Architectures of Deep learning with Pros/Cons and its applications

Share on:

In today’s rapid growing world, Artificial Intelligence unbelievably enhanced the human life in many aspects. But if we talk about the advanced AI then it undoubtly indicate toward the deep learning. Deep learning is a base for ANN or we can say that DL is a subset of ANN. Deep learning works on algorithms which are inspired by the structural and functionally similar to human brain neural system known as artificial neural system (ANN). DL is a field in which more than one hidden layer are used for the manipulation of complex data. It is the hierarchical arrangement of huge data in order to manipulate for the better performance than ANN.

Deep word in the deep learning denotes the depth of the machine learning algorithms where ANN is known to be a shallow part. The rise of deep learning from the few years is just because of good computational capabilities and for the effective arrangement and utilization of huge amount of unsupervised data. Deep learning has number of architectures on which it works:

Deep Neural Network (DNN)

Deep neural network is a type of ANN in which various types of hidden layers are present between the input and output layers which are used for the modelling of complex data. It is a feedback-forward architecture in which the data flow in a linear way without looping back of it.

Deep Learning
DNN

The main problems in the DNN are Overfitting and computation time. Overfitting can be avoided by using weight decay or sparsity methods while computation time can be improved by using graphical processing units (GPU) and by considering learning rate, batching, etc.

Recurrent Neural Network

It is also a kind of neural network in which information cycles are created by the computational units of ANN which it represents dynamic temporal behavior. Looping is present in the network of RNN which is used for the persisting of data.

Deep Learning
RNN

This architecture of ANN is used to connect all the pervious information in the current task which is going on if the gap between them is small. But if the gap is large between the previous information and the point where it is needed then a particular type of RNN is used which is known as Long short-term memory (LSTM). These connections between previous information and current tasks helps to solve the real time situations or problems effectively.

Convolutional Neural Network (CNN)

In deep learning, CNN is a multi-layered feed forward interconnected network which is used in the extraction of information of visual imagery or we can say that it is used in the area of image detection or recognition.

Deep Learning
CNN

CNN extract all the information or parameters of image which is further used for many applications like facial recognition, optical character recognition, identification procedures, surveillance, etc. However, traditional ways for information extraction uses particular methodologies but CNN automatically detect the filters of images and extract information from it on its own.

Boltzmann Machine

This architecture is a neural network all the nodes or networks are interconnected just like the neurons. This architecture mainly used for taking inconsistent decisions. In this architecture of DL, the nodes whether they are visible of hidden all are connected to each other.

Restricted Boltzmann Machine

Deep Learning
Restricted Boltzmann

This architecture is just a kind of Boltzmann machine in which the hidden and visible nodes are connected to each other but the two or more hidden nodes are not connected to each other and also visible nodes are not permitted to connect to each other. It is used in the enhancement of the performance of speech recognition, etc.

Deep Belief Network

It is a graphical model in which hidden layers are present to convey the abstract representation of features of input data. The hidden layers are not connected with any neuron or computational layer but the hidden layers are connected to each other. It is mainly used for the training purpose of the network in a layer-wise manner by following the hierarchical graphical modelling.

Deep Autoencoder

This type of neural network is used to encode or decode the features of the input data in the Deep Belief Networks. Basically, it is the encryption and decryption of features of data (showing less features) without changing the sense of information of data when it flows from the layers of deep belief network.

Deep Learning
Deep Autoencoder

Typically, the units of the hidden layers between the input and output layers are usually less than the units of input and output layers and of which are same in the input and output layers. This encryption and decryption of data features helps to identify with minimal features of data to represent the data and called as dimensionality reduction.

Applications of deep learning

Object Tracking

Deep Learning enhances the great advancement in the field of object tracking. These object tracking devices are mainly used in tracing the specific person or a vehicle at the traffic signals which greatly help the crime branch departments to trace the criminals. Object tracking also detect and identify the class of vehicle with the help of Convolutional Neural Network (CNN).

This application which is dependent on deep learning is also used for identifying the distance between the cars and their speed in self- driving car systems and also experience the vertical and horizontal distances between the objects. With the help of object tracking we can also identify the size and position of the object in the CCTV camera footage. This application of object tracing is also used in the medical departments to detect the tumors in human body in the cases of cancer suffering peoples and useful for human well-being.

Smart Surveillance using Edge and Cloud Computing

Smart Surveillance through CCTV cameras by using deep learning is another good application. As the traditional surveillance record and save the all the footages on hard drives and hence, decreases its life. The new surveillance uses the cloud space to store the footages for easy backup and reduce the risk of data loss. But this method requires more data bandwidth and good upload bandwidth too. But it is costlier fulfil the demands of higher data usage to upload the footages to the cloud.

