Neural Nets and the Human Mind
If you are familiar with the terms Artificial Intelligence, Machine Learning , Deep Learning etc , you might know that these so called “Machine Learning Algorithms” have taken the world by storm , infiltrating every possible sector of modern technology and making their implementation rather an obligation than innovation. Such is the power of artificial intelligence. Now some may argue whether the machine learning algorithms we use today are autonomous enough to be called AI , but that is a very broad debate and not the focus of this article.
In this article, I have tried to explore the deep intricacies and relationship between neural networks and the human mind.
How is it that when you start typing something on Google , the browser already guesses what you’re looking for? How does your mobile camera recognize a face in the vast ocean of minuscule pixels that we call a picture?How does a virtual assistant like Alexa, Siri or Google Home achieve complex tasks like scheduling meetings, fetching weather information, reading bedtime stories, etc ? The answer to all of this, as you may have guessed - is Machine Learning. More specifically, Neural Networks.
What is a Neural Net?
The purpose of this segment is to teach someone completely new to Deep Learning what a basic Neural Network is. Although the actual math involved is way more complex, the highlight here is not how they work , but to try to understand their connection to the human mind.
The above is the representation of a neuron- the simplest unit in a neural network. What a neuron does is, it takes in some input values from all the neurons in the previous layer (in this case the previous layer to our neuron(y) is the blue one i.e. x1, x2 , x3..) and multiplies each input with the weight associated with the neuron the input is coming from. So here, x1 is multiplied with w1 , x2 is multiplied with w2 and so on. All these multiplications are added inside the neuron y.
y=x1.w1 + x2*w2 +x3*w3……xm.wm
This summation is then converted to an output by using activation functions such as sigmoid, tanH , reLU etc. What are these activation functions? Let me explain by giving you an example of the sigmoid activation function.
The sigmoid function converts any input (in our case , the summation-y is the input to the sigmoid function) to a value in the range 0 to 1. As simple as that. You add the products of the inputs and the respective weights , pass them through the activation function , and voila! your neuron has given you an output. Now that we know about neurons , let’s see how networks use them.
This is as simple as neural networks get. Each grey circle here is a neuron and as you can see, there are three layers in this network -the input layer, the hidden layer and the output layer.
Now neural networks learn to do tasks by themselves, but how do they learn? through data of course. We will discuss more about the learning from data in the next segment but know for now , know that data is fed to the neural network through the input layer , processed in the middle layers and then outputted by the output layer.
For example if you want a network to recognize cats and dogs from images, the labelled images(labelling images means that we give the neural network the correct answer) of the cats and dogs are fed to the neural net, processed , and the layer makes a guess as to whether it is a cat or a dog. Now since it has the answer to the image , it adjusts the weights between the neurons in such a way, that its accuracy increases.
The Human Mind- A giant neural network
Have you ever observed a toddler playing? How they are always so curious and so hungry for interaction with the world around them. The brain of a toddler is like a simple and untrained neural network which starts learning as soon as they are born. Since the network is very simple in early stages , the first things that the toddler learns are accomplishments of very basic tasks such as moving, eating, crawling etc. For example-
When a toddler holds a ball, he observes it. Its weight, its shape, how he feels holding it , the curvature around his skin etc. This observation is part of his initial learning process . This observation is in fact ,what helps him learn. The experience of the toddler with the ball trains his brain for the next time he encounters one. Now when he sees a ball, he remembers what it is and how it will feel to hold and throw it. As the child plays more and more with the ball, his training builds . A teenager is almost an expert at handling and throwing balls and although one may think of throwing balls as a very simple task, the learning that you had to go through for it, was anything but simple.
As i said earlier, the brain of a child is like a simple neural network achieving basic tasks like walking , eating , speaking etc. As the child grows, he/she learns more complex processes such as hand eye co-ordination , speaking fluently, eating without spilling etc. People with experience in training neural networks will agree that it is relatively easy to train a simple network , and as you add more layers and make the data and the end output more complex, the learning rate slows down. Childhood is the best time to learn a new skill because the simple neural net which is the brain is relatively untrained as compared to an adult and thus it becomes easier to bend the neurons to your will. When the same child grows up, the learning becomes difficult with the complexity of the task. Complex problem solving, decision making and judgement making are few examples of complex tasks that even we as adults do not stop learning and building upon.
Okay so now we know that the brain learns all the time. But how exactly does this learning happen?
The answer is- the brain rewires itself. Just as the weights and connection between neurons change in a neural network, the brain also modifies the connections between its billions and billions of neurons to give rise to complex networks that have the power to not only retain the training they have undergone earlier, but to also build on that training , and to increase the complexity of the end process.
