Artificial Intelligence (AI) for Executives and Top-Level Managers - Udemy Excerpts

https://www.bluelife.ai/consulting

Sections:

Computer vision Natural Language Processing Reinforcement Learning Deep Learning Robotic process automation

Deep Learning -

Machine Learning (ML) - Discipline within AI that teaches computers how to make predictions based on data.

what is DL?

Family of ML algorithms based on Artificial Neural Networks

DL is a sub part of ML and ML is a sub part of AI

Without having to put any framework on our data, we let the Deep Neural Network to learn on its own. (like human babies learn about things)

If you feed huge images of dogs and cats to computer system - Just by observing the data and comparing the labels, the Deep Neural Network will learn these features on its own.

ML algo gets accuracy of 95% but Deep leaning accuracy is 99% so its very powerful, however you will need huge amount of data to train.

Why do we need Data Scientists or ML engineers to build Deep Learning Algo?

Input output type neural network

Input layer -------------> hidden layer-----------> output layer

Defining the number of layers and adding number of neurons in each layer.

The different ways that neural networks can be constructed are called architectures. each use case has different layers and neurons which is a creative work done by human.

Finding right neural network architecture is actually a very creative process thats why Machine Learning Engineers and data scientists who build deep learning algo are in high demand.

Once setup Neural network will do work on its own and learn.

How DL Neural network is build?

How NN (Neural network) works under the hood?

NN are very technical and complex, but principles and initution are quite simple. If we hear NNs are trained meaning their weights are being trained / modified.

Suppose we have 3 layers NN

NN have memory, which is effectively weights within these neurons. These weights can be adjusted and that effect how the calculations between neurons perform so basically the input layer is used as a weighted sum in each one of these neurons in the middle there and thats how the middle layer neurons are calculated and the weights is what we can adjust and then the final ouput layer is weighted sum of the neurons in the middle and thats how that final neuron is calculated and the weights there can be adjusted.

when performing the training part of a neural network, you have a lot of supervised data.

Input of Dogs and Cats are fed and those pictures having 0 and 1 are put in vectors and these go thru several layers before arriving at final deciion layer which tells final result is actually a cat or a dog. we pair the label to result. If the prediction is incorrect the weights need to be adjusted and thats called Back Propagation.

training is a iterative process where data is fed and back propagation done before we can say its tailored to predcit whether an image is a dog or cat.

Same approach can be applied to find whether customer is likely to churn within the next 6 months or not.

input can be customers demographic, affinity to purchase, geolocation, preferences, purchase history.

Once trained in the end, we will have a neural network designed to predict customer churn.

How to use DL with others areas of AI such as Reinforcement learning, Computer vision and NLP.

AL encapsulates ML and ML encaps Reinforcement learning, ComputerVision and NLP.

CV, NLP and RL don't always require DL in order to function, however if we use DL along with these CV,NLP and RL it means we are essentially adding the power of Deep Neural Networks to those technologies. If we add DL then these tech becomes Deep Computer Vision, Deep NLP and Deep RL. Adding DL to these technologies can substantially improve accuracy of these algorithms.

End result we can add DL to these tech.

10 case studies

10 use case where Deep Learning is used

1. google AL worked on large data set to develop a DL that can properly identify metastasis with 89$ accuracy compared to 73% by doctor.

Next to Cardio Cancer is huge issue

2. Spotify uses DL for recommendation engines. Compares songs listened by people with 2 billion songs spotify has - to detech patterns within the vast amount of data to find users with similar music taste further improving the selection. thus spotify increased its number of active users from 75 to 100 million.

3. goldspot discoveries - used DL for mineral exploration, earths non renewable natural resources are getting harder to find, companies had to dig very deep to find and cover more area both of which are costly and time consuming. number of goldspot discoveries dropeed drastically. Using DL to analyse geological data, GoldSpot discoveries developed a predictive method for finding gold deposits".

• they were able to correctly identify 86^ of existing gold deposits in Quebec
• despite working off data from just 4% of the total surface area.

4. Digitial Domain - Deep learning for Visual effects (VFX) - VFX are used in blockbuster movies to everyday commercials,

Character Thanos in Avengers game used Visual effects Used DL trained on higher as the scans of actors face to render the character in real time.

5. Ayasdi - Deep learning for anti money laundering - Money Laundering txns total approx 1-2 trillion dollars annually, an incredible 2-5% of the global GDP. Money laundering stops distribution of narcotics, terrorism, human traficing,

6. Deep Instict for Cyber Security - DL Powered security platform - found 12 infected endpoints - 10% of devices detected 1% cpu usage it tested 99% sucessful at finding and preventing incoming malware. prevented 140 cases of incoming ransomware

7. Doxcel - For productivity tracking - around 80% of construction projects are delivered over budget, which typical delivery times 20 months behind schedule" instant visibility is hard to get by humans. doxcel used drones and robots to collect information from construction site uses deep learning to analyze data collected from robots and drones that traverse a construction site. this can detect deviations from the planned schedule and prompt the project team to act before delays start to pile up.

result - 38% labor productivity improvement and 11% under budget.

8. Amazon Rekognition Deep learning for facial recognition - still in infact criminals faces can be instantly matched and traced.

9. Zestimate - deep learning for real estate prices - AI used for apprising a land or house and esitimate cost. Zestimate uses a DL algo to analyze home charactersitcs, unique features, and market data to predict the price of almost 100 million homes in the US

10. Zestfinance - for loan approval - Deep learning algo analyzes a potential customers credit score and determines their risk of defaulting.

