Excursion to Artificial Neural Networks Introduction to Artificial Neural Network 

Presented By Mayank Misra A data geek passionate about AI | ML @Mayank_skb 

Line Up • Neural Network • Inspiration to Artificial Neural Network • Why Artificial Neural Network • Machine Learning vs Artificial Neural Network • Types of Neurons 

Population of Neurons, interconnected by synapses to carry out specific function Neural Network 

No content

A Story 

No content

No content

Artificial Neural Network A computer model based on structure and functions of biological neural network. In itself it is not an algorithm, but rather a framework for many different machine learning algorithm to work together and process complex data. ANNs are considered as non- linear data modeling statistical tools where the complex relationship between input and output are observed or patterns are found. Likewise the biological neural networks, the functional unit of an artificial neural network is a Neuron or rather an Artificial Neuron. 

Artificial Neuron Artificial Neuron (fundamental unit of an artificial neural network) 

Inspiration to Artificial Neural Network 

• To model non-linear and complex information. Neural Networks have the ability to learn and model such non-linear and complex relationships. • Unlike other prediction techniques, they do not impose any restrictions on the input variables (ex: like how they should be distributed) • They do not requires explicit feature selection. • Also they are widely used in pattern recognition because of their ability to generalize and to respond to unexpected inputs. • They are used in many applications such as language translation, animal recognition, text summarization, handwriting recognition etc. which is used a lot. Why Artificial Neural Network 

Machine Learning vs Neural Network No need to do Feature Selection learn high level features in an incremental fashion End to End Solution e.g. YOLO net feed-in input and have localized output, if using SVM then need to have an bounding box algorithm then object identification algorithm Training and Testing speed longer time for training due to large number of parameters but testing scenario is reverse, KNN test time increases on increasing the size of data Model Interpretation Neural Network models are usually black box, don’t know what they are modelling and how Performance Neural Network performance increases with increase in amount of data best to deal with big data, ML becomes stable 

Types of Neurons Perceptron Sigmoid • A single layer neural network • Takes multiple binary inputs and produces a single binary output { Output 1 if σ=0 > 0 -1 otherwise 

Shortcomings of Perceptron • Perceptron can only do binary classification • Perceptron can only classify linearly separable set of vectors The XOR problem addressed by Minky and Papert 

Types of Neurons Perceptron Sigmoid • Sigmoid addresses the short comings of Perceptron • Using Perceptron we cannot have layered architecture • Sigmoid is continuous and differential 

Learning Procedure Sigmoid Neuron zi = sigmoid(∑wi xi + b) where sigmoid is Sigmoid(z) = 1/1+− Hence zi = 1/1+−(∑wixi + b) And using backpropagation it updates its weights 

Comparison between Step function and Sigmoid 

T H A N K S F O R L I S T E N I N G mayank_skb