Era of AI (Deep Learning) and harnessing its true potential
Artificial Intelligence (AI) AI Augments our brain with infallible memories and infallible calculators Humans and Computers have become a tightly coupled cognitive unit Instrumental in disrupting various industries and closely connected with research Medicine Research Environmental Catastrophe Eliminating Poverty
About Me. Analytics Consultant Lead - Opera Solutions Senior Consultant - PwC Principal Data Scientist - Infoedge Senior Data Scientist - Experian (APAC) Gaurav Kumar
Current State of AI In 2016, companies invested $26BN - $39BN in AI 3X External Investment growth since 2013 AI adoption is greatest among stong digital adoptors Embedded in almost all devices we interact with - Cars, Mobiles, wearing devices, equipments etc. Amazon Kiva - $775M Click to ship cycle time Netfilx Recommendations Impact $1B annually Google Company Strategy Mobile first - AI First
Machine Learning - Background Machine Learning is a type of Artificial Intelligence that provides computers with the ability to learn without being explicitly programmed. Labelled data Machine Learning Algorithm Data Training Prediction Learned Model Prediction Various techniques that can learn from and make predictions on data
Machine Learning - Learning Approaches Supervised Learning: Learning with a labeled training set Example: email spam detector with training set of already labeled emails Unsupervised Learning: Discovering patterns in unlabeled data Example: cluster similar documents based on the text content Reinforcement Learning: learning based on feedback or reward Example: learn to play chess by winning or losing
Types of Machine Learning Problem Types Problem Types Comparision of ML Classifiers Classification (supervised predictive) Regression (supervised predictive) Clustering (unsupervised descriptive) Anomaly Detection (unsupervised descriptive)
What is Deep Learning? A class of machine learning algorithms that use a cascade of multiple non-linear processing layers where higher level features are derived from lower level features to learn different representations of the data in each layer A deep neural network consists of a hierarchy of layers, whereby each layer transforms the input data into more abstract representations (e.g. edge -> nose -> face). The output layer combines those features to make predictions Edges Nose, Ear, eyes Faces
A new concept? 1958 Perceptron Convolution Neural Networks for Handwritten Recognition 1998 Google Brain Project on 16k Cores 2012 awkward silence (AI Winter) 1969 Perceptron criticized 1995 SVM reigns 2006 Restricted Boltzmann Machine 2012 AlexNet wins ImageNet
What Changed? Big Data (Digitalization) Computation (Moore s Law, GPUs) Algorithmic Progress
Why so popular? Traditional ML Input Data Feature Engineering Traditional learning Algorithm Deep Learning ML Input Data Deep learning Algorithm Applications of Deep Learning Speech/Audio Processing Computer Vision Natural Language Processing
Deep learning vs Humans Human Performance Human Performance ImageNet: The computer vision World Cup Deep Learning in Speech Recognition
How does it work? Feed-Forward and Backpropagation Optimization Algorithms - Comparison
General Architecture - Deep Learning Convolution layer is a feature detector that automatically learns to filter out not needed information from an input by using convolution kernel. Convolutional neural network shares weights between local regions Recurrent neural network shares weights between time-steps. Maps input sequence to output sequence. The output vector s contents are influenced by the entire history of inputs.
Requirements Large data set with good quality (input-output mappings) Measurable and describable goals (define the cost) Enough computing power (AWS GPU Instance) Excels in tasks where the basic unit (pixel, word) has very little meaning in itself, but the combination of such units has a useful meaning.
Popular Applications Deep Q-Learning (Reinforcement Learning) Deep Learning for Games (Environment by Google DeepMind and OpenAI Gym Alpha go - trained deep learning model Image beats Localization the world champion in Go Image Completion Image Image Segmentation Image Captioning Image Compression Image Transformation Image Sharpening Syntax and Semantics Image Colorization Image Augmentation Adding Emotions Summarization and Text Generation Image generation via Generative Fine Adversial Arts Networks Album Covers (Inceptionism)
Current State of Recommendation Engines Collaborative Filtering Peer Information Preferences Information from various users Generation of User x Item (Utility) Matrix Methods of performing Collaborative Filtering: User-User similarity Item-Item similarity SVD decomposition and Similarity Some drawbacks but reco. are relatively more relevant Content Based Recommender Profile Information User declared content used to recommend NLP applications on Text (or Image data) Methods of performing Content Matching: Creation of User Features Creation of Item Features Cosine similarity between user and item Time consuming and difficult to productionize
NLP - Application You shall know a word by the company it keeps (Firth, J. R. 1957:11) Embeddings are used to turn textual data (words, sentences, paragraphs) into high- dimensional vector representations and group them together with semantically similar data in a vector-space. Thereby, computer can detect similarities mathematically. Woman Man Aunt - Uncle King - Male + Female Queen Human - Animal Ethics
Create word embeddings or using off the shelf Easiest approach Use learned word embedding (GLOVE Stanford) Google News 3 Billion word corpus Twitter 3 Billion tweets Wikipedia 6Billion word corpus
Recommendation using Word/Doc Embeddings For every user/item based on the content, 1.Calculate Word2vec for each word in document 2.Use weighted average (weights from TF/IDF) 3.Doc2vec to find document representations 4.Cosine product of User x Item to rank-ordering
Adoption in RecSys NLP embeddings used in content based recommendation engine to match relevant text metadata between users and products. Model based recommendation mimicking collaborative filtering with multiple variants proved and adopted in industry Change of paradigm from static overnight computed recommendations to dynamic realtime recommendations using current session information
Thank You
Real Time Recommendations (Session Info)