# Need For Deep Learning: Edge Over Classical Approach

## Explaining the need for deep learning networks to solve the modern age problems using data

## Best-In-Class Performance

Deep networks have outperformed traditional ML techniques in numerous areas, including speech, natural language, vision, and gaming, with accuracy levels that are significantly higher.

For instance, the graph below illustrates the picture classification accuracy of various approaches using the ImageNet dataset; the blue color denotes traditional machine learning (ML) methods, while the red color denotes a deep convolutional neural network (CNN) method. Here, deep learning completely defeats traditional ML.

## Scales Effectively with Data

When there is more data, deep networks scale considerably better than traditional ML algorithms. The straightforward yet potent illustration of this is the graph below. More data is frequently the greatest recommendation when trying to increase accuracy with a deep network. This quick and simple repair doesn't perform nearly as well with traditional ML algorithms, and more sophisticated techniques are frequently needed to increase accuracy.

## Adaptable and Transferable

Unlike traditional ML methods, deep learning approaches can be used in a variety of domains and applications. First, using pre-trained deep networks for various applications within the same domain is now efficient thanks to transfer learning.

For instance, in computer vision, object recognition and segmentation networks frequently use feature extraction front-ends that were trained on pre-trained image classification networks. The full model's training is facilitated by using these pre-trained networks as front-ends, which frequently leads to better performance in a shorter amount of time.

In addition, as speech recognition uses many of the same deep learning concepts and methods as natural language processing, understanding how to apply deep networks to NLP isn't too difficult given the same foundational knowledge.

## No Need for Feature Engineering

Classical ML methods frequently necessitate significant feature engineering. Typically, exploratory data analysis is performed on the dataset initially. A dimensionality reduction may then be performed for ease of processing. Finally, just the best features should be passed on to the ML algorithm. When employing a deep network, there is no need for this because data can be passed straight to the network and usually achieves good performance immediately away. This completely avoids the lengthy and difficult feature engineering stage of the process.

# What Is Deep Learning?

Deep Learning is a branch of Machine Learning that use Neural Network ideas to handle highly computational use cases involving the processing of multidimensional data. It automates the feature extraction process, with very little human participation. A neural network is essentially a collection of neurons and the connections that connect them. A neuron is a function with many inputs and just one output. Its job is to accept all of the numbers from its input, apply a function to them, and then deliver the result to the output.

# How Does A Neural Network Work?

To grasp neural networks, we must split them and understand a perceptron, which is a neural network's most fundamental building block. To categorize linear data, a single-layer neural network is utilized. It comprises four vital parts:

- Inputs
- Weights and Bias
- Summation Function
- Activation or transformation Function

The following is the underlying logic of a Perceptron: The inputs (x) from the input layer are multiplied by the weights ‘w’ applied to them. The weighted sum is formed by adding the multiplied values. The weighted total of the inputs and their weights is then applied to the appropriate Activation Function. The activation function converts the input to the desired output.

## Weights and Bias

When an input variable is sent into the network, the weight of that input is allocated at random. The weight of each input data point reflects its significance in predicting the outcome. The bias parameter, on the other hand, allows you to fine-tune the activation function curve to produce a precise output.

## Summation Function

Once the weights are allocated to the inputs, the product of the input and weight are calculated. The Weighted Sum is obtained by adding all of these products. The summing function accomplishes this.

## Activation Function

The activation functions' principal goal is to map the weighted sum to the output. Transformation functions include activation functions such as tanh, ReLU, sigmoid, and others.

# Neural Networks Explained With An Example

Think of a situation where you need to create a CNN that can divide images into two categories:

- Class A: Contains dog pictures
- Class B: Images of creatures other than dogs

So how do you build a neural network that can distinguish between dogs and other animals?

The first step in any process is to prepare the input for processing by processing and converting it. According to the image's dimension in our scenario, each dog image will be divided into pixels.

For instance, if the image has a resolution of 60 by 40 pixels, there will be 2400 pixels altogether. The input layer of the neural network receives these pixels' representations as matrices.

The perceptron in a CNN works similarly to the neurons in our brains in that they absorb input and process it by sending it from the input layer to the hidden layer, then to the output layer.

Each input is given a starting random weight transmitted from the input layer to the hidden layer. Following this, the inputs are multiplied by the relevant weights, and the sum is supplied as input to the following hidden layer.

Each perceptron in this situation has a bias value assigned to it, corresponding to the weighting of each input. Additionally, each perceptron undergoes an activation or transformation function that decides whether or not it will be activated.

Data transmission to the following layer takes place via an active perceptron. The data is transmitted via the neural network until the perceptron gets to the output layer (forward propagation).

The output layer determines whether the data belongs to class A or class B by deriving a probability.

If you liked the article, please help share and spread the word. Also, don't forget to connect with me on Twitter.