AI for general- Andrew Ng

Data

1. Unique for each business

2.  A------>B You can decide what is A and what is B

3. Be thoughtful while collecting data.. not all data are useful

4. Data cleanup important, else it will become garbage in-> Garbage out

Machine learning vs data science

1. Machine learning is the ability to learn without explicitely programmed

2. Data science is about extracting knowledge and insight from given data

As an example, 

• ad software uses ML model

• Data science team can tell that travel industry is not properly utilized and so, CEO can direct sales team to focus on travel industry

Data scientist can use ML models for their work, they can even use non-ML tools

AI company

1. Stretegic data acquisition.. for example free products simply to collect data

2. Unified data warehouse including GDPR compliance

3. New roles based on task need

Playbook to transform to AI company

1. Execute pilot projects to gain momentum

2. Build in-house AI team

3. Provide broad AI training

4. Develop internal (employees) and external communication (stakeholders)

What AI can do now
1. Anything a person can do with 1 second of thought can be automated (already or near soon)

,What AI can't do

1. Understanding signal from human gesture.. like stop sign, pick-up request sign

What makes AI problem easier

1. Learning a simple concept (<1 sec thought)

   1. Self driving car

2. Lots of available data

AI behaves poorly if

1. Data pattern deviates in test compared to training.. for example, x-ray image position tilts in the test 

Workflow of machine learning project

1. Collect data

2. Train the model

3. Deploy

Workflow of data science project

1. Collect data

2. Analyse data

3. Suggest hypotheses and action

AI knowledge & Domain knowledge

1. AI experts

2. Domain experts

3. Cross domain team who is combination of both above 

How to choose an AI project

• Think about automating tasks, not jobs

• What are the main driver of business value?

• What are the main pain points of the business?

Note that you can make progress even without big data

• For example, defensive web portal

Anaysis before finalising project

1. Is it feasible?-> Technical diligence

2. Is it valuable? -> Business diligence

   1. Lower cost

   2. efficiency

   3. Revenue generation

   4. Create brand new business

   5. ...

