Topic : Deep Neural Networks, and their Applications .
Ms. Harshika
What are Deep Neural Networks
(DNNs)?
A Deep Neural Network (DNN) is a part of Artificial intelligence
(AI) that works like a human brain. It is made up of many
layers of connected nodes called neurons. These neurons help
the computer learn and make decisions on its own.
(DNN) is a advanced type of Artificial Neural Network (ANN)
that contains many hidden layers between the input and
output layers. It is designed to learn from large amounts of
data and recognize complex patterns automatically- just like
the human brain.
Definition:
• A deep Neural Network (DNN) is a type of Artificial neural network
(ANN) that has multiple hidden layers between the input and output
layers.
• These layers allow system to learn from complex patterns and
relationships in data.
Structure of a DNN.
A typical DNN consists of:
1. INPUT LAYER:
- It is the first layer of the network.
- It receives the input data(for example, an image, sound or text).
- Each neuron (node) in this layer represents one feature or value from the input.
- It only passes the data to the next layer without changing it.
2. Hidden layers:
- These layers are present between the input and output layers.
- They are called hidden because they are not directly seen from outside.
- Each hidden layer processes the input data, finds patterns, and passes the information forward.
- The more hidden layers a network has , the “deeper” the neural network is.
- These layers perform complex calculations using weights, biases, and activation functions to learn the
data.
3. Output layer:
• It is the last layer of the network.
• It gives the final result or output, such as classifying an image as “cat”
or “dog”.
• The number of neurons in this layer depends on the type of output
required.
How DNNs work?
• (Step by step working of DNNs)
1. Input layer:
The process starts when data is given to the input layer.
Example: if the task is to recognize an image, the input layer receives all the pixel values of that
image.
2. Forward propagation:
The data moves forward through the hidden layers.
Each neuron in these layers does some mathematical calculations using weights and biases.
An activation function is applied to decide whether a neuron should activate or not.
This helps the neuron detect important features and patterns in the data.
3. Output layer:
-> After passing through all hidden layers, the final layer gives the output result.
-> Example: The output could be “cat” or “dog”, “spam” or “not spam”, etc.
4. Error calculation (loss):
The network checks how much its output is different from correct answer.
This difference is called error or loss.
5. Backpropagation :
To reduce the error, the DNN adjust the weights of neurons.
It sends the error backward through the network and updates the weights using a method
called backpropagation.
This helps the model learn and improve accuracy.
6. Training repeats:
-> Steps 2 to 5 repeat many times(called epochs) until the network gives accurate results.
Example:
• The input layer takes the images.
• The hidden layers learn features like ears, fur and eyes.
• The output layer finally predicts. – “This is a cat”.
Key Characteristics of DNNs:
• Learn from large datasets automatically.
• Capable of feature extraction without manual programming.
• Handle non- linear relationships in data.
• Require high computing power (GPUs, TPUs).
Applications of Deep Neural
Network ( In detail) :
1. Computer vision:
• Image classification- DNNs (especially CNNs) identify objects in images (e.g, cat vs dog)
Example: Google photos image recognition.
• Facial recognition- used in smartphones and security systems.
• Object detection- used in autonomous vehicles to detect pedestrians and traffic signals.
2. Image recognition:
• DNNs- especially convolutional neural networks(CNNs)- are widely used in
computer vision tasks.
• How it works:
• The network takes an image as input(as pixel values).
• Early layers detect simple features(edges , colors).
• Deeper layers detect complex structures(faces,
objects , scenes )
The final layer classifies the image into categories(eg. Dog, cat or tree)
Examples:
• Facial recognition systems(used in smartphones and security systems).
• Medical imaging (detecting tumors in x-rays or MRIs)
• Autonomous vehicles (detecting pedestrians, road signs, and obstacles )
3. Speech recognition:
• DNNs- especially recurrent neural networks(RNNs) and long short-term
memory(LSTM) networks- are effective in speech and audio processing.
How it works:
• Audio input is converted into a sequence of sound features (spectrograms)
• The DNN learns the temporal patterns in the speech signal.
• The output is a sequence of words or phonemes representing the spoken sentence.
• Example:
• Virtual assistants: siri, alexa, and google assistant use DNNs to convert speech to
text.
• Transcription services: automatically converting lectures or meeting into text.
• Language translation: real –time speech- to- speech translation (e.g, google
translate)
4. Self – driving cars:
• DNNs are used to make cars see and understand their surroundings using cameras and sensors.
• They can detect traffic lights, pedestrians, and other vehicles to drive safely.
• Example: Tesla and Waymo use DNNs for autonomous driving systems.
5. Medical Diagnosis:
• DNNs helps doctors analyze x-rays, MRI scans, and other medical images.
• They can detect diseases like cancer, tumors, eye problems early.
• Example: Google’s DeepMind Health uses DNNs to detect eye diseases from retina scars.
6. Recommendation systems
• DNNs learn your interest and suggest what you might like next.
• Example: YOUTUBE, NETFLIX, And AMAZON use DNNs to recommend videos, shows or products.
• Example: spotify suggests songs you might enjoy based on your listening history.
7. Natural Language Processing
(NLP):
• DNNs helps computers understand and generate human language.
• Example: Google Translate uses it to translate text between different
languages.
8. Agriculture:
• DNNs are used to monitor crop health, detect plant diseases, and predict yield.
• Example: Drones with AI cameras analyze fields to help farmers take better care of crops .
Advantages of DNNs:
• High accuracy in complex tasks
• Automatically learns features (no manual programming).
• Can handle massive and unstructured data.
• DNNs used in almost every field- from Healthcare to entertainment.
• DNNs are very good at finding hidden patterns and relationships in
large datasets.
• For example: Detecting spam emails or identifying similar- looking
faces in a photo collection.
Limitations:
• Requires large amount of data.
• Needs high computing power.
• Can be difficult to interpret.
• Training time is long.
• Required skilled experts.
• Expensive to build and maintain.
• It’s not easy for beginners to handle.
Popular Deep Learning
frameworks:
• Software tools or libraires that help us build, train, and test Deep
Neural Networks. (DNNs).
1. TensorFlow : developed by google.
2. PyTorch: Developed by Facebook (Meta).
3. Keras : high level library that works on top of tensorflow.
4. MxNet: developed by Apache.
5. Microsoft Cognitive Toolkit : developed by Microsoft.
Future scope:
• More efficient Architectures (like transformers)
• Explainable AI (making DNNs more transparent).
• DNNs may improve fraud detection,risk management, and financial
forecasting.
• Future AI will understands human language better, enabling smarter
chatboxs, translators, and voice assistants.