Replicating a Human
In Feb 2016, we saw the first live try of this, when Hanson Robotics launched first AI based social humanoid robot Sophia. It was a good try, but AI is still very far from its final destination. Well, let’s try to figure out how far is AI from that destination.
So, what make the humans complicated? Is it the ability to sense, the ability to think or the ability to perform?
Let’s first focus on the ability of humans to sense. It can be broken down to 5 basic senses touch, sight, hearing, smell and taste. If we try to look at Sophia, it has tried to replicate 2 of theses senses i.e. sight and hearing, majorly focusing on conversations between humans. Do these 2 senses matter the most or the other 3 are very complicated to replicate? Even AI can be divided into 2 major categories i.e. “Strong AI” & “Weak AI”. Strong AI specifically focuses only on automating intellectual tasks which humans can do e.g. detecting emotions, analyzing real time videos, understanding sarcasm etc. Since the introduction of neural networks & deep learning, world of strong AI has totally transformed. Most of the strong AI solutions are based on deep learning. Let’s see where Strong AI has reached in replicating human vision and understanding natural language capability.
Vision : It’s so easy for us to identify an object e.g. a door, an apple, a car etc. but when we tell a machine to do the same task, we realize the challenge. Even trying to find from where to start is challenging. Well, fortunately research is going on it from a very long time and AI has came very ahead in it. In fact, we have coined out a special term for this very task of ‘machine understanding images’ i.e. Computer Vision. Every big tech giant is working on their own solutions and algorithms for this. Google had recently launched its application “Google Lens” which can literally give you the more information about an object than a human can. But, that only makes the machine more informative not more intelligent. For a predictable challenge, machine may even surpass human. But, is it really possible to train a machine to reach the same level of understanding a human can get from an image? Do we even have the database for that? There are many areas of Strong AI where computer vision is trying to transform the things, but is unable to reach upto a satisfactory level. One of them is autonomous driving. Even though, a lot of research is going on this, it would take a huge amount of time to bring it on production level.
Natural Language : Since the introduction of word2vec, which enabled us to capture the semantics of the sentences, AI started to transform the world of Natural Language Processing but not upto the level of humans. It was upon the introduction of LSTM cells and CNNs, which brought the AI closer to solve the intellectual natural language problems like information summarization, information extraction, chat-bots for customer services & assistance etc. We even have the products like Google Alexa, which can directly intake the data in raw form and perform the task. But all these are limited to some specific tasks. One of the way to check how far Strong AI has reached as compared to human, is a Turing Test.
Mitsuku, continuously 4 times winner of the Loebner Prize Turing Test, is said to be the best conversational AI. Chat with it here. While it performs very good at casual talks and jokes, the difference becomes clear when sarcasm comes into play. Researchers are predicting, conversational AI may reach upto level of humans by 2029.
Conclusion
We can never expect a accuracy of 100% from AI/ML solutions. In case of problems where human intellect is required, expectations fall more. Although, accuracy and applications of strong AI are increasing very rapidly. But, we are observing developments only in very specific tasks. Training the robots to predict the unpredictable tasks where human intellect is needed, is the most challenging aspect of Strong AI. Nobody is sure, whether robots can really reach the human level of intellect or not.
