Machine Learning – “Field of study that gives computers the ability to learn with out being explicitly programmed.”

Data Science – “Science of extracting knowledge and insights from data”

– Arthur Samuel (1959)

What are the key factors driving our motivation to learn and implement AI for businesses today?

Motivation behind AI is its success and impact on various critical business functions across many industries today. AI is changing the way, people live and interact with society, how businesses operate and compete. This success is a result of 2 to 3 decades of committed and sincere efforts from many researchers and scientists in the fields of AI and computing. In addition to business, data and technology innovation, this AI research community is investing additional time to address so many adoption barriers around explainability, various kinds of biases, and many more social and ethical challenges that AI could face while we experience the benefits of this constant innovation. Looking at the world ahead of us, all of us must face this AI experience and contribute to it in different ways. So, if you have not started this journey, you need make couple of initial steps to contribute and take advantage of this innovation.

Artificial intelligence, machine learning and deep learning are doing something that once seemed unimaginable a few decades before. They are transforming heavily regulated industries, such as healthcare and life sciences, financial services and trading, and many others. For healthcare and life sciences, Artificial intelligence takes the role of an experienced clinical assistant who helps doctors make faster and more reliable diagnoses. We already see AI applications in the areas of imaging and diagnostics, and oncology. Machine learning has the potential to improve remote patient monitoring. AI algorithms can take information from electronic health records, prescriptions, insurance records and even wearable sensor devices to design a personalized treatment plan for patients. These AI-related technologies accelerate the discovery and creation of new medicines and drugs. There is a broad consensus amongst insiders that healthcare is being transformed for the better because of AI. The opportunities and potential are limitless. Healthcare is going to be one of those industries that is elevated and made better by machine learning and artificial intelligence.

Who should get involved in this transformation?

Decision of building an AI company should start at senior executive level such as CTO, CIO, CDO, CAIO. Role of a CLO is evolving rapidly to shape structure of the AI organization. One of these roles along with AI leadership should take initiative to design AI strategy. Business leadership from certain high impact functions such as sales, marketing, and finance should be part of this initiative governing the decisions and closely monitors the impact at every stage of the progression

This can be used as a guidance document for senior executives, AI managers, AI Engineers, and other business and technical teams who closely on these AI projects. Senior executives and project managers receive guidance around leading an AI strategy, identifying the value drivers for business case, knowing several factors which impacts the company culture, society, and customers. AI execution team can be use this as guideline for prioritizing right use cases, designing, and creating the right pipelines, and implementing best practices for implementation and operation of these AI projects

What is the scope of the content here?

Main intent here is to share guidance to the new teams on how to get started with AI strategy and journey. It is also important to get through the hype and get realistic view of what AI can do in your scenarios. This document covers the following topics

1. AI Overview, Terminology, Definitions
2. Building an AI strategy
3. Executing AI projects
4. Driving Adoption
5. Other considerations and resources

1.    AI Overview?

1.1  AI Introduction

AI is divided into two classes, 1) ANI – Artificial Narrow Intelligence, 2) AGI – Artificial general intelligence. Smart speakers, self-driving cars, web search, AI farming and factories are examples of ANI. AGI is some thing that human can do involving emotional intelligence. There is a lot of progress made on ANI, but current state of the research is far away from replacing people using AGI. For the rest of the discussion, we will refer to AI as ANI in general. Overall value of AI is estimated to be $13 trillion by 2030.

1.2  Machine Learning

Machine learning (ML) is a statistical technique used to predict the target variable based on the set of the key features available within data. ML can be implemented in 3 scenarios, 1) Supervised Learning, 2) Unsupervised Learning, 3) Reinforced learning. Supervised learning expects input to output mapping available in the test data to predict the output for new inputs scenarios. Spam filtering, speech recognition, machine translation, online advertising, self-driving car, visual inspection are examples of supervised machine learning. Unsupervised learning predicts the output as clusters based on the test inputs with out any given outputs. Customer segmentation is a classic example of unsupervised learning. Reinforced learning is based on the rules and rewards for doing the right action. Learning chess, operating robot is an example of a reinforced learning.

