between the neurons in same layer maybe they are going to learn better as you are asking one neuron to limit itself to some thing as some other neuron is learning a different thing.
And we should learn from the current architecture also, espescially backprop. If you give knowledge to someone who applies it for some task and it fails, the student should go to the teacher and frankly say that “Sir you have taught wrong” and in this way maybe we are deteriorating our relation with our teacher but this feedback would help others also.
And what if this universe is a neural network, learning something for some bigger creature. If that be the case, I think we are learning greatly in the direction of science.
