Intro AI - Tagged PDF

Artificial Intelligence An Introduction: What is it and why should we care? M. Shukla  Intelligence? M. Shukla  3 Types of AI 1. Artificial narrow intelligence, aka Weak AI, typically does one job well 2....

Artificial Intelligence An Introduction: What is it and why should we care? M. Shukla  Intelligence? M. Shukla  3 Types of AI 1. Artificial narrow intelligence, aka Weak AI, typically does one job well 2. Artificial general intelligence, aka Strong AI, this is as smart as a human 3. Artificial super intelligence, surpasses human intelligence M. Shukla  Context  A subset of Computer Science  Two broad areas  Symbolic Reasoning The domain of expert systems Human reasoning hard-coded in (nested if-then)  Machine Learning Statistical Learning Speech Recognition Natural Language Processing Deep Learning M. Shukla  Influential ML Minds  Some names you might see, but there are many more…  Geoffrey Hinton  Michael I Jordan  Andrew Ng  Yann LeCun  Yoshua Bengio  Demis Hassabis See: https://medium.com/@lijiang2087/10-most-important-people-in-artificial-intelligence-in-2017-4120ef6cbec7 M. Shukla  More Context From the mid 1980s Eg Sigmund and gigo M. Shukla  Why Things Stalled  Boolean logic functions are a fundamental part of how computers operate A B AND OR 0 0 0 0 0 1 0 1 1 0 0 1 1 1 1 1 Here we have true represented with a 1 and we have false represented with a 0 M. Shukla  More Boolean(s?) http://www.ambrsoft.com/Equations/LogicalGates/LogicalGates.htm M. Shukla  AND 0,1 1,1 0,0 1,0 A B AND 0 0 0 0 1 0 1 0 0 1 1 1 M. Shukla  OR 0,1 1,1 0,0 1,0 A B OR 0 0 0 0 1 1 1 0 1 1 1 1 M. Shukla  The Problem with… XOR! A B AND OR XOR 0 0 0 0 0 0 1 0 1 1 1 0 0 1 1 1 1 1 1 0  Exclusive OR caused us a significant problem for some time… Essentially OR but not AND  Alternatively it as when A or B is true but not A and B M. Shukla  XOR! 0,1 1,1 How and where do you draw the division line? 0,0 1,0 This is an example of a non-linearly A B XOR separable problem. Minsky and Papert 0 0 0 pointed this out in 1969. 0 1 1 1 0 1 1 1 0 M. Shukla  Biological Neurons There are c.100 billion neurons in a human brain Biomimicry M. Shukla  An AI Neuron  Basic Perceptron model Input 1 Weight 1 Input Weight 2 2 Node Output Input Weight n Weighted sum Σ n and Activation Function M. Shukla  Activation See also Activation Functions section from: https://towardsdatascience.com/coding-neural-network-forward-propagation-and-backpropagtion-ccf8cf369f76 https://medium.com/@shrutijadon10104776/survey-on-activation-functions-for-deep-learning-9689331ba092 p.125 from 'Deep Learning for Computer Vision with Python' by Adrian Rosebrock M. Shukla  Solving XOR 0,1 1,1 Multiple linear classifiers can divide up data which didn’t seem separable by a single straight line, a non- linear plane if you like 0,0 1,0 A B XOR 0 0 0 0 1 1 1 0 1 1 1 0 M. Shukla  Deep Learning  Multi layered This shows Feed forward but Back propagation needs also to be considered M. Shukla  Backprop Make Your Own Neural Network by Tariq Rashid M. Shukla  Learning Supervised learning The easiest way. Can be used if a (large enough) set of test data with known results exists. Then the learning goes like this: Process one dataset. Compare the output against the known result. Adjust the network and repeat. Unsupervised learning Useful if no test data is readily available, and if it is possible to derive some kind of cost function from the desired behaviour. The cost function tells the neural network how much it is off the target. The network then can adjust its parameters on the fly while working on the real data. Reinforced, learning The ‘carrot and stick’ method. Can be used if the neural network generates continuous action. Follow the carrot in front of your nose! If you go the wrong way - ouch. Over time, the network learns to prefer the right kind of action and to avoid the wrong one. https://appliedgo.net/perceptron/ M. Shukla  Suit You - Fitting  Overfitting  Underfitting This is essentially about how much you train your Artificial Neural Network. Too much and you risk it only recognising the training data set. Too little and it simply wont be very good at what it is supposed to do. M. Shukla The metaphor here is guitar string cannot be too tight or too loose  Supervised  A labelled training data-set is initially used  Once acceptable results are gained then unlabelled data can be introduced  Typically used for classification and regression https://towardsdatascience.com/supervised-vs-unsupervised-learning-14f68e32ea8d M. Shukla  Unsupervised  Unlabelled data is introduced with the ANN finding its own inherent structure/labelling  Typically used for clustering or association https://towardsdatascience.com/supervised-vs-unsupervised-learning-14f68e32ea8d M. Shukla  Hybrid  Typically has only a small amount of labelled data for training  A combination of supervised and unsupervised approaches can be applied  Many real world applications fall into this category often due to lack of labelled data M. Shukla  Reinforcement  Has become popular more recently  No labelled data for training as such…  Rather, a method to quantify performance is introduced using a reward signal  For an example using Mario see: https://recast.ai/blog/the-future-with-reinforcement-learning-part -1/ M. Shukla  And then..  Convolutional Neural Network  Generative Adversarial Network  Recurrent Neural Network  Long/short-term memory networks  Recursive neural network  Natural Language Processing M. Shukla  Intelligence? “How much do we think about chimps?” Elon Musk M. Shukla https://youtu.be/ycPr5-27vSI?t=726 (Trigger warning! Joe Rogan uses adult language)  See also…  https://towardsdatascience.com/what-the-hell-is-perceptron-626217814f53  https://www.youtube.com/watch?v=ntKn5TPHHAk  https://www.youtube.com/watch?v=b99UVkWzYTQ&list=PLjJh1vlSEYgvGod9wWiydumYl8hOXixN u  http://bigdata-madesimple.com/machine-learning-explained-understanding-supervised-unsupervise d-and-reinforcement-learning/  https://www.youtube.com/user/keeroyz/featured M. Shukla