Supervised learning can be used to implement, 1) regression – to predict the continuous valued output (e.g.: Price of a house given square foot and built year), 2) classification – to predict discrete values (e.g.: Classifying tumor  to be malignant or benign based on the tumor size)

There are various algorithms and optimization techniques available to implement machine learning including linear regression, multi variate regression, logistic regression, gradient descent, regularization, neural networks, back propagation, support vector machines, k-means clustering, dimensionality reduction, anomaly detection, recommender systems, stochastic gradient descent, online learning, etc. Below are some of the core algorithms in different categories. Each of these algorithms along with optimization techniques will be discussed in detail in separate discussions.

Supervised Learning:

Regression: Linear regression, Logistic regression

Classification: Decision trees, Random forests, Naive Bayes, Neural networks, Support vector machines

Recommendation engines: User based affinity, Item based affinity

Unsupervised Learning:

Clustering: K-Means, K-nearest neighbors

Association rules: Market basket analysis

Deep learning and ML are used interchangeably sometimes. Deep learning is implemented for problems that requires complex interconnections between several input groups and variables. In a nutshell, ML technique implemented using neural network with multiple layers (perceptron) is referred to as deep learning. Some of the use cases include, 1) demand prediction based on the price, shipping cost, marketing, material type, 2) identifying the face of a person given a image with 1000×1000 pixels.

Machine learning is perfect subset of AI. Deep learning is a perfect subset or special case of ML. AI and data science overlaps to some extent. Boundary between machine learning and data science is little fuzzy today, mainly in terms of how companies are calling and referring to it. Some companies may solve all mapping and decision insights through a common expertise, but some may have very clearly isolated roles for ML and data science. Output of the machine learning is a piece of software that helps predict outcome. Output of data science is slide deck with business insights and recommendations. One classic example is online advertising system. Machine learning processes all ad-clicks 24 hours a day to predict for right targets. Data science provides insights around right sales teams and channels to drive more sales

1.3  Working with Data

Quality and dimensionality of data is especially important to implement AI systems. Typical data set considered for supervised learning consist of several input variables called features, output variable called target variable that needs to be predicted for new scenarios based on the trained model. Data can be acquired using, a) manual labeling, b) from observing behaviors, c) download from websites and partnerships. Data can be generated by users or machines

Data needs to maintain good quality to bring value of it. Data can have different quality issues including incorrect labels, missing values, inconsistent measures, etc. Data can be of structured (e.g.: data bases) or unstructured data (e.g.: images, audio, free hand text). Data engineering techniques need to be applied to process various kinds of data and to maintain accepted quality required to run AI processes.

There are various preparation techniques available in pre-processing.

Cleaning – Removing the bad or irrelevant data records that could add bias or wrongly influence the results

Standardization – Formatting of the dates, names, consistency of reference data (e.g.: state abbreviations), currencies, etc.

Binning (ranges) – Converting continuous variable into ranges (e.g.: converting the customer targets into 3 classes/buckets based on volume of sales or number of customers in each bucket)

Indicator variables (text to numeric) – Used for algorithmic performance or consistency (e.g.: Converting male/female to 1 and 2)

Data imputation – Providing data for missing values

Centering and scaling – Adjusting the data ranges so that they are in the same scale with out any bias in weights for different input variables. For example, sales volume is in the range of thousands, ages in the range of tens. These are all converted to the scale of 0 to 1 or -1 to +1 depending of the input data points.

Text: tf-idf (Text frequency or Inverse document frequency) – Used to convert text data to numerical data for prediction purposes.

1.4  AI Company

What makes an AI company?

Just by implementing machine learning or deep learning, company can not be classified as AI company. To be an AI company, it needs to satisfy the following criteria at minimum

• Strategic data acquisition
• Unified data warehouse or data lake
• Pervasive automation
• New roles (e.g.: Machine learning engineer, data scientist, etc.) and division of labor

1.5  Practical Limits of AI/ML

Senior executives are over-inflating the expectations of AI. Academic literature publishes only success stories. No one is publishing the practical limits of where AI can and can not do well. In general ANI scenarios where we can come up with answers with out deep thinking or emotional intelligence. ML can take care of most of the supervised learning use cases discussed above including spam filtering, speech recognition, machine translation, online advertising, self-driving car, visual inspection.

