Welcome to Kaytek's Easy AI (Artificial Intelligence) Landing Page
Artificial Intelligence is nothing but Augmented Intelligence
Published Articles on AI / ML (Machine Learning) / DL (Deep Learning) (cont)
New ! Baby steps on an AI journey - 31st October 2019.
New ! MIT Sloan Management Review Twitter Chat Interactions - Implementing Artificial Intelligence - 3rd September 2019.
New ! One of the industries where AI is poised to have a huge impact in the future is the retail sector. Perhaps not only a negative impact due to perceived job losses caused by AI based automation, but also a positive impact caused by higher productivity. Can Retailers learn fast ? - The Mystery Of The Misplaced Shoe...Can AI Help ? - 31st May 2019.
Article Can I get a rAIse ? - CHAAI Se Charcha - An appraisal discussion between a fictitious human worker and an organization's CHAAI (CHief Appraisal AI Officer) - 30th January 2019.
Understanding the Resnet (Residual Network) Block in Convolutional Neural Networks (CNN) for AI - Computer Vision
For solving Artificial Intelligence (AI) Computer Vision problems, Convolutional Neural Networks (CNN's) have been always popularly used. Within CNN's there was a breakthrough a few years back when the Resnet (Residual Network) Block was introduced as an architectural innovation to reduce the problems of adding extra layers in the network architecture.
The original technical paper released on Resnet unfortunately did not contain an easy explanation of the terminology used in the Resnet Block diagram (Figure 2) in the same. To help ease understanding of the same, an article Resnet Block Explanation with a Terminology Deep Dive along with an accompanying presentation has been released on Medium. - 23rd October 2018.
Neural Networks are the secret sauce of Artificial Intelligence (AI) - Article (Yet Another) Neural Network Terminology Upto WX + B Stage - 27th September 2018.
Artificial Intelligence (AI) Maths captures real world knowledge - Article Entity Embeddings package real world knowledge for Artificial Intelligence (AI) algorithms - 1st August 2018.
Whatsapp Meets Google Artificial Intelligence (AI) - Article You are obsessed with Whatsapp analysing more than 700000 lines of Whatsapp chats using Google AI's Natural Language Cloud Services - 27th March 2018.
New ! 6th June 2019 Presentation on Entity Embeddings at Deep Learning Meetup in New York, USA.
More Event Photographs (courtesy Kris Skrinak)
Kaytek Founder Director Mr Mahesh Khatri spoke at the Deep Learning NYC Meetup Group in New York, USA on Thursday 6th June 2019 on the topic Entity Embeddings & Pytorch in the Enterprise". The event was organized by Mr Kris Skrinak of Amazon Web Services (AWS) & Ms Pallavi Gadgil,the leaders of the Deep Learning NYC Meetup Group. Thanks to them and also Amazon for hosting the above meet at their New York office. Also, a big thanks to to all the participants who came and actively interacted on the topic.
Presentation Overview - Entity Embeddings is an upcoming technique for applying deep learning to tabular data. It involves representing the categorical data of an information systems entity with multiple dimensions.
It is being used in several production systems at companies such as OpenAI, Google, Instacart, Twitter, etc. In his presentation, Mr Mahesh Khatri presented the concept along with examining it's usage in the following 3 papers : Kaggle Competition winner papers - Artificial Neural Networks Applied to Taxi Destination Prediction (Yoshua Bengio’s team - 31st July 2015) & Entity Embeddings of Categorical Variables (22nd April 2016) & Google Research paper - Deep Neural Networks for YouTube Recommendations (16th September 2016).
Youtube Video Contents & Timeline
0:00 - Why Talk of Entity Embeddings ?
2:45 - What Are Entity Embeddings ?
