AI Rhyme But No Reason: Part 2
Objective: To explain the unexplainability of AI in lay person terms
So, here is a potentially mind-blowing fact.
For all the amazing properties of modern AI , and for all the incredible predictions that they are able to make to create the perception of “intelligence”, there are very few scientists in the world who may be able to rationalize and verbalize the steps a neural network takes to arrive at its prediction.
There is a current DARPA program (XAI) that puts it aptly:
Dramatic success in machine learning has led to a torrent of Artificial Intelligence (AI) applications. Continued advances promise to produce autonomous systems that will perceive, learn, decide, and act on their own. However, the effectiveness of these systems is limited by the machine’s current inability to explain their decisions and actions to human users.
Sure, researchers are trying, but current state of art thrives in part on blind faith, mesmerized by the improving accuracies of predictions.
Here is an explanation for this artificial disability that you may find familiar.
Recognizing Your Dad
Suppose that a close family member, say your father, is coming to town, and you are scheduled to pick him up from the airport. Due to a last moment emergency, you are unable to go. So, you ask me if I could pick him up.
I ask, “how will I recognize him?”
You respond, “I don’t have a photo on me but let me give you a verbal description”. You write out your verbal description, giving hair color, ethnicity, height, build, distinguishing features, and so on.
Now, imagine me standing at the arrival gate comparing the travelers emerging with my sense of your verbal description. I would analyze their appearance for aspects that match your description. I would recognize those aspects using my independently developed sense of ‘tall’ or ‘white curly hair’ or “well-built” and so on.
With that as a backdrop, let’s continue.
Compare that to how you recognize your family member. It is completely different. You hardly need any analysis. You don’t ponder about the last time you saw your dad from the right side, at a particular angle while he was dragging a suitcase to identify an approximate match in your mental database. You do not seek a specific memory amongst the millions of your dad to find the closest match. You JUST recognize him.
That instant familiarity comes from intuition. Intuition is the synthesis of experience. Your recognition is based on years and years of growing up and watching your dad from so many angles. Recognition is not memory recollection. You know that I am right because your recognition is split-second at any given moment, and because you will recognize your dad even if he wears a hat you’ve never seen him wear before.
Your verbal description will never compare to the high confidence you have in recognizing your dad. You could write a 10 page description for me, and it won’t stand a chance against your own intuition. Those words have NOTHING to do with how you recognize him.
Intuition is not explainable in words.
In a previous post, I discussed intuitive machines. Machine learning synthesizes data into patterns to make predictions on previously unseen data. Those patterns are the machine’s intuition. Those patterns is similarly unexplainable (thus far). In essence, during training, the machine focussed on the objective — improving prediction accuracy. Keeping track of why it learnt what it learnt was never set as its objective.
Now, are you able to transfer your intuitive sense of your dad to me to make recognition easier? No. You are alone in the universe in your unique, personalized synthesis of ‘dad’ memories. You have no words to explain even to yourself what comprises this mental thing called intuition that recognizes family members. You just do it. Do you agree? Do you think even a sibling shares your exact synthesis? Your experience of your dad, not to say anything about your vision and brain function, is unique to you.
The Tower of Babel and the Vulcan Mind Meld
Ever felt like saying to someone, ‘I wish you’d just get me’? Has human language created barriers? For literary flair, is this a side effect of the biblical curse from the Tower of Babel? Stephen Curry, Tom Brady, and Katie Ledecky cannot transfer their intuition for playing their game. They might coach someone, but that someone has to build his or her own intuition through experience.
We haven’t talked about the nature of AI synthesis in depth yet. But, I can tell you this : a machine can transfer its intuition from a backend system to a smartphone. There is also a technique called “Transfer Learning” where an expert AI can teach a novice AI some common intuition, before letting it specialize towards its own task.
While AI catches up and in some ways, surpasses humans, will humanity evolve to a higher form of communication? The Star Trek Vulcan mind-meld comes to mind.
Returning to the ground, this hidden nature of human intuition is also a core, and bewildering aspect of the outcome of Machine Learning. I wish I could transfer my synthesis of AI to you without so many words.