The credit for the immense hype that Artificial Intelligence has created in last couple of years mostly goes to the advancements that happened in the area of “Artificial Neutral Network” and their successful implementations in different fields like healthcare & medicine, banking & finance, manufacturing & retail and many more. Although Artificial Neural Network as a concept is not new and has been prevalent since 70’s but the real implementation has only been possible over last couple of years due to various reasons.
Artificial Neural Network which is essentially inspired by biological neural superstructure of the human brain has many similarities if we believe that human brain functions in the way our findings has been put forward by the researchers. Artificial neural networks have demonstrated amazing capabilities in many cases and outperformed human brains in couple of instances like “ImageNet Challenge” where an AI algorithm can recognize different images better than humans, play complex games better than humans like “Google’s DeepMinds AlphaGo & IBM’s Deep Blue”, detect cancer better than a doctor and many more. But at the same time Artificial Neural Networks also have serious limitations. Even before going into those limitations lets first discuss how human brain works that we know so far.
There are so many things about the human brain that we don’t know, but of course there are few things that we do know. The most complex and advanced biological system that exists in the world is the human brain, which has immense capabilities compared to it’s size. If you believe, God be the creator of this universe, then this piece probably is the most sophisticated part of his creation.
The nervous system is the central part of the human body which transmits signals back and forth to different parts of the body that helps coordination of our voluntary and involuntary movements. At a cellular level the nervous system consists of a special type of cell called “neuron”. Arguably there are near about 100 billion neurons in human brain and they are connected to each other through a wire like structure called Synapse, which acts like a pathway for electrical signals to travel from one neuron to the other.
As per the information available, the protein synthesis occurs at the nucleus of the biological neuron called Soma considered to be the centre of human cognition. The signal emitted by the neurons are believed to be partly electrical and partly chemical which is a byproduct of the phenomena that happens due to the protein synthesis inside it.
Now to cut short, human neurons receives the inputs to be processed from the outside world through our five senses, i.e. ear, eye, nose, skin and tongue. Of course the inputs received through these organs are encoded into electrical signals before being send to the neurons to be processed. Interestingly, the biological neurons works on a threshold based principle. This means whenever the input signal received by a neuron after it’s chemical synthesis crosses a particular threshold limit, it fires the signal to the next neuron. The next neuron then takes this signal as it’s input and does the same process which continues through millions of subsequent layers of neurons to arrive at a conclusion on the piece of information that been processed. Wow, isn’t it amazing !
Have we ever even realized that such complex processing happens inside our brain even when we just try to differentiate between a cat and a dog in our daily life. This is just an example but human brain has many beauties and you will be amazed as you try to explore more and more about it.
If we draw an analogy between the internet with the human brain and if we consider a neuron in our brain as a webpage in the internet; a human at any age has around 100 billion neurons in the brain but internet has 10 times more, around 1 trillion webpages. Of course internet is a much bigger entity than a tiny human brain. But lets now compare the complexities of both in terms of number of connections or you can consider it as number of hyperlinks. Internet has over 100 trillion hyperlinks and an adult’s brain has around 300 trillion which is 3 times that of internet and amazingly a child’s brain has 10 times the number of connections in the entire internet, 1000 trillion connections.
Scientists have revealed that child’s brain at a very early stage from zero to one year used to have around 100 billion neurons and only few million connections. But these connections exponentially multiplies during the development stage of the brain till the age of 6(six) years at the rate of around 1000 (thousand) connections per second. That’s how a child learns and the learning process is essentially through building more and more connections among the neurons.
So when you teach something to your child or when he or she learns new things by observing some events or incidents, in the process, millions of new connections are established immediately among the neurons in your child’s brain. Most of these connections at the learning stage believed to remain intact and are not erasable throughout lifetime. That’s probably the reason we are always told not to teach wrong things to our children as it is going to be very difficult to alter or erase at a later stage. That’s my personal view !
An interesting fact about the human brain is its constantly changing, constantly evolving and establishing millions and billions of new connections, redefining existing connections and removing irrelevant connections among the neurons across different parts of our brain. This is the process how human brain learns new things, adds experience to those learnings and draws conclusion on the incidents about our life.
But what about imaginations? How we can imagine something which we never had seen or experienced ? How we can relate our experience from one incident to a completely new situation which we have never experienced in our life. How will you explain these amazing capabilities of our brain. How about love, affection, empathy and our feelings for others. Can those be mimicked inside a semiconductor? Probably yes but not anytime soon.
Back again to Artificial Neural Networks, they have got numerous advantages but they also have many limitations. These limitations can be broadly classified into two categories — Scientific or Mathematical limitations and General limitations.
Scientific or mathematical limitations includes issues with- “Piece-wise Linear Curve”, “Flat Activations”, “End-To-End-Learning”, “Learning Many Orthogonal Functions” and “Universal Approximation Theorem”which are beyond the scope of this article. But there are other general limitations too like requirement of huge relevant dataset for training deep neural networks; requirement of huge computational power for training and substantial time requirement for training such models which you probably all are aware of. Again hyper-parameter tuning for the deep learning models is sometimes a tedious process although some automation of such activities are now available but still needs maturity.
To overcome the above mentioned limitations researchers has also suggested few alternatives such as “Geoffrey Hinton’s Neuron Capsules”, “Yann LeCun’s Energy-based models” and “Zhi-Hua Zhou’s Deep gcForest” but all those are yet to hit the market in big way.
But more than all the above, Artificial Neural Network and for that matter the AI itself is lagging the most important and the most aspired capability what is called “General Intelligence” which is far beyond what we have achieved with AI till now, the “Narrow Intelligence”. Lets understand what exactly that means.
An AI system can be far better than human in performing some activity it is made for but it doesn’t have the awareness what it is performing. Also it can’t bring in the experience from some other context to the current context or relate experiences. Human brain can do this very well and many more amazing things which is referred as “General Intelligence”.
Good news is scientists at Karolinska Institute known as the (Royal) Caroline Institute of Sweden in a paper titled “An organic electronic biomimetic neuron enables auto-regulated neuro-modulation” has claimed to build a fully functional neuron by using organic bioelectronics. This artificial neuron contains no living parts but it’s capable of mimicking the functions of human neuron cell and communicate in the same way as our human neurons do.
As an AI professional and researcher I strongly believe that we will surely be able to reach to a stage where our machines will have General Intelligence not necessarily as capable as human brain. Trust me that day we will be able to address many unsolved issues for humanity with AI and that will be truly an era of “Man and Machine Partnership” for the humanity.