Machine learning: The new cult in the town or an entirely new world?
AI is dealing with its own internal problems and it is helping us to deal with/manage various real-world problems. Is artificial intelligence an underdog or it has reached its full potential
The need for this new world
(from stable diffusion Prompt: The whole new planet of artificial intelligence)
The study of artificial intelligence, engineering an intelligent machine to a given problem is a cult with rising popularity. We are exploring new scopes to apply these discovered and engineered synthetic intelligence methods and this exploration is a whole new world in itself. At this point, we can say this world is neither naive nor limited in terms of opportunities, exploration, application, discussion, or even disciplines. Artificial intelligence has enormous power and potential to be applied to fields ranging from cinema to aerospace, from scientific discoveries to products and services that we need every day.
This is not the journey where we have roads to explore the known this is a journey where we have to find the path to the unknown. To explore that unknown we are having tools of the present known. One obvious question is if we have tools of the present known why do we need to explore that unknown? Why should we go to find new paths, new places, new worlds, and new possibilities? The answer to that is our present knowns and present tools are not enough to solve all of our problems
Artificial intelligence is dealing with its own internal problems and it is helping us to deal with/manage various real-world problems.
Is artificial intelligence an underdog or it has reached its full potential? If artificial intelligence is powerful then how much and what makes it stronger?
The sun of this new world
(From stable diffusion prompt: “The whole new world of artificial intelligence with neural networks as the sun“ )
If artificial intelligence is a whole new world then neural networks are the sun of that world. A strong force that powers the whole planet of artificial intelligence. We can think of neural networks as thinking machinery, like any other machinery it has its own mechanisms. This mechanism tries to learn from a given data. To learn from data is to understand data. To learn is to refine your understanding of that data in iterations. The success of this learning of neural networks is judged by its ability to predict optimal answers for questions that the network was not provided during learning and these new questions are related to questions on which the machine was trained on. If that neural network has learned well then it has past datapoint to answer these new questions.
The currency of this new planet
(from stable diffusion prompt: data as the currency of artificial intelligence)
Data is indeed the currency of artificial intelligence. Datasets are records of events that contribute to the outcome of that action or step that the network needs to take to solve the problem. Let us just consider the problem of self-driving cars. Which events contribute to the outcome of driving? We, humans, drive by our past experience and our ability to see and observe the physical world around us. This ability helps us to make decisions like when to stop a car when to slow the car, when to overtake and when not to. Driving a car has two aspects one is a human's ability to see using their eyes and the other is a learned experience that analyzes the scene and takes the optimal decision. Now for Artificial Intelligence to learn driving it has to learn to analyze the scene and understand how to drive a car from human driving videos. The collection of human car driving videos will be the dataset of this problem to the neural network.
Neural networks work on datasets that have some correlation between their inputs and outputs. If this condition of correlation between its input and output is true then there are always some patterns to that data, if known we can easily predict the output even for inputs to which outputs are unknown. Neural Networks utilize this same property. Neural Networks attempt to find the correlation between inputs and outputs. They attempt to map input and output with a mathematical function. If neural networks find the right function then they can always be right in a given problem no matter how unpredictable it may even seem to the human mind.
Any data has multiple variables for example in the car driving video one variable is road another one is road obstacles, traffic signs, etc. In order to correctly analyze the scene machine needs to distinguish these variables and classify them into categories.
Data are broadly divided into 2 categories. First is linearly separable data, the data which can be separated using a straight line. This type of dataset is a fairly simple dataset without any complexity. Let's say we have data of students' grades on one axis and we have a pass or fail on the other axis we can clearly draw a line on the graph which will separate failed and passed students. Using this line we can clearly categorize data. The second category is non-linearly separable data. The data can not be separated using a straight line. This is a complex dataset. Our self-driving car dataset falls in this category.
Power of the sun
(from stable diffusion prompt: “neural networks shining as bright as the sun“)
An Artificial Neural Network is a mathematical function that takes some input and returns an output. This mathematical function has many individual mathematical units called neurons. Different neurons are connected to each other with different connection strengths (weights), the greater the strength stronger the passage of signal from one neuron to another. The neural network has 3 main components one is the input layer, the input layer is the section of the neural network that is responsible for taking inputs. The last section of the neural network is the output layer, the layer which returns the output. In between, there are hidden layers. Neural networks learn to perform a particular task in the process of training. During the training process, we give neural networks input and they return output, and we judge the accuracy of network output by comparing it with the expected output. If the network’s output is inaccurate we measure the inaccuracy and tell the network about that inaccuracy. Networks adjust their internal parameters (those mathematical functions ) according to the inaccuracy and by the end of the process, it learns to perform that task with the desired accuracy.
If neural networks are this capable then they can learn to perform any task given a dataset. Then reality is more powerful than any sci-fi imagination/prediction. In theory, it is the case but in practicality, it is not. But why, what are those challenges that are limiting us from doing it, utilizing all potential of a neural network theory?
Now is the time to return to the question that we asked earlier how powerful is a neural network? Before answering this let me reframe the question, When will the network converge? When will a network find desired output to a given input? In the case of linearly separable data answer is simple. The perceptron rule says for a linearly separable data network without any hidden layers can find optimal parameters that can predict output to all inputs accurately. Given any linear data neural networks without any hidden layers can learn to predict data with the desired accuracy
But now what about non-linearly separable data? For non-linearly separable data, we have a delta rule which tries to find optimal parameters for a network in given non-linearly separable datasets
In the real world, we have linearly separable and non-linearly separable data, we know what kinds of neural networks converge on linearly separable data, and for non-linearly separable data. So for a neural network to perform a given task we need a dataset that describes the task, decides what will be the input to the network and what outputs we expect and we have a really powerful tool that learns to perform that task
This is going to be a series of blogs where I will be discussing Neural Networks. I try to start with intuition, real-world relevance, underlying questions, mechanism, and theory of that topic so stay tuned. Subscribe to my newsletter so that you will not miss any blogs