The Artificial intelligence era has almost come full swing, potentially advancing and revolutionizing all aspects of life. The various fields being affected by the AI revolution is increasing exponentially.
Data processing is a core aspect of artificial intelligence. Modern computers can now process massive amounts of information in increasingly less time. However, a question to consider here is whether these systems understand the information being processed. Most importantly, will these systems ever be able to think themselves.
Here I take a different outlook on the topic and introduce how discussions in modern philosophy can assist us in sculpting new deep learning models.
What Modern Philosophers Have To Say
We time-travel all the way back to 1641 to a skeptical Rene Descartes sitting by the fire and writing his philosophical masterpiece, Meditations on First Philosophy.
Amongst many things, Descartes states an interesting theory on the composition of his thoughts. He states that his thoughts are nothing more than an amalgamation of his perceived world. The theory in its self is very simple and can be visualized in a numerous number of ways. For instance, let’s take a strictly imaginary creature, the unicorn.
A Mythical Unicorn
What Descartes was simply stating was that imaginary creatures like the unicorn, are just compositions of entities we have perceived, like horses and horned animals.
Now, why is this relevant to intelligent systems thinking? Well, simply put, intelligent systems are already excellent at perceiving objects, so they already contain the fundamental building blocks of thoughts. More specifically, intelligent systems, that are categorically deep learning models, are outstanding in identifying objects from input data. Input data can be provided in numerous ways, all of which are typical sensory data to humans. For instance, a pizza can be seen, touched and smelled (some would say heard, but that’s another issue).
Object classification models from visual and audio data, like images and recordings, are already commonly used in various fields. Classification models relying on other forms of sensory data are being experimented with. For instance, odor detection machine learning models, capable of detecting odor patterns and classifying them based off of training data.
Essentially, we already have the availability of the fundamental building blocks of thoughts. We now need a way to progressively develop our sensory perception into deeper thought.
Our building blocks have been built on sensory perceptions, in supplementation, we assume our intelligent system has basic fundamental mathematical and data processing capabilities. Now let’s work through a small example of how this elementary system can be utilized to develop thought. We start with the classification of a simple image depicting one apple. We assume that the model training set was saturated with enough information to correctly classify the image correctly. In doing so, the output representation can be a string, something along the lines of “Single Apple” or “1 Apple”. Simple parsing algorithms can now determine the numeric value assigned to the entity under inspection, in this case, the apple has a numeric value of one. The intelligent systems, in turn, can now expand and augment it’s already large training dataset by altering the numeric value and creating a sensory perception to illustrate that change.
However, instead of focusing solely on numerical value characterization, more emphasis should be directed towards property identification. With the identification of properties, we can extract valuable information from different sensory perceptions and merge them into abstract ideas. The key part here is the identification of properties, which can range from colors to supplementary features, like a hat. What this allows the system to do is to add and remove features from sensory perceptions, over and over again. Layering the perception and then classifying it to augment the already existing database of knowledge. Creating a cyclical pattern of learning for the system, making the part of the learning process autonomous.