Artificial Intelligence: Look Ma, No Hands!
To Begin With The End..
I am having a hard time making a machine suffer. It being an egoless version of me.
Strengthening learning of a machine by giving it negative feedback is tricky. Doable in some sense, but one perambulates to quantify and correlate its suffering with its exact learning 🙂
The Human Connection to Artificial Intelligence — Two Of a Kind
Reinforcement machine learning deals with replicating human learning patterns of rewards and punishments on a machine. But then it is debatable if all humans learn alike from rewards/punishments :\ And that is why we stay enamored by kids. How do they learn what they learn? They never knew the language before; they never knew any of the concepts of this world before. How did they learn, then? Exactly? Why do people learn differently?
For punishing machine to be a vantage point, its suffering needs to be isolated to machine only, with no impact to any human in any way! When a machine makes a bad choice, it learns better for next time surely, but for that particular instance when machine was learning the lesson, perhaps a user paid the price. That is the trick.
The Winning Hand
Still the winning hand is with artificial intelligence.
There are some big ideas we are executing on for spreading financial prosperity. Some of these ideas we attempt to solve for by artificial intelligence have otherwise no known perfect solution for all possible scenarios that may occur. The data is too huge, too complex, of unknown new variety. There is a need for dynamism in algorithm itself, for machine to adapt to environment.The winning hand with artificial intelligence as in these cases the machine error rate will be far far below the human error rate., and machine success rates will be far far above the highest possible human success rate.
This decade will see many experiments with AI in lines where above hypotheses stay true.
Drawn on Google AutoDraw — My attempt at drawing a brain 🙂
Using Google AutoDraw AI to choose a better brain 🙂
Artillery For Artificial Intelligence
For every man who is called sexy, no one mentions the hundreds of grueling workout hours that went in to bring out the sexiness 🙂
So it is with famously sexy data science job, the outcomes of AI always seem sexy. Because intelligence is sexy! 🙂 But like any engineering field, it involves work and discipline.
AI in Steps
Most common steps towards creating artificial intelligence are –
- Know the Domain, what you are solving for
- Study the data — Data Mining
- Cleanse , Normalize Data, develop tools
- Choose a Model
- Test with Few Models —> Shortlist the Optimum Models — >Pick the best Model
- Train/Fine Tune/AB Test The Model
- Correct If Model Overfitting or Underfitting
- Quantify The Model — Monitoring Errors, Learnings, Positive Impact
Based on the understanding of the domain you are solving for and data knowledge, one is well equipped to select models that would work best. Some examples –
- k-Nearest Neighbors
- Linear Regression/Polynomial Regression
- Support Vector Machines (SVM)
- Decision Trees, Random Forests
- Hierarchical Cluster Analysis (HCA)
- Expectation Maximization
Recurrent Neural Networks
There are readily available algorithms for different modes of machine learning in different languages, platforms.Some examples –
The End Game
It is so perfect that it cannot be a coincidence.We are following the plan to the T!
Mathematics -> Algorithms -> Programming -> Big Data -> Stable, Cheap, Usable Cloud Infrastructure -> Artificial Intelligence -> ??
The decades spent in each were important, and so was the sequence.Without the work and the time spent in each precursor step to make it stronger, the next steps were simply impossible or weak to execute.
AI wave had come before too, but that was premature as both Big Data and Cloud Infrastructure were still evolving. This is the true beginning of the decade of Artificial Intelligence.This is the decade to make AI solid.For some it will be also the decade to wonder what comes next?
There is no time like present time to contemplate future 🙂
The algorithms, and programming we were doing for some decades now were imparting good amount of intelligence to machine. It was the beginning of having machines as an extension to our brains.But that was not artificial intelligence still.
When we programmed machines in a manner that intelligence of the machine seems to improve upon experiences, we arrived at “Artificial Intelligence”.
A machine can never do “X”
When we talk of artificial intelligence, we imagine a machine capable of everything a human can, and if it cannot what, and if it can, what all? This is a good way to set limits for near and long term AI expectations.
For example, some think, A machine can never do “X” where
X => display emotion
X => have fun with me
X => show intuition
X => be brilliant on day 0; falter on day 5; seemingly thoughtful on day 11; ambiguous on day 15; quiet on day 20; brilliant again on day 31;
But for all you know, maybe someone out there is working on exactly that..
Weak AI is a machine that behaves intelligently. Strong AI is a machine which behaves like a human. There is work ongoing in both.
Eliza And Alexa: A Sister Act
I had played with Eliza a decade back. Though it was only a NLP tool with stemming, tokenization, parsing and rhetorical conversation implementation, it made you feel as if you were talking to a human being, the best kind who listens and understands..
And coming of age to silky feminine voice of Alexa who does what you ask her to do if she understands you, and artificially gifted that she is, she gets better at understanding you over time. I hear females jokingly mention that Amazon Alexa pays more attention to their husbands than to them If machine can invoke jealousy, well, AI is getting stronger 🙂
AI trends increasingly show that our leaning to machines is not only for help but also for a feeling of companionship.
Machines have learnt to learn. That is a big deal.
Next, will the Artificially intelligent machines help evolve human learning? What do you think?
Mathematics -> Algorithms -> Programming -> Big Data -> Stable, Cheap, Usable Cloud Infrastructure -> Artificial Intelligence -> ???