What AI can or cannot do today?

• AI cannot write a 50-page report by analyzing the market or conducting market research
• If toy arrived two days late after a kid’s birthday, it could help process the refund request and send a refund back to credit card. But AI can not send a consolation email by apologizing and offering any special discounts on next sale
• With just 1000 examples, system can not be trained to generate a sympathetic response. Sometimes it may send junk if the model is over trained on inputs
• For self-driving car, AI can only detect the cars or objects, but it cannot detect human gestures, bike turn signals
• For X-ray diagnosis, AI can diagnose pneumonia from ~10,000 labeled images. But it cannot diagnose pneumonia from 10 images of medical textbook chapter explaining pneumonia

ML tends to work well when learning a simple concept and there is a lot of data available. ML tends to work poorly when learning a complex concept from small amounts of data and it is asked to perform on new types of data

Recommendation: Please conduct a technical diligence before deciding with AI implementation for any specific use cases. This technical diligence can be couple of weeks of assessment to validate the feasibility and to agree on the acceptance criteria

2.    Building AI Strategy

2.1 AI Transformation Playbook

If you not already into AI, you can follow high level steps below to create AI transformation

1. Execute pilot projects to gain momentum
2. Build an in-house AI team
3. Provide broad AI training
4. Develop an AI strategy
5. Develop internal and external communications

Execute Pilot Projects: More important for the initial project to succeed rather evaluating based on value and other factors. This builds confidence given this is new area. Show traction with in 6 to 12 months. These can be in-house or outsourced

Build A Team: AI function can be under CTO, CIO, CDO, or CAIO. Onboard functional teams from multiple business units to work with this AI team collaborate in building the AI strategy and transformation

Provide AI Training: Training content should be catered based on the role level. Executives and senior leaders should learn key AI use cases, impact on business, AI strategy, and resource allocation. Leaders of division working on AI projects should learn setting the project direction (technical and business diligence), resource allocation, monitor progress and evaluating the success. AI engineers should learn to build software, gather, process and clean data, and execute on specific projects

Develop AI Strategy:

• Leverage AI to create an advantage specific to your industry sector
• Design strategy aligned with the “Virtuous cycle of AI” by following the feedback loop of on boarding more users, gathering more data, ultimately to create better product
• Consider creating data strategy aligned with your organizational objectives to support strategic data acquisition and to create enterprise wide access and consolidated view through unified data warehouse across your business functions. This ensures better value through AI and analytics initiatives
• Create network effects and platform advantages. In industries with “Winner take all” dynamics, AI can be an accelerator. It can help set competitive advantage.

Develop communications: Develop a communications plan to onboard internal teams and manage external relations. These include investor relations, government relations, consumer/user education, talent/recruitment, internal communications

Key considerations:

• Be realistic about what AI can and can not do given limitations of technology, data, and engineering resources
• Pair engineering talent with business talent and work across functionally to find feasible and valuable projects
• Plan for AI development to be an iterative process, with multiple attempts needed to succeed
• Work with AI team to establish timeline estimates, milestones, KPIs, etc.
• Don’t wait until you collect tons of data, but start your journey early with the data that you have and experience the behavior of the system, adjust necessary inputs, gather more data as available and required
• You do not need to wait to build entire AI team, just start with light team with some initial training to start experimenting and assessing the feasibility of some of prioritized concepts and use cases

2.2 Roles in AI Team

It is especially important to understand AI pipeline across key use cases. This knowledge helps in estimation and resources allocation. In a simple smart speaker implementation, there are 4 components involved including 1) Trigger work / wake word detection, 2) Speech recognition, 3) Intent recognition, 4) Execute a joke. Step # 4 is special program to execute a command. There can be several other functions such as play music, volume up/down, make call, current time, units, conversion, sample question, etc. to be implemented as programs.

For self-driving car, there are 3 foundational steps involved, 1) Car detection, 2) Pedestrian detection, 3) Motion planning. There can be other components such as lane detection, object detection, traffic light detection, trajectory prediction. But Motion planning is special program that helps steer / accelerate or break.