Reinforcement Learning

What is RL?

There are 3 main groups of algos in ML

Reinforcement Learning - Providing a reward or punishment , example teaching a robot to walk either give reward as 1 if it maintains balance and walk or -1 if it doesnot. robots to play soccer, play chess. this is the future of ML

Supervised Learning - Teach a machine to search - classification dog or cat? Categorize new images. first provide label data to algo

Unsupervised learning - Discovery of new patterns, Clustering of customers into groups based on their similarities.

Comparision between RL

Reinforcement learning - Its very close to human intelligence. Baby trying to walk - it stands up with left leg and falls down - gets pain - baby learns to not to do it - instead of baby is able to walk few steps and stand up and not fall it gets happy (reward). Child keep doing the same so walks long time. Convert this to system Pain is a punishment and happy is reward.

6 core advantages how it become cuting edge

John Langford - supervised learning popular researcher.

Advantage of RL

1. RL doesnot need masive amout of data

2. RL is innovative unlike supervised learning, which only imitates patterns in the original dataset" - algo can do better than teacher - RI can come up with new solutions never considered by humnas also.

3. if data is baised then supervised is based on it so baised - RL is free from biases.

4. RL combines - Exploration - machine tests new approaches on the fly to find better solutions.

Explotation - machine exploits the best solution it has found so far. Other algo require training and redeployment but RI keeps going...

5. RI is goal oriented - Objective oriented and not like find dog or cat etc.

Example : Robots playing soccer and drive car. Algo maximising investment on the return.

6. RI does not require retraining because adapts to new environment automatically

how RL helped marketing fields

Alibaba

1. Recommender Engine - with real time based - Dynamic

   Netflix is not dynamic but Recommender engines can be used but online shopping is the right use case as changes occurs very fast.

2. Optimizing advertiesements.

3. Selecting best content there by maximising profits. Moonrise

4. Increase lifetime customer value -

5. Inverse RL - shopping - save money or feature or brand

Use cases of RL in business

NLP

NLP is the field of AI concerned with the interactions between computers and human (natural) languages, in particular how to program computers to process and analyse large amounts of natural language data.

But before we dive into NLP we need to understand the two types of data that exist: Structured and unstructured. We humans communicate predominantly in unstructured data, specifically via natural language. And that’s why, based on various sources, unstructured data accounts for 70-90% of data in the world. Most of this data contains text in some form – whether written or audio.

There are Two parts to NLP : Natural Language Understanding (NLU) and Natural Language Generation (NLG).

NLU - refers to mapping the given input from Natural Language into formal representation and analysing it. If the input is in the form of audio, then Speech Recognition is applied first to convert it into text. Then the hard part starts – interpreting the meaning from the text. Human language is complex, some words, for example, “bank” or “leaves” can carry different meaning depending on the context they are in. This is where advanced machine learning models such as Google’s state-of-the-art BERT come in to interpret the meaning.

NLG - If the text needs to be put into audio, like in the case of Siri or Alexa, then Speech Generation comes in, but this is not always required

Banking - Operational costs saving from using chatbots will grow from $209 m to $7.3 billion by 2023.

Google BERT -

Chatbots

Banking - Operational costs saving from using chatbots will grow from $209 m to $7.3 billion by 2023.

• Real estate agent bots
• Loan calculator bots
• Life coach bots
• politician bots
• investment advice bots
• home services bots
• appointment - setting bots
• employee on boarding bots

Simple chatbots - some predefined questions and answers

ManyChats -

Advanced chatbots - dependes on NLP specically NLU

watson assistant

autodesk uses - watson assistant.

Analyse document and extracts product related information from them at 97% accuracy.

Autodesk - chat bot - 1000k users rep Boston consulting - NLP to analyse survey data...

JPMogan uses NLP for document review, COIN (Contract intelligence) uses nlp

youtube - content management - illegal conent spreading..human moderator replaed by NLP..

8allocate - sentiment analysis...how customers feel about your brand allows you to adjust your sales and marketing strategy

Virtual processing assistance -

SKAI - customer feedback analysis

Booking.com - uses powerful translation so non english speaking people are able to use..

Computer Vision -

Computer vision is a field of artificial intelligence that trains machines to understand and interpret the visual world. The barcode scanner is one early example of CV in use camera identifies a face and auto-focuses on it; when Facebook auto-suggests whom to tag in a photo; even when a car drives automatically – these are all examples of CV in action.

Two types of CV exist: Classic and deep learning

Classic CV - relies on pre-built libraries of features. It collects images, labels them according to similar characteristics, and groups them in a dataset, or library of features. We have to choose the right library for our purpose: To detect, say, a face in an image, a library of facial features will identify the eyes, nose, and mouth according to its preexisting trait sets. The beauty of classic CV is that it returns results quickly and accurately.

Deep learning CV - requires neural networks to function, specifically convolutional neural networks (CNN). The key difference between the two types is that classic CV uses features that we have already input into the library, while deep learning CV generates its own library of characteristics by using its CNN.

Use cases

1. Image classification
2. Image segmentation -
3. Object detection
4. object tracking -
5. image generation
6. edge detection -
7. face detection
8. facial recognition
9. optical charater recognition
10. pattern detection
11. feature matching -

Harvest Croo Robotics Computer vision for automatic crop harvesting -

Ebay Image Search - take photo and search

https://www.bluelife.ai/consulting