Understanding how machines learn and why AI is not intelligent
With the advancement of technology, there are now machines and bots with artificial intelligence that are created to do many things that are common to humans. These machines are created with intelligence and the ability to learn just like humans. But while we know that humans are taught the things that they know through various learning platforms, and as well as from their life experience, how do these machines learn?
We know that machines are programmed to carry out specific sets of instructions to do certain tasks. But these machines are now able to learn on their own, through a process called machine learning.
This is the use of artificial intelligence to provide computers with the ability to learn automatically and also to improve automatically from their experience to do things and know things they are not pre-programmed to do and know.
To simply explain what machine learning is, it is a combination of the human ability to make crucial decisions, analyze, and learn from their experience with the ability of the machine to take care of numerous data. Machine learning is basically about adding a human touch to the machines.
Parts of Machine Learning
Machine learning has different parts and elements to it. To completely understand what machine learning is and how it works, you need to understand all of these elements. The elements include data, task, model, loss, learning and evaluation.
Data: This is the first step to machine learning. Machine learning can’t happen without a datum first. It is what the machine acts upon. It’s like an input to the machine, and without input, there is no output.
There are a number of data types in machine learning. The data can be a table with numerical values in it, videos, text, audio, images, or from many other probable sources. Whatever data type it is, different methods (such as speech signal frequency, the value of RBG per pixel in a picture, etc) can be used to encode them. This improves the data and makes it a high dimension.
To acquire the data, there are also a number of methods that are also used. A website with a large database, such as Google AI can be used. You could also use freelancing websites, where you can easily get people to provide you with paper writing services, but in this case you will hire the people to help you acquire data. For instance, you could give them hundreds of images containing different foods and food products with an instruction to write specific details such as the food ingredients, information on allergy, weight, etc.
Task: Now that you have your data, the next step in machine learning is to assign tasks. You assign this task in a way such the data can be used in a lot of ways. You can use the data to determine the relationship between 2 data of different texts. This relationship is then used to produce other data such as using email content to determine what the automatic response will be etc.
There are 2 types of tasks; the supervised task and the unsupervised task.
The supervised task is one in which you are responsible for supplying the data, which is the input, and the result of performing the task, which is the output. This is so that the machine learns this pattern and is able to use your method in other cases. This type of learning is mostly responsible for most of the success that has been recorded in machine learning, and it uses both multi-class and binary classifications.
On the other hand, the unsupervised task is one in which you only supply the computer with input, then depending on the algorithms or a number of rules, it is able to generate meaningful output.
Models: “y=f(x)” is the relationship between the input data and the output the computer generates given that y is the output and x is the input. When it comes to finding the functions that are between different data, it is sometimes easy like y=mx+c, which is the straight-line equation. But in most cases, it is more difficult than that and might even contain variables with orders greater than 25.
The machine makes use of these complex equations for calculating the equations and getting its parameters’ values. When it comes to making a model, the terms used include regularization and overfitting.
Loss function: Functions can be derived and slotted in as values of input and output (x and y). These parameters can, however, have changing values based on the user type, and this, therefore, leads to developing different models. A method had to be devised to be used in determining the model that’s the best from the solutions that are available.
The loss function method is the method that’s usually being used and it is L = Sum. This is the square of the result of subtracting experimental output from the actual output.
You can take the best model of the several loss functions computed for use. There is loss such as KL divergence, and loss of Square error.
Learning algorithms: This is actually where the problem of searching for the model’s parameter is solved. It is not feasible to search with brute force; especially for models that have a high order. The reason for searching for these parameters’ value is for the optimization of that loss function. Algorithms such as Adagrad, Adam, RMS prop and Gradient Descent can be used.
Evaluation: This is the last stage of machine learning. After going through all the stages before this, a system of the benchmark is needed for the different models of machine learning. To evaluate this, the predicted output is compared against the true output. This then gives the percentage of accuracy. A second approach is to choose a little number, which contains the answer, from a huge amount of output.
F1, Recall and Precision are some elements needed for this method. At this stage, different tests are carried out to determine how each model performs. This is because of the different approaches involved in determining the solution, and you only require the best of them all.
Is Artificial Intelligence Really Intelligent Then?
Artificial intelligence is centered on the creation of computers that have the ability to carry out different tasks that would ideally require the intelligence of a human, such as decision-making, language translation, recognition of speech, and visual perception, etc.
Going by this definition, it is hard to argue that artificial intelligence is really intelligent. This is because the machines are only created to follow a set of rules or instructions to carry out the tasks that they do.
For instance, the machine might be able to make certain decisions and perform specific tasks, but it is unable to tell what the task it is carrying out really is or why it has to carry the task out or what the purpose of the task it. As a matter of fact, it cannot even say what it is doing. Yet, it can use a set of algorithms to do the task, and even do it faster and in a more efficient way than humans. Is that really how intelligence is defined? Being able to follow a set of rules to do a task but not really knowing what the task is or being able to think out of the box?
The point is, there is a problem with artificial intelligence and it is very limited. Computers are pre-programmed to follow a particular algorithm or set of rules to do what they do. They are still confined within those algorithms. You might argue that the machines learn. No doubt they do. They are able to take in numerous data, analyze it and try to learn from it. But the only reason they can take in this data and learn from it is that they are programmed to. They are still limited by their algorithm and also the data they have been exposed to.
Unlike human intelligence that learns both consciously and unconsciously using its senses, these machines cannot feel. They are programmed to do what they do and built to achieve results. They are also not able to develop emotional intelligence like humans which actually makes them weak.
You might be able to replace a human worker with a machine that is very efficient and super productive. But in reality, they are just built to achieve certain goals and carry out some tasks faster than humans probably can. But they are not intelligent on their own and can never be. The fact that they are automated doesn’t make them intelligent. That is the mistake that many people that believe in AI make.