What are eyes but a mechanism of nature to see nature, in fact that is not really the eye is it. The brain perceives, by any definition of the word, the brain makes conscious sense of the world around us. The eyes are a tool, sculpted in evolutionary history far previous to the sensory cortex, were for the purpose of mindless sensory input. The reason behind creation is a question many a mind has pondered, and the truest answers will be evident in their perseverance through time. Eyesight to the blind is a good analogy for the way machines must learn to see in order to be diligent in their pursuit of classification and generalization. I actually named this blog post after a song on one of my favorite albums, 'Tommy', by The Who. I think that there is a pretty cool, loose be that, but still cool connection between the concept of the album and the concept of the machine learning methodology and goal. Tommy is the story of a boy who is blind, deaf, and cannot hear; his absence of these skillsets gives Tommy a different perspective on life. There is a story from the blind to the seeing integrated into the story and the album truly has a deeper concept than the top layer of the story. It is about being able to see the deeper picture in life, which is ultimately subjective and which is why it is so impressive and difficult to write an album over. Music and art are the connections between this physical known world and the spiritual unknown. We really do not know why Beethoven's music makes us feel so happy.
Technology has the capability to see far beyond the eyes of people. Technology has precision of sight down to the atom, but what is vision without understanding? What is the power of recording audio if there is no ability to process it. Well there is the ability to process it, and that is ingrained into our evolutionary neurophysiology, and in our ability to process information. The goal of machine learning, deep learning, and artificial intelligence is the process of mimicking that ability, mimicking understanding. Understanding and learning are guided by the psychological models we create in order to understand life. We fundamentally understand what is harmful to us and we tend to avoid such things. Fear is the best predictive modeling system we have ingrained into us. Fear is a wall, or a barrier that limits our movement in life, and even our movement though our own psychology. Fear is an important part of our vision in life, even if it is the limiting factor. This suggests that we are not performing up to our fullest capabilities in life, because there is the limiting factor of fear that controls us to some extent.
Imagining a state of the art robot who is basically human, the goal is to conceptualize how to create something like this. Disregard facial features, and everything physical when it comes to humans. Think about things like speech and touch, aka the important things we need to think about. In this paragraph I will talk about speech and in the next paragraph I will talk about touch. Just as a disclaimer this is not a very heavily explored realm of thought... at all, actually really never explored so take it lightly with the criticism please. Spectrograms, it is common knowledge that a spectrogram is a visual, image representation of changes in frequency over time... imperfectly be that. How similar is a northern accent to midwestern accent saying the exact same thing, would there be a continuous fluid transform from the midwestern spectrogram to the northern one. In other words would there be some continuous, differentiable even, function that could be applied to the midwestern spectrogram to transform it into the northern one? Would there be some way to bend certain frequencies as a function of time to in turn transform the accent? The road to innovation starts with questioning, query about the nature of a certain phenomenon. There probably is a way to physically bend the frequency representation of a spectrogram in order to change the inflection of certain words to further change the accent of an audio clip. Change in accent does not mean change in meaning and that is what I essentially want my program to understand. There is no fluidity in the truest form of understanding which a sentence represents, there is merely the physical fluidity of physical change in the spectrogram. Just as there is no change in meaning with change in accent, there is no real change in meaning with changes in language. "我会走我的狗" and "I will walk my dog" have the same meaning, it is just the expression of that abstract thought that changes with language. Similar to the change in accent representing the physical fluid change in frequency representation in a spectrogram, there is some level of fluid change on some higher level of abstraction of thought that is much more complex and human with language. This is a very hard problem to solve but the point I want to get across is asking questions is integral to solving problems.
Now introducing the problem of how to interact with objects in a human like fashion. How can we teach an AI to predict how an unknown object will move. The answer, just like to all other ML problems is generalization. Can we generalize to that extent? Is it possible to have a machine develop such a broad and complex understanding? What is necessary to have, is a good engineer, the best engineer. The idea of creation is essentially described by almost every religion is God, in one form or another. This creator, this engineer of human behavior is analogous to some from of God. It can be hard to wrap your head around, but this needs to be a very aware, and smart inventor.