Here are set of example roles that can get involved across broad variety of AI projects.

• Software engineer – helps implement special programs such as joke execution, ensure self-driving reliability.
• Machine learning engineer – helps implement A->B mapping for speech recognition, object detection
• Machine learning researcher – helps extend the state of the art in ML which includes advanced functionality. Sometimes the above two roles are combined into a single role called “Applied ML Scientist”
• Data scientist – Examines the data and provide insights. Make presentations to executive team
• Data engineer – Organizes data. Make sure data is saved in easily accessible, secure, and cost-effective way
• AI product manager – help decide what to build; what is feasible and valuable

Get started with a small team including 1 machine learning engineer/data scientist, 1 software engineer, or pull an existing team member and train that person on machine learning depending on the initial need for assessment and piloting.

2.3 Taking Your First Step

With the knowledge that you gained so far here; you can jump into couple of first steps towards your AI journey.

• Gather a few motivated folks who can join any of the AI courses. Create a reading group to learn from each other.
• Gather some business use cases, pick that is critical and quick to implement by aligning and brainstorming with business team
• Hire one or two machine learning and data science people to help based on the type of use cases and intensity of initial work planned
• Hire or appoint an AI leader (VP AI, CAIO, etc.)
• Discuss with CEO and board possibilities of AI transformation. Create business case justifying the value of the AI on business and other drivers such as being effective, being more competitive, if it were good at AI?

2.4 Use Cases or Applications

Based on consolidation of several surveys conducted, please find several use case categories along with some examples.

Computer Vision:

1. Image classification/object recognition/face recognition systems: used in many applications including, a) identifying persons and pets in home security camaras, b) face recognition for computers and security systems
2. Object detection: used in self driving cars to detect objects like cars, bicycles, trees before determining motion planning
3. Image segmentation: used in photo editing applications to detect noise and unwanted patterns from raw images
4. Tracking: Used to track social distancing compliance and notify violation based on the distance in the rule

Natural Language Processing:

1. Text classification: used for sentiment recognition from customer reviews and email spam classification filtering
2. Information retrieval: used in smart web search based on the context of the user
3. Name entity recognition: used for detecting names and key words in search applications
4. Machine translation: used for language translation (e.g.: Spanish to English)
5. Parsing and parts-of-speech tagging: Used in word processors, resume builders, linked-in application to correct grammar in sentences

Speech:

1. Speech recognition: used in speech to text applications such as medical transcription
2. Trigger word / wakeword detection: used in smart speakers like amazon echo, google nest, etc.
3. Speaker ID: used to store voice prints or voice biometrics to uniquely identify a speaker like identifying fingerprints
4. Speech synthesis: used in text to speech/TTS applications

Robotics:

1. Perception: used to figure out what is in the world around you. This is also use case for self-driving car.
2. Motion planning: finding a path for the robot to follow. This is also used in self-driving car
3. Control: sending commands to the motor to follow a path. Used to control acceleration in the self-driving cars.

General machine learning:

1. Unstructured data: As we had already seen, this is used for processing image, audio and text applications including face detection, speaker ID, and language translation
2. Structured data: Predicting trends in business activity like sales trends, classifying tumors as cancerous or not, segmenting the customer profiles

Unsupervised learning:

1. Clustering potato chip sales: Identifying segments of potato chip brands based on the price per packet and # of packets bought by each customer
2. Identifying cats from YouTube videos: Given data (with out specific desired output labels), find some thing interesting about the data

Transfer Learning:

Learn from task A, and use the knowledge to help on task B

1. Self-driving car: There is not of data about detecting car with thousands of images. The learning from this car detection model can be applied to detecting Golf carts though you have fewer images to work with

Reinforcement learning:

Use a “reward signal” to tell the AI when it is going well or poorly. It automatically learns to maximize its rewards

1. Enabling a robot helicopter to learn flying: Use GPS, accelerometers, and compass to track movements and provide inputs/rewards to enable helicopter learn flying on its own
2. Helping computer learn to play chess by providing rewards or correct moves

GANs (Generative Adversarial Networks):

GANs are a type of neural network that relies on two different components, one to generate content (adaptive network) and the other to test it (discriminator)– thus “generative adversarial.’ The discriminative network attempt to distinguish between real and generated content, helping the generator learn through each iteration.

