Artificial neural networks are programs inspired by the human brain. A variety of industries are now utilizing them, but a few leading the adoption of ANN include the manufacturing, security, and healthcare industries.
Artificial neural networks, also referred to as ANN, are programs inspired by the human brain and enable deep learning applications to learn from large unstructured datasets (in the form of text, images, and videos). A variety of industries are now utilizing artificial neural networks, but a few leading the adoption of ANN include the manufacturing, security, and healthcare industries.
Source: ScienceDirect
The fundamental component of a neural network is an 'artificial neuron', just like the biological neurons in a human brain. Simply put, an artificial neuron takes different inputs, processes data, and passes the output to other neurons in the network.
Weights
A weight is assigned to each feature at the input layer. These weights are then optimized to minimize the loss function. Scalar multiplications of weights and their weighted sum is provided as an input to the activation function in a neuron.
Bias
Optimizing the weights only affects the steepness of the sigmoid, so in order to shift or move the curve, ‘bias’ values are required. Bias values shift the activation function, which is critical for successfully training the deep learning model.
Activation Function
Whether a neuron fires or not depends on the output of the activation function and the weighted sum of inputs along with bias added to it. Activation function introduces non-linearity into the output of a neuron. There are various activation functions like Sigmoid, ReLU, Tanh, and Leaky ReLU.
Activation Function | Function | Plot |
---|---|---|
Sigmoid
|
|
|
Tanh
|
|
|
ReLU
|
|
|
Leaky ReLU
|
|
|
Real-world data is fed to the neural network for learning at the input layer. While no computation is performed here the data continues to be provided to the hidden layers.
Each computation and optimization of weights and biases from forward pass and backpropagation algorithms are performed on hidden layers. Different variants of the gradient descent algorithms are used to optimize the weights w.r.t the errors. There can be multiple hidden layers with different numbers of neurons at each layer which further transfer the results to the output layer.
This layer provides the predications made by the model. Loss is calculated by comparing the model’s predictions with the expected results. This loss is minimized using gradient descent algorithms.
Computer vision provides computers with the ability to visualize, identify, and process objects in images and videos. Computer vision is used for object identification, medical image analysis, event detection, video tracking, motion estimation, etc. The most commonly used algorithms for computer vision are variants of Convolutional Neural Network (CNN).
Convolutional layer is the core of a CNN. This layer has different learning parameters like filters/kernels, Padding, Stride, and Channels. These parameters enable teams to process, analyze, extract important features from image data and train the deep learning model.
Our team at QASource has extensive experience in building and testing deep learning systems and custom neural networks using CNN architectures like ResNet, AlexNet, and OneNet. We use various deep learning development libraries like TensorFlow, PyTorch, and Theano Keras. Our engineering teams have experience in developing and testing deep neural networks for different applications like computer vision, object detection, and object identification.
We would love to hear your feedback, questions, comments and suggestions. This will help us to make us better and more useful next time.
Share your thoughts and ideas at knowledgecenter@qasource.com