8:32 - Importance of Entity Embeddings
9:34 - 2 Perspectives - Word Embeddings & Real World Tabular Data
10:15 - Word Embeddings
27:00 - Real World Tabular Data
34:55 - Machine Learning Library Support
39:39 - Artificial Neural Networks Applied to Taxi Destination Prediction
42:33 - Entity Embeddings of Categorical Variables
45:24 - Deep Neural Networks for YouTube Recommendations
52:30 - Industry Usage - Twitter, OpenAI, Healthcare, etc
55:13 - Aricles, Summary, Call To Action
FastAI References in the above talk :
14:07 — Size of Embedding
26:23 & 35:35 — Fastai Library Support function — add_datepart
37:38 — Jeremy Howard on Embedding size
44:37 — Rachel Thomas on ‘Rossman Stores Competition’ paper
52:30 — Jeremy Howard on commercial & scientific opportunities
New ! Medium Article on the above talk - An Entity Embeddings sharing with New York's AI Community - 29th June 2019.
Application of Entity Embeddings - Article Collaborative Filtering — Understanding embeddings in User Movie Ratings - 20th December 2018.
Entity Embeddings Deep Dive - Entity Embeddings are used by some of the largest and smartest organizations on the planet like Amazon, Facebook, Google, Twitter and many more in their gigantic internal production scale machine learning systems.
TwimlAI (This Week in Machine Learning & AI) just published Mahesh Khatri's presentation on Entity Embedding Deep Dive at their North America Meetup on 13th November 2018.
An explanation of Entity Embeddings & usage in the following research papers was covered :
- Cheng Guo's paper - Entity Embeddings of Categorical Variables which was a 3rd prize winner at a Kaggle competion
- Yoshua Bengio's paper - Artificial Neural Networks Applied to Taxi Destination Prediction which won the 1st prize at the ECML/PKDD discovery challenge &
- Google AI Research paper - Deep Neural Networks for YouTube Recommendations
All of those who are interested in learning about contemporary usage of Entity Embeddings may find the presentation useful.
A Basic Knowledge of Neural Networks is needed for understanding this presentation.
Thanks to Sam Charrington & TwimlAI for their support in hosting this event. - 10th December 2018.
Viewpoint - Approaches to AI / ML / DL Education - Specific Technology versus Generic Learnings
One of the valid concerns expressed by experts is the choice of technologies to focus for people entering this field. In one specific podcast, the discussion was on Pytorch versus Tensorflow.
There is already an abundance of AI / ML / DL educational offerings and technologies out there. The pace of innovation will not slow down.
The learning approach should always be to extract generic learnings from specific AI / ML / DL technologies so that future learnings and re-learnings are easier and faster when the next AI / ML / DL tool or technology arrives.
It may be much much tougher, take a much longer time, but the results in terms of long term conceptual learnings will be worth the effort. DL Giants such as Geoffrey Hinton spent years toiling away before getting meaningful results. The impatience and haste shown by many beginners to the field reminds one of the kindergarten story of The 3 Little Pigs.
Viewpoint - Thoughts on FastAI Courses - The 'Top Down' Versus 'Bottoms Up' Approach
Jeremy Howard, the founder of FastAI with over 30 years of ML & coding experience is obviously well qualified. He has positioned the course as different from all the other courses in the market via a 'Top Down' approach which is code heavy and digs into the Maths whenever required on a need to basis. However,'Being different' need not mean 'Being easier to understand'.
For beginners; the 'Top Down' approach parachutes learners immediately to the peak of Mount Everest. Without having struggled to the top. From the peak, we get a great view standing on top of the mountain of FastAI and Pytorch based on his 3 decades plus experience. The results of Fast AI as demonstrated by consistent world class contemporary benchmarks are indeed remarkable. They give us an immediate starting point to use this world class library.
We are dazzled with a 'Shock & Awe' feeling. However, we have not really climbed our way to the top. We need to do a detailed deep dive into the code. Which means that a bottoms up approach will help people who are struggling currently to get their basics & foundations right. The hierarchy of code understanding on a bottoms up basis should be in this sequence : Python - Numpy - Pytorch - FastAI.
It is interesting that Jeremy Howard has planned the next version of fast.ai part 2 to be a 'Bottoms Up' approach.
Last updated on 1st November 2019.
Created on 17th October 2018.