It is a type of unsupervised training in which the adaptive network attempts to fool the discriminator, improving accuracy and helping the machine learn what constitutes an acceptable degree of accuracy.

So, what does this mean for you? GANs is a new way for neural networks to generate accurate, high-quality content through unsupervised training, making it cheaper and faster to produce content required for deep learning training.

1. Text to image synthesis: Creating images based on the text you type.
2. Data manipulation: manipulating photos to alter or remove part of the photograph. GANs feature could allow us to manipulate model distribution, removing the sunglasses for better recognition
3. Drug discovery: Drug discovery is a slow process. Traditionally, a single hypothesis is tested over a range of years if not decades with massive human input and substantial resource investment. GANs, on the other hand, can rapidly generate novel biological components (e.g.: novel proteins) to test hypothesis simultaneously
4. Art: GANs can manipulate images based on training data to create new works of art that mimic the training rules and add unique details that would please buyers
5. Privacy maintenance: GANs is already being used to create alternative examples to share in instances where we need to share sensitive data with a third party. The generation can approximate the data without revealing anything potentially compromising
6. Music generation: It is expensive and time-consuming to score scenes, and expensive and time-consuming to license already created music for that same distribution. Enter GANs. It uses a discriminator to learn the distributions of melodies and exploiting prior knowledge to generate new melodies based on our parameters. Example is a MidiNet which competes against Google’s MelodyRNN

Knowledge Graphs:

Knowledge graphs provide framework which can represent complex relationships between people customer groups, different elements, etc. and helps navigate and analyze these relationships to gain analytic value

Background: Combinatorial chemistry has produced huge amounts of chemical libraries and data banks which include prospective drugs. Traditional methods fail to address this question, “how do we take advantage of this data to identify prospective nature of a compound as a vital drug?

Solution: Knowledge graphs, however, provide the framework which can make drug discovery much more efficient, effective, and approachable. This radical advancement in technology can model biological knowledge complexity as it is found at its core. With concepts such as hyper relationships, type hierarchies, automated reasoning, and analytics we can finally model, represent, and query biological knowledge at an unprecedented scale

3       Building AI Projects

Starting AI project involves multiple core components, 1) workflow of a project, 2) gathering and prioritizing projects, 3) Organize data, 4) Creating team for execution. This activity is common for any companies of any industry or type (for profit or nonprofit).

3.1 AI Workflows

Data science and Machine Learning projects have different workflows because DS projects involves deep actionable insights, whereas machine learning projects focus on just detection.

DS Workflow Examples: (Insights and Recommendations)

Example 1: Optimizing a sales funnel

Sales funnel involves, Website -> Product Search -> Shopping cart -> Checkout

Key steps in the data science project involves the following steps

• Collect data: Figuring out different market sources, opportunity, pricing, demand, key features, countries, cultures, patterns, etc.
• Analyze data: Checkout dropouts, average time spent, any other misses and reasons. Observe other trends, patterns related to seasonality, holiday, events, and other timing, overpricing, etc.
  1. Iterate many times to recapture and reanalyze data periodically after feedback and to capture changing events and trends
• Suggest hypothesis and actions (deploy changes, actions and reanalyze changes): Hypothesis may generate some insights. Check out policies may have some influence

Example 2: Manufacturing line

Process involves: Mix Clay with water -> Shape Mug -> Add Glaze -> Fire Kiln -> Final Inspection

Most common problem in this process is detecting damage or defects and find upstream elements that are causing more of these defects

Key steps in the DS workflow:

• Collect data: Who supplied the data, what is the moisture content by category of soil, humidity in the batch processed, temperate used to make the mug
• Analyze data: Understanding what led to damage based on the features above (color and type of clay, place of the clay, time in the day, temperature, humidity, etc.)
• Suggest an action: Determine the ideal temperature, procuring the right soil, keep the right humidity levels, etc.