Build vs buy

• ML project can be in-house or outsourced

• Data science project is mostly in-house

• 
  

Working with AI team

• Define acceptance criteria well

• 
  

Real life problem with data

• Insufficient data

• Mislabelled data

• Ambigious labels

AI technical tools

• Tensorflow

• Pytorch

• Keras

• MXNet

• CNTK

• Caffe

• PaddlePaddle

• Scikit-learn

• R

• Weka

Week-3 Building AI in your company

Case study- smart speaker

Steps

1. Wakeword detection

2. Speech recognition

3. Intent recognition

4. Execution

Example

• Tell a joke

• Set timer for 10 mins

Case study - Self driving car

1. Sensor

   1. Image/Radar

   2. GPS

2. Object detection

   1. Car detection,  -> Supervised learning

   2. pedestrian detection -> Supervised learning

   3. Lane detection

   4. Traffic light detection

3. Motion planning 

   1. Software to tell about path and speed to avoid any accident

4. Steer/Accelarate/Break

Roles of an AI team

• Software engineer

  • Example - Joke execution

• Machine learning engineer

  • Gather data

  • Train the ML model

• Machine learning researcher

  • Extend state of the art in ML

• Applied ML scientist

  • Does work of ML engineer and ML researcher

• Data scientist

  • Get insight from the data

• Data engineer

  • Organize data

  • Data storage in secure, easy access and cost effective way

• AI product manager

  • Decide what AI product to build

AI transformation playbook

1. Execute pilot project to gain momentum

   1. Might not be the most valuable

2. Build AI team

   1. CEO-> AI team (Under CTO, CIO, CDO or new CAIO)

3. Provide AI training

   1. Curate content, not create

4. Develop AI strategy

   1. Woking on strategy as 1st step is not correct

   2. Build virtuous cycle

      1. Better product -> More users -> More data -> Better product

   3. Consider creating a data strategy

      1. Example is....Free photo service ...

   4. Create network effects & platform advantage

5. Develop internal and external communication

   1. Investor relationship

   2. Government relationship ... for example healthcare sector AI

   3. Consumer/user education

   4. Talent/Recruitment

   5. Internal communication

Avoid AI pitfall

• Be realistic about AI.. what can do. what can't

• Don't let ML engineers alone work ... pair them with business counterpart

• Don't expect that it will work first time itself

• Think you need superstar AI engineer

Taking your first step in AI

Some initial steps

• Get friends to learn about AI

• Start brainstorming projects

• Hire a few ML/DS people to help

• Hire or appoint an AI leader

• Discuss with CEO about possibilities of AI transformation

Survey of major AI applications

Computer vision example

• Image classification/ Object detection.... For example, CAT

• Object detection.... Tells the position of object

• Image segmentation .... draws precise boundary on object instead of rectangle

• Tracking

NLP

• Text classification

  • Sentiment recognition

  • Information retrieval.... Web search

  • Name entity retrieval ... For example, name of country, phone number, person name

  • Machine translation

  • Others -> Parsing, part of speech tagging

Speech

• Speech recognition

• Wakeword detection

• Speaker ID

• Speech synthesis .... convert text to speech

Robotics

• Perception ... Sensing what is around you

• Motion planning... Finding a path to follow

• Command ... Send command to the motors

General machine learning

• Unstructured data ... Audio, video, text

• structured data ... table 

Survey of major AI technologies

Clustering example

• Cluster of Potato chip sales 

Transfer learning

Benefit is that it needs less data in second step

Example: First learn car detection. Then train for Golf cart detection..

Reinforcement learning

Uses reward signal. Good reward is via +ve number reward.. Bad reward is via -ve number reward

GAN (Generative Adversial Network)

Synthesize new image

Knowledge graph

Example is right side panel for celebrity query... It is done by creating knowledge database containing name, DOB and other info

AI and society

Realistic view of AI

• Too optimistic

  • Not right that super intelligent robot killers are coming soon

• Too pessimistic

  • Not right that AI winter is coming soon

• Just right

  • AI can't do everything, but it can transform the society

Limitiation of AI

1. Performance

2. Explainability is very low - For example, how to know if AI is telling right about finding in x-ray 

3. Biased AI through biased data

4. Adversial attack 

Bias in AI

Example

• Discrimination against women

• Discrimination against dark skinned people

• Discrimination against minority

Combating bias

It happens due to bias in data

• Technical solutions

  • Zero out bias in the words

  • Use less biased inclusive data

• Transparency and auditing processes

• Diverse workforce

Adversarial attack on AI

Example

• Tampering image

  • Fooling that a bird, hammingbird is hammer

  • Fooling that Hare is Desk

• Physical attack

  • Temparing stop sign board

  • Keep something nearby banana will make AI fool that it is not Banana

• 
  

Adverse uses of AI

• DeepFakes

  • Creating fake video for someone claiming that they did it

• Undermining of democracy and privacy

  • Oppressive survelience

• Generating fake comments

• Spam vs anti-spam and fraud vs anti-fraud

AI and developing economics

• Developing economics gives lower level of job which will be terminated due to AI automation 

  • Solution - Leapfrog.... Train workforce so that some of them jump for upper level of work

  • Good sign

    • Some developing economy  saws better adoption of mobile payments, online education 

How developing economics can build AI

• Currently US and China are leading, but all AI communities are still immature

• Focus on AI to strengthen vertical economics

• Public-private partnership to accelerate development

• Invest in education

AI and Jobs

• Number of Jobs created is higher than number of Jobs eliminated

Solutions

• Safety net for finance by basic income

• Lifelong learning

• Politicial solutions

• Learn AI as well along with what you are doing

  • Help to become more valuable due to having domain expertise