The limits of our current machine learning concept

First, to be clear, I am not talking about the actual machine learning techniques and models like RNN, CNN, LSTM, and such. Rather, I am talking about the bigger concept being machine learning (including deep learning) in general.

Our current concept of machine learning involves training a model based on some set of data. There is a "ground truth" in the training process. This is not a problem when trying to differentiate between a cat and a dog, or to recognize the letters of the English alphabet. These data sets have distinct and discrete categories.

But our world is not all black and white.

For example, a matchmaking app that uses past data of couples to recommend potential partners. Based on a list of attributes, the model learns what kind of people pair up. It then uses that model to recommend partners to singles. But to the person receiving the recommendation, is this the partner that he or she really wants? What if the model has not taken into account certain attributes? What if the model has assigned weights that are actually different from the target person's? After all, each person is unique, and the attribute space can be extremely large if we keep breaking it down.

 

What happens when our choices become those recommended by a machine? Instead of opening up our options, are we instead confining them to whatever is the limit of the model being used by the machine? Instead of fitting training data to a model, will we end up fitting humanity to a model's limited data space?

This question is always there in the realm of what we call the arts. The non-technical part of our lives. This includes language, which is a major realm of machine learning research. Two main areas are research in machine translation and text generation. But... there is no "right" answer in these realms. "I love you" can be translated as "愛している", but it can also be "月がきれいですね" if you really consider the entire cultural and literary context. But we are creating models that come up with the "right" answers, which in such cases, usually means selecting the best out of several possibilities. Eventually, as we keep selecting a single "right" answer, will we end up reinforcing the model to narrow its selection to that single "right" answer? Will we enter a self-fulfilling reinforcement loop that actually narrows our options, our creativity? Will we end up with translated texts that follow a certain style and uses only certain words and expressions? Will we end up generating texts that follow a certain fixed pattern?

And language shapes the way we think. We can have great ideas, but we can only express them, communicate them to others, through words. The words we can use, then, naturally limits what we can express. We can think, but a thought that cannot be manifested in the physical world remains confined within our own. Will our reliance on AI in the realm of language end up limiting the human mind?

As the character John Keating said in the movie Dead Poets Society, "We don't read and write poetry because it's cute. We read and write poetry because we are members of the human race. And the human race is filled with passion. And medicine, law, business, engineering, these are noble pursuits and necessary to sustain life. But poetry, beauty, romance, love, these are what we stay alive for."

Will our reliance on AI end up sustaining life but not make it worth living for?