ML workflows examples: (A -> B mapping)

Example 1: Speech Recognition (Echo/Alexa or Google)

Process: Provide Audio clip -> detects if the device should respond to Alexa key word

Key steps in the workflow involves:

• Collect data: Data of words including (Alexa) and other words (Hello, others) from different people with right distribution across geographies and accents
• Train the model (A> B): For each audio, figure out the outcome
  1. #1, Alexa
  2. #2, Hello
  3. #3, Alexa
  4. Iterate many times until the process is good enough

• Deploy the model: After deploying the model, get the data back, maintain and update the model. Example: It is not able to capture certain accents like British, then model needs to be retained.

Example 2: Self driving car:

Process: Provide Image of the car relative to road -> Detect the position of other cars

Key steps in workflow:

• Collect data: Collect bunch of car pictures. Draw rectangles around cars. It is usually done with software
• Train the model: Software may find tree as a car

Deploy the model: Observe the safety. Capture data where it is not detecting the cars, ex: golf cars or 3 wheeler not detected well. Tree is detected as car

3.2 Use cases and Data Applications

In many industries including Mfg., agriculture, healthcare, data is digitized. Some of the channels such as surveys, doctors’ notes, and transcription leaving good availability of the data. AI is helping in many job functions. DS and ML work together in many aspects. Here are several functional segments where data can be leveraged effectively to address various use cases

Sales functions:

• DS: optimize sales funnel
• ML: Automated targeting/Dynamic targeting (in Pharma), Automated lead sorting
  1. Who are the priority profiles and influencers (CEO – High, Marketing managers – Med, Office staff – Medium, Other connections – low)

Manufacturing:

• DS: Optimize manufacturing line
• ML: Automated Visual Inspection
  1. Find scratches, defects, dents, quality in the manufacturing line as quickly and close to the start will result in huge success

Recruiting:

• DS: Optimize Recruiting funnel
  1. Email outreach -> Resume screen -> Phone Screen -> on-site interview -> Offer
• Automated resume screening
  1. Issue is that many wrong profiles are passing through phone screens and resume scans
  2. Screening resumes and building custom phone screen recommendations to filter candidates better in these two steps results in huge efficiency
  3. Ethical fairness / Bias is another consideration that needs to be addressed. It is getting better.

Marketing:

• DS: A/B testing
  1. Optimize their web sites for right features to catch customer attention
  2. Red theme vs green themes for buttons or other marketing content. User Experience plays a role too.
• ML: Web site recommendations
  1. Recommending right product

Agriculture:

• DS: Crop Analytics
  1. Analyzing optimal conditions for crop (solar conditions, weather, water price in the market, soil condition, what to plan, where to plant, maintain soil, nutrition, erosion, a lot of other factors)
• ML: Precision agriculture
  1. Spotting the weeds for picking and killing
  2. Figuring out planting distance

3.3 Selecting and Prioritizing an AI Project

There are two important considerations, 1) what AI can do? 2) What is valuable for business? You need to take the intersection of these two for prioritization. AI experts (AI & Deep learning) and Domain Experts (Sales and marketing functions and healthcare, agriculture, and product domains) should form as AI cross functional team to brainstorm and help prioritize on these investments

Brainstorming framework:

 Automating a task vs. a job:

Think about automating the tasks rather than automating jobs. Some of the examples include call center routing, and radiology

Call center routing example:

• There are many job activities involved in call centers including picking up phone calls, answering emails, answering chats, Processing refunds, others
• Pick the task that is more efficient and can be automated. Example: Email routing, Chat response, etc.

Radiologist example:

• Radiologists perform many jobs including Reading X rays and other diagnostic films, engaging new learning and education, training juniors, discussing with patients, and engaging in research activities
• Pick 1 or a few such as Automation of X-ray reading and detection

Key business drivers:

What are the main drivers of your business values? Find a solution that augment this

Other key considerations:

What are the other main points in your business?