Also, the cloud services for additional storage requires the great cost according to per month costing. So, to avoid all these expenditures there is a great option to consider the smart surveillance using the deep learning algorithms to identify the specific important footages to process and send it to the server for uploading and saving it to the cloud.

This system of smart surveillance is designed in such a way that it can scale, tolerate failures and work with multiple devices and cameras. The architecture of this surveillance having 4 stages that is Device layer, Pre-processing Layer, Processing Layer and Cloud layer.

In the field of Fault Diagnosis

Deep learning tends to be very efficient in the field of fault diagnosis. As we know that deep learning works by extracting features from the given data. So, we have to save a time for extracting features by setting essential and appropriate parameters of deep learning network. A map should be prepared by considering original map to the failure data classification point-to-point.

This practice helps in the extraction of artificial features. This technique would be helpful when there is a lack of fault samples in the aeroplanes and other equipment by using the advantage of migration ability of deep learning. By using rotating machinery ground test bed, we can easily use it in deep learning for state recognition, fault diagnosis, etc.

Realizing Super Resolution applications on integrated GPUs

Recent Advancement in the deep Convolutional neural network (CNN) enables the deep learning to brought super resolution image on low powered integrated GPUs. With the help of CNN we can easily produce the best super resolution images or applications by adding the extra details in the extracted features of the images which can easily maintained and realize on the Intel’s Iris graphics which an integrated graphics.

Instead of delivering large inputs, we can divide it in smaller tiles in order to use small memory bandwidth which results in the improvement of performance as well. By doing this, on integrated graphics we can easily optimize the SR single images to SR videos with upto 44 FPS quality without any complications and frame drops. By watching towards its compatibility, CNN also work efficiently with high end dedicated graphics as well.

Advantages and Disadvantages of Deep learning

Advantages

1. Deep learning can focus on that it tends to target without affecting its statistical modelling.

2. We can teach it to learn only a particular task or a particular set of tasks rather than allowing it to learn thing by its own.

3. It can understand or learn from unsupervised things and act according to it without any supervision in order to instruct it to work.

4. It is capable of making new images based on its previous memories.

5. Not dependent on computation power due to which its performance remains good and get awareness in any task more quickly and effectively.

6. It is extensible which means we can add more models of learning by just adding the more layers to its neural network.

Disadvantages

1. The comparison of its achievement with respect to a hand-crafted method is very different and harder.

2.  The data based on approximate statistics as not 100% accurate.

3. The efficiency is very low and have some difficult problems too.

4. Difficult to evaluate its performance regards to the real-world problems and applications, as more detailed testing techniques and analysis required for its validation and implementation.

5. Its training requires huge amount of data on which it get trained as it thinks about thousands of images and videos.

6. It is expensive as it requires large amount of memory and other computational hardware to store and compute thousands of data sets on which it is trained (implementation of algorithms).

Effects of AI on the life of Human Being

Positive Effects of AI

1. It helps the human to do their work with greater efficiency and precisely. It helps to augment their work and substitution of AI in the repeatedly or dangerous tasks hence reduce the loss of life. For example, Use of NPU, Deep learning and ANN based robotic machines in the prediction of Land mine bombs by the military.

2. AI can save the countless hours of humans and ensures the increment in their productivity instead of managing various un-necessary issues related to their daily routine jobs. For example; Use of AI to make autonomous transportation and resolving traffic handling issues can greatly influence the life of humans by saving their precious time and also improving the job productivity too.

3. Better AI facility in the field of Healthcare greatly influence the life of human being by reducing the risks of wrong or late diagnosis of diseases hence reduce threatening to human life. For example, the use of AI diagnostic machines in the diagnosis of earlier prediction of cancer stages and helps to cure it before it becomes too dangerous.

4. The use of AI in criminal investigation for identifying main culprit instead of innocent one. For example, using fingerprints found on crime spot or by face recognition techniques we can easily match them with doubtable peoples to identify the original criminal.

5. Using of AI in the field of sports for training purposes will improve the game play of the players in a very significant way. For example, for training of a batsman the AI ball thrower machine can easily detect the angle and speed of the ball according to the position and posture of the player and deliver the difficult delivery to him hence improve his/her game play significantly.

Negative Effect of AI

With the positive impacts of AI, there are some negative impacts of AI too. With the enhancement of AI in the life of human being it will significantly decrease the employment opportunities to the individuals.  For example; According to PwC, there will be decrement of 7 million existing jobs in the UK from 2017-2037. These 7 million jobs will be replaced by AI but there will be also possibilities of new jobs offering to peoples but not confirmed that up to how much extent. This uncertainty will be a big challenge by upcoming of AI opportunities in the life of human beings.

Share on:

Leave a Comment