Even as you are reading this article, every neuron in your brain is changing itself, modifying it’s connections with the billions of other neurons , building on the experience you have gained since you were born. This is how learning happens. Every time you learn, your brain rewires itself! This process is just like the training of a neural network.
Data is everything
You must have heard the term ‘big data’ and you might know that today, nothing is more important to tech giants than data of the consumer base. Why is this so ? Because data is what drives the ‘machine learning algorithms’ . Because data is what trains the neural networks, because without data , learning is impossible.
You must have come across verification captchas on the internet like the one below.
Answering the above captcha might not seem like a big deal to you but ask a child to do it and although they may get it right, the answer will come slow.
Fun fact: did you know that every time you answer a captcha like this, you are helping to train the bots or the neural networks in place, all around the internet! You are in fact helping prepare data to train the models that actually serve you! Next time your ML enthusiast friend starts bragging about the models he has made, tell him that you have trained big and important models yourself.
Just like the data we prepare for our artificially made models is of utmost importance , the human mind is also a greedy network wanting to be fed data, in huge amounts, every single moment of your life. The data is building you, helping you learn , turning you into the intelligent machine you are. And this brings us to the final aspect of a human mind learning from data around it, the input layer or the five senses.
The five senses
We know that a neural network has a input layer , through which the data enters the network in a form that the model can understand and process such as- tensors . Data then passes to the hidden layers where the training happens , weights change and finally the output layer gives out a prediction whether it be a cat/dog classification , or completing a sentence based on previous words.
Since we have compared the human mind to a giant neural network, it is only fair if this network has a input layer as well. The five senses of sight, sound, touch , taste and smell make up the input layer to our mind.The sense of sight is the most important of all the five senses, because the maximum variation in data comes through this sense. A good dataset to train your neural network on is one which contains different types and examples of data rather than monotonous data. The latter type can cause overfitting to the model which means that the model may give correct results when the data is of the same type as the one it was trained on, but when data deviates from the training data norms, the accuracy takes a severe hit. Our eyes are the most active when we are learning most of the important processes in our life. From walking to eating, from learning to ride a bicycle to playing a piano, from interacting with other fellow humans to typing on a computer , almost every task requires our eyes. The sense of sound is also pretty important but it was more so back in the days when humans had to hunt for food in the wilderness focusing their ears for sounds of prey because their survival depended on this sense. Touch is also among the most valuable input senses that humans possess. It keeps our hands away from hot stoves, prevents us from feeding boiling milk to our babies and helps interact with every object we come into contact with. Although the sense of smell and taste do not contribute as much to the major learning of a human neural net, they have their own value. These senses contribute more to the finer pleasures of life than the growing and learning process. The smell of fresh flowers , the taste of good wine and such finer experiences of life come through these senses of smell and taste.
Now that we know about the input layer , the data , and the neural network where all the processing takes place, let us take an example putting all the pieces of the puzzle together to see the bigger picture i.e. pattern recognition.
Pattern Recognition
Let us take the example of learning to read. The eyes are the most important sense here. A child learns to read by learning the alphabets first. He does so by repeatedly looking at the letters,trying to pronounce them and writing them. Each epoch of this reading, speaking and writing trains his mind, improving upon his accuracy until he perfects the art of writing the alphabets. After his network has undergone this relatively simple training, he now learns to string these words together , and here , the sense of sound takes a major role. When he speaks these works , he starts recognizing patterns in these words, how different combinations of the same set of 26 letters give rise to the multitude of words making up the English language. This pattern recognition is very important because this is what helps a child recognize a human face or an animal or a toy. In adults , these pattern recognition becomes very very powerful. The vast amount of training they have undergone and the vast amounts of data they have processed helps them see patterns all around them. Predicting whether it is going to rain based on past experiences with rain is pattern recognition. Connecting memories and experiences to predict the outcome to an event is pattern recognition. Sherlock Holmes finding out that the man in the caravan is Professor Moriarty just by looking at the dust of chalk on his sleeve is pattern recognition!
Conclusion
I would like to say that though i have first explained how neural networks work and then compared the human mind to them , it is actually the other way around. The human mind inspired the neural networks. Why do you think those tiny functions are called neurons anyway. The human mind is an incredibly complex structure that produces every thought, action, memory, feeling and experience in this world. It is nothing short of miraculous that evolution gave us humans such a powerful tool which we today try to replicate into artificial minds ,which we otherwise call- neural networks.