You can program without Big data. But it never hurts to have more data. Data makes some businesses defensible (web search). With small data, you can still make progress. For coffee mugs example: 1 M works well, but you may not have that many examples. But having 10, or 100 or 1000 should work fine to start building a starting point. Do not give up, learn, and adapt from the small data through feedback. This help build the capability through small size investments

• Look for some early wins by taking small and high values project to drive confidence and motivation in teams. Share stories of success and encouragement.
• Since some of the projects takes months to execute, effective prioritization around timing, complexity, business value, data availability is a key to success
• Show due diligence: Technical and business diligence are both importance to validate, assess and prioritize these projects

Technical diligence:

• Can AI system meet desired performance? Ex.: 99% accuracy possible
• How much of data is needed? Is the data available in the desired format and quality? Level of effort required to performing cleaning
• People engagement and skills required
• Timeline (how long is the best timeline)

Business Diligence: Two import themes

• Improve current business
  1. Lower costs through automation, shortening timelines, lowering efforts
  2. Increase revenue through efficient business processes
• Setting up new business or new strategy or new event

Ethical Diligence:

Please refer to below sections regarding ethical considerations and social impact on AI to drive adoption and address several potential issues around various kinds of bias

Build vs Buy (aged old question in the IT world)

• ML Projects can mostly be done in house or outsourced. Most details to be come in the AI transformation play book
• Data science project are commonly done in-house since they mostly connected to business teams. Some business consulting organizations who are deep experts in business can help with these.
• Another thumb rule: avoid rebuilding the industry standard capabilities if they can be easily integrated in your systems.
• Thumb Rule: Do not sprint in front of the train. Always leverage industry assets and standards. Do not over engineer or re-invent standard processes in the industry.
• A few weeks of assessments will always help before jumping directly into implementation

3.4 Working with an AI Team

You have cool ideas, but do not have a team. Do not worry. Building expertise is essential. Follow gradual approach to build this team. With the help of a few quick on-line programs, you can help build this initial team who can work on some quick assessments to get quick wins

Define an acceptance criterion. Goal is to detect product defects with 95% accuracy. Question is how to measure this accuracy? As Business team, provide an AI team with a test data set on which to measure their performance. Consult AI experts to design this process

Another question is around how to leverage the statistical models to even estimate the system performance. You don’t need to address this question right away, but as you get mature in capabilities, this can help achieve performance and improve the model quality through this automated validation.

How AI team think about the data? AI Teams treat the data into 3 categories

• Training data
• Validation data
• Test data

Common pitfalls to avoid:

• Do not expect 100% accuracy on any model.
• Do not over fit the model (Optimize the model through iterations or regularize the models to remove bias and overfitting)
• ML can have limitations
• Insufficient data
  • Effectively chose algorithms as the scenario and data availability as some of the algorithms perform extremely well with small data sets
  • Get more data from operational sources if feasible
  • Create some test data or have experts produce more examples as required
• Mislabeled data – Data is incorrectly labelled as defect item in the training data set
  • You will need to spend some time in doing the data wrangling, cleaning, normalization, or standardization
• Ambiguous data – How small scratch is interpreted by business vs AI team. Need better business alignment while you are doing the AI modeling
  • Better governance or consensus on decisions

3.5 Technical Tools for AI Team

Here are some AI open source frameworks to leverage to build AI/deep learning models.

• TensorFlow
• Pytorch
• Keras
• MxNet
• CNTK
• Caffe
• PaddlePaddle
• Scikit-learn
• R
• Weka

Other resources:

• Refer to research publications under “ArXiv”, pronounced as Archive
• Refer to open source repositories: GitHub is popular one. One of the example sources in GitHub is Face recognition. You will find several downloadable resources, mostly open source on GitHub. Feel free to use these in your projects. Check for license

CPU vs. GPU:

• CPU: General computation. Intel and AMD are the common manufacturers. Used for low intense workloads
• GPU: Graphic intensive processing for videos, fancy graphics for video games, processing large NN/DL algorithms. Nvidia, Qualcomm, Google are the most common manufacturers for these. Used to high intense workloads for high performance computations and heave deep learning jobs

Cloud vs. On-premises POV:

Cloud deployments make the process fast. You will need to manage the infrastructure and safe issues effectively.

Edge deployments:

Computation happens with in the device to avoid latency of the processing and transfer rates. Most found example is self-driving cars where processor is installed within the CAR itself. Home smart speakers are another example for edge deployments.

