Introduction to Deep Learning

Prerna Aditi
7 min readJul 7, 2018

--

Deep Learning is an application of Machine Learning (based on Neural Networks) that deals with deep neural nets and intricate algorithms encouraged by the structure and role of a human brain. It is a machine learning technique that trains computers to accomplish what comes naturally to humans: learn by example.

One of the best examples of Deep Learning could be driverless cars. Being key expertise behind driverless cars, it enables them to identify any stop sign or differentiate a lamp-post from a pedestrian. It is also used for voice control in devices like TVs, tablets, phones, and hands-free speakers. Since it has the ability to obtain those results that weren’t achievable before, it’s gaining popularity day by day and that too for a good reason.

image source: fotolia.com

With the help of deep learning, any computer model can easily learn to classify straight from the reverberation, content, or pictures. State-of-the-art accuracy can be attained by deep learning models, and can sometimes exceed the human-level performance. With the help of a large set of labeled data and neural networks, architecture models are trained.

· Why do we need deep learning?

1. Feature Extraction: Deep learning algorithms take large volumes of data as input. After this, it analyzes the input for extracting out features of an object and then identifies similar objects. Thus, it avoids the manual procedure of feature extraction.

2. Performs Complex Algorithms: Using deep learning we can deal with complex algorithms.

3. Processing Huge Amounts Of Data: It can work with an enormous amount of both, structured and unstructured data. The more we have data (labeled or reference data) the better the system will do. When dealing with such a huge amount of data we need to ensure good performance of the model and so we require deep learning for this.

At higher levels, Deep learning attains identification accurateness than it had before. With the help of this, consumer electronics can meet user expectations and it is important for security-vital applications such as driverless cars. Recent advances have progressed deep learning and thus it works better than humans in some jobs like classifying entities in images.

Though deep learning was introduced in the 1980s, there are two major reasons why it has currently turned out to be useful:

1. Huge amount of labeled data is required by Deep Learning. For example, millions of images are required by driverless cars along with hours of video.

2. Considerable computing power is also required by deep learning. Comparable structural design of high-performance GPU is competent for deep learning. This facilitates the development group to lessen the training time for a deep learning network when combined with clusters or cloud computing.

· Applications of Deep Learning

Deep learning has its applications in industries from automated driving to medical devices.

1. Automated Driving: Deep learning is used by automotive researchers for identifying objects such as stop signs and traffic lights automatically. Moreover, deep learning is used for detecting pedestrians. This can help in reducing accidents. One best example of it is Google cars. These type of automated cars are actually fed with video of surroundings and is supposed to determine if there are any obstacles or any cars or if it's driving in lane etc.

2. Medical Research: Deep learning is used in cancer research for the automatic detection of cancer cells. Suppose there is a patient who needs to be diagnosed if he/she has cancer. For this, we need a cancer specialist. But we can’t find the specialist in every hospital. In such a case, things can worsen. To help in solving this issue, deep learning can play a tremendous role. If we use deep learning here, then the system can detect the disease or perform initial screening quite easily and automatically, without waiting for any specialist. An advanced microscope has been built by Teams at UCLA which yields a high-dimensional data set used for training a deep learning application for identifying cancer cells accurately.

3. Robotics: Deep learning is used in robotics for training robots to act like humans. Nowadays, robots are everywhere. They are both knowledge-oriented as well as industrial robots.

4. Electronics: Automated hearing and speech translation uses Deep learning. For example, deep learning application is used in home assistance tools that act in response to our voice and know our preference.

5. Machine translation: We have a lot of information these days and sometimes they are in a particular language. It becomes quite tough for humans to translate each and every information or any document into all possible languages. Suppose we go to any country and we come across a signboard. But we don’t know what is written over it, as we might not know the local language. So, we can use deep learning for this task. We have an application that uses deep learning which can do the task of translating the information to our preferred language.

6. Industrial Automation: Deep learning can be used around heavy machinery for ensuring the safety of workers by detecting if people or any objects are within the unsafe range of machines.

7. Aerospace and Defense: Using deep learning objects can be identified from satellites that locate and spot secure or insecure zones for troops.

And there are many more applications of deep learning.

· How Deep Learning Works

Since, most of the deep learning processes apply neural network architectures, which is why we refer to deep learning models to deep neural networks. The term deep generally refers to the number of veiled levels in the neural network. There are 2–3 hidden layers in the traditional neural networks, whereas the deep networks may have as many as 150.

Large sets of labeled data and neural network architectures are used for training deep learning models. They learn directly from the data without the help of manual feature extraction.

Convolutional neural network (CNN or ConvNet) is one of the most popular types in deep neural networks. A CNN convolves discovered characteristics with input data and uses 2D convolutional levels, making the architecture suitable for processing 2D data, such as images.

The manual feature extraction has been eliminated by CNNs so that there’s no need for identifying attributes used in classifying the images. The CNN extracts characteristics straight from images. The significant features are learned whilst the network guides on a set of images. Deep learning models are highly accurate due to this automated drawing out feature for computer vision jobs such as entity categorization.

· Difference Between Machine Learning and Deep Learning

Deep learning is a subset of machine learning. It uses neural networks and is suitable for dealing with a large amounts of unstructured data. And since it uses neural networks therefore the feature engineering is carried out automatically.

image source: blog.aimagnifi.com

A machine learning function starts with manual extraction (done by data scientist) of significant features from pictures. With the help of these features, a model is created and then objects in the image are sorted using these features. Deep learning workflow helps in extracting the appropriate features automatically from images. Deep learning carries out “end-to-end learning” — that means any network is provided raw data and also the task to be executed, for example, classification and then it gains knowledge of how to accomplish this automatically.

One major benefit of deep learning networks is- as the volume of our data becomes enlarge, they (deep learning networks) regularly keep on improving.

What is a Neural Network?

image source: data-flair.training

Since we discussed so many things about deep learning, so now let us talk a bit about neural networks too because deep learning uses neural networks. As said above, deep learning deals with algorithms encouraged by the structure and functioning of the human brain. Now we need to understand that the human brain works with the help of neurons. A human brain contains millions of neurons interconnected to each other. An artificial neuron network (also called a neural network) is a way of simulating the human brain. They receive signals as input from other neurons or other parts of the body and based on certain criteria they send signals to the next neurons.

The smallest unit of any artificial neural network is an artificial neuron. The artificial neuron has a central unit that receives the input. Like, if it is doing image processing the inputs could be the pixel values.

There are hundreds or thousands of artificial neurons called processing units in ANN, that are interconnected by nodes. These are comprised of input and output units. Various forms and structures of information are received by the input units and the neural network efforts to learn regarding the information presented for producing an output report.

So, this was all for this article.

--

--