Machines that Think and Feel

What is the Difference Between Machine Learning and Deep Learning

While Machine Learning (ML) Algorithms can learn and evolve gradually, they need guidance. For example, if the algorithm makes an incorrect guess, the human intervention must adjust. On the other hand, Deep Learning Algorithms can learn on their own whether their predictions are correct with the help of artificial neural networks.

Another way to distinguish between Machine Learning (ML) and Deep Learning is to look at how they learn. For example, you want to teach a machine to classify photos of dogs and cats. If you’re using the Machine Learning (ML) Model, it needs to be fed with structured data so that the algorithm learns certain features that differentiate the images of both animals. The algorithm has the information to distinguish between the two animals with each labeled image. On the other hand, the Deep Learning Model does not need to provide structured data or labeled images. Artificial neural networks allow algorithms first to detect and then comprehend the different characteristics of each animal.

After the images are processed through different layers of deep neural networks, the system develops specific descriptors to help classify animals and pictures. In the next step, the other outputs processed by each layer of the neural network are combined with categorizing the images effectively.

Another notable difference between Machine Learning (ML) and Deep Learning is the hardware requirements. The complexity of the calculations and the amount of data processed are relatively lower for Machine Learning (ML). Therefore, Machine Learning (ML) Programs can be run without much computational power. On the other hand, Deep Learning Systems require significant computational resources and graphics processing units (GPUs).

The time it takes to train Deep Learning, and Machine Learning (ML) Algorithms is also different. As you can imagine, Deep Learning Algorithms take time to prepare due to big data and complex computations. While training a Deep Learning System may take several hours or even weeks, a Machine Learning (ML) system can be introduced in a matter of seconds or hours.

Neuron and Deep Network

The picture above gives some idea mathematically. The smallest unit of Deep Learning is a neuron. (F = X1*W1 + X2*W2 + … + Xn*Wn) But we do know that insertion neurons give more accurate results for more complex problems. We call this a deep neural network.

Weight Calculator

Let’s consider a single neuron. Our formula is expressed as Z=wX+b. X is the input, b is the bias, w is the weight, Z is the prediction. Let’s remember this: Artificial Intelligence (AI) tries to find the most accurate variables w and b. I said most accurate because we have correct results for hundreds of X entries. The Artificial Intelligence (AI) ​​will generate Z outputs with a random value for w and b. Then he will see that the Z value is too far from the correct value. Then it will change the w and b slightly. Then he will look at the results again and say, “A very good, I came a little closer,” then he will change a little more in the same direction. This way will determine the w and b values ​​that can produce the closest Z value for the entire data set by changing it many times.

The graph above ideally shows the variation of the loss value according to the w and b values.

We talked about the learning process with a single neuron network so far. To solve more complex problems, we connect these neurons one after the other and side by side to form a deep network.

So far, we have focused on w’s, b’s, number of hidden layers, number of neurons in layers, etc. We call these Hyperparameters.

Data

First, we have to have a dataset, and that dataset should contain inputs and outputs.

There are some properties that the dataset must provide. First, it must be big enough. If you have a limited data set, you can enlarge your data set by using data augmentation methods.

In addition, the data set must completely contain the problem you want to solve; it must come from the same distribution.

Layer Types

So far, we have always talked about the layer type we call “Dense” or “Fully Connected.” As their name suggests, these can give good results in the face of flat layers where all neurons are interconnected and problems that can be considered linear. But apart from these, we can talk about a few more layer types in general.

The first of these is “Conventional Layers.“

For example, applying a filter to a picture means multiplying a shot in a matrix with a filter (linear product) matrix.

We call Artificial Intelligence (AI); We were giving you a neuron before, and you were trying to find the value of w and b in them. Now we provide a 3–dimensional matrix (filter). Calculate this and find the most accurate deals.

Again, equivalent to the previous layer logic; Just as we could give more than one neuron in a layer, now we can provide more than one filter. This method is called the “Conventional Layer.”

Another layer type is “Recurrent Neural Network. “

It is often used in situations such as predicting future elements. Stock market predictions word translations are the most famous areas. It also has a unique structure.

Output Layer

The output layer is now the layer that produces the expected result. Softmax is used in examples such as number recognition and classification.