4       Driving AI Adoption

4.1 AI Impact on Society and Ethics

It is important remove fears in Society from AI being killer intelligent. At the same time do not remove faith on its potential to solve super critical issues which we could not have solved otherwise. You need to hit the right balance between being too optimistic and too pessimistic. AI cannot solve everything today, but it started to transform many industries. Unless you experience its power and follow its evolution, it is exceedingly difficult to convince people to use it. It is particularly important to educate around reality vs. hype around AI.

Here are some consideration and limitations that creates barriers around AI adoption. But by working through change management and practical demonstration, these issues can be addressed.

• Performance limitations: Some key considerations around system requirements is important to address these performance limitations and build confidence
• Explainability: It hard but sometimes doable. Neural networks and deep learning models can solve most complex problems, but at the same time they are complex to explain with hypothesis. A lot of researchers are working on this topic. Often it not impossible.
• Biased AI: A lot of data stored in web is making AI to be based towards predictions. To be explained later.
• Adversarial attacks on AI: to be explained later

Discrimination Bias:

AI is learning unhealthy stereotypes from data on the internet. Which may make the predictions and ultimate decisions biased towards one group or other. Here are some examples of it.

• Hiring tool that discriminated against women
• Facial recognition working better for light-skinned than dark-skinned individuals
• Bank loan approvals show bias based on the social interactions
• Toxic effect of reinforcing unhealthy stereotypes

Combating this kind of bias requires following steps

• Technical solutions to use less biased and/or more inclusive data, incorporate more features to remove these bias scenarios
• Transparency and/or auditing processes
• Diverse workforce to create less biased applications

Adversarial physical attacks on AI:

AI system perceives differently and reads differently than what humans do, which may have the objects or things classify differently. Computer systems are much better than decoding QR codes. There are advantages and disadvantages to this. Combating these kinds of attacks incur extra costs in building more components around this. Here are some examples of physical attacks

• Keeping florescent stickers on the stop sign will make the system not detect as stop sign. This may impact road systems and self-driving cars in future
• Keeping another shiny gadget next to banana will make system incorrectly classify it as banana

Adverse users of AI:

There are some adverse uses of AI as well. People do some deep Fakes by making fraud transactions, synthesizing videos, etc.

• DeepFakes: Synthesizing video of people doing things they never did
• Undermining the democracy and privacy, Oppressive surveillance
• Generating fake comments on blogs and other sites
• Spam vs. anti-spam and fraud vs. anti-fraud

4.2 AI Impact on Economies and Jobs

Developing economies can jump directly leveraging the existing research. China and India are examples where they jumped straight to mobile phone in a leapfrog way. Same with mobile payments and online education. Developing economies can use leapfrog approach leverage latest innovation that happened around the world. Mobile phones, mobile payments, and online education are few examples of how some of these developing countries adopted quickly using this leapfrog approach. Here are a few ideas around how developing countries can build AI

• US and China are leading, but al AI communities are still immature
• Focus on AI to strengthen a country’s vertical industries
• Public-private partnerships to accelerate development
• Invest in education

AI Impact on Jobs:

AI is automation on steroids. It is having accelerated impact on jobs. It is important to note that AI is creating far more jobs than that it can displace by 2030. Accordingly McKinsey global institute 555-890 mil jobs created by 2030, while 400-800 mil jobs will be displaced, given net positive effect through AI adoption.

AI is still applied at task level, but not he jobs level. AI will not fully replace the job of radiologist, instead it can help automate some of the routine tasks like image or x-ray reading to help predict the pneumonia or other indications. This means that Radiologist does not need to worry about changing his profession, rather he can leverage this technology to better focus his time on human interactions, coaching his teams, conducting research, and provide inputs to system to automate some of the route tasks to save him time and leverage AI potential

5       References and Learning Aids

5.1 References

GANs (General Adversarial Networks):

https://medium.com/@ODSC/6-unique-gans-use-cases-24cab2aa924d

https://insilico.com/

Drug Discovery Knowledge Graphs:

https://blog.grakn.ai/drug-discovery-knowledge-graphs-46db4212777c