Blog - 20 Must Knows About AI
Return to AI Research
1. What is AI? - Certain vs. Uncertain
2. How does AI get its intelligence? - Expert System (algorithms) vs. Machine Learning
3. Where is AI Working? - Selling us stuff
4. What are the primary roles of AI? Cognitive Assistance (Decisions) vs. Autonomy (Intelligent Automation)
5. Should it be called AI if Siri, Alexa, Google or Cortana helped?
4. Is AI going to be smarter than humans? - Artificial General Intelligence vs Billions of Narrow AI
4. What can AI do? - Cognitive Assistance vs. Autonomy
6. How does machine learning learn? ML, DL & RL
7. What can AI do similar to Humans?
8. What can AI do better than humans?
9. What does AI do worse than humans? context, takes a long time to learn, needs lots of data
10. Why are Autonomous Vehicles taking so long?
11. What are the limitations of data? - Passive vs. Active Data sets
12. When will AI become mainstream like the internet?
13. What is Intelligent Automation - Robotics (Physical) vs. Process (Digital)
14. Will AI drive Workforce Disruption and Reduction - Citizen Developers
15. What AI is a Must in Managing Complexity - 50 complex challenges
16. Why AI Gives Us a change to Keep Up With Complexity - complexity is growing
17. Why AI Requires Quailty Data Sets - Large for Deep Learning
18. Why AI is biased
19. AI is based on incomplete data
20. AI doesn't understand
21. AI or Humans Can't Accurately Predict an Uncertain Future
BLOG
1. What is AI?
There may be as many definitions of Artificial Intelligence as there are people. Merriam-Webster defines Artificial Intelligence as:
a branch of computer science dealing with the simulation of intelligent behavior in computers, the capability of a machine to imitate intelligent human behavior
While I was once a computer engine, I struggle understanding some of the explanations of AI that surface at the top of my Google search results. These explanations may be accurate and useful for Machine Learning engineers and data scientists, yet offer little value to most people I'm trying to help. Like knowing what a car can do to drive it, these executives and managers need to know AI basic to make decisons on leveraging AI to improve their businesses or improve clinical diagnoses, prognoses and treatments. This simple definition of Artificial Intelligence may be a helpful start:
a computer making decisions or assisting humans in making decisons when uncertainty exists. computer-based decisions or advice is based on both human learning and machine learning
If you withdraw money from the ATM, a fixed algorithm will complete a certain decision of moving money out of your account. If you Google your bank, the search engine will be 99.5% certain you want Catskill Hudson Bank rather than a 0.5% chance your want a bank on the Hudson River.
If the probability of the correct decision is 100% certain, such as dispensing the $100 to your ATM based on your request, this is not considered Artificial Intelligence. This is simple conditional (if-then) reasoning that can be represented by computer logic or even built into a mechanical device. The ATM may seem intelligent, yet it is simply following a specific instructions like a wall switch turing on a light.
Google search providing a link to Catskill Hudson Bank on top of the search results, rather than a news story of an accident on the bank of the Hudson River is based on both human learning and machine learning.
It is important for people to understand that the foundations of AI is human learning, machine learning and a hybrid of the two. More on this in the next "must know" AI basics.
2. How does AI get its intelligence? Expert System (Human Learning) vs. Machine Learning
One May 4 1997, Deep Blue beat world champion Garry Kasparov in chess. IBM programmed Deep Blue to use brute computational force to evaluate millions of positions. While reported as artificial intelligence catching up to human intelligence, this exposed the limitations of using expert systems in AI.
An expert system gets its intelligence from a human expert. The expert knowledge is represented by an algorithm that will make decisions in uncertain situation.
Machine learning gets its knowledge from a set of data. The algorithm is generated by a computer program rather than an expert. The algorithm represents knowledge based on the data.
An expert system would be an algorithm based on a human expert's knowledge that if someone applies for a $200,000 mortgage with a 700 credit score, makes $100k per year, they will would have less than a 5% rate of default.
A machine learning model would be an algorithm based on hundreds of thousands of mortgage holders' applications credit scores, annual income, and mortgage amount along with historical default of these mortgage applicants. A copmputer program generates an algorithm that represents this set of data.
While this explains where AI gets its intelligence, it also illustrates the flaws of AI. The expert system approach quickly falls apart when you want to factor in default rates by zip codes or other parameters. It is too difficult to represent many factors into algorithm. Machine learning can easily factor in the default rates by zip codes and thousands of other parameters, yet it is limited by the underlying data set then defines that algorithm. The data represents the past, has built in biases, and lacks the nuances of reason. If the person with a low credit score just won twenty million dollars in the lottery, it would recommend denying the applicant as it would probably treat lottery winners as flaws in the data because it has too few of them.
Gary Marcus
Expert Systems, Symbolic AI may be the greatest ooportunity for AI. The are almost 2 million healthcare studies produced each year. Yet this intelligence in burried in millions of silos of PDFs. There is a huge opportunity to code this into Symbolic AI.
unlies al or factorinand presented by if-then rules such as a person applying for a mortgage with is a computer system emulating the decision-making ability of a human expert.[1] Expert systems are designed to solve complex problems by reasoning through bodies of knowledge, represented mainly as if–then rules rather than through conventional procedural code.[2]
3. Where has AI been proven? - Selling us stuff
While many AI predictions continue to come up empty, AI has almost entirely taken over most of what we touch on our smart phones. AI on our smart phones learns your behaviors, what you like, what is important to you and what is going on in your life. This information is used to compete for your attention, keep you engaged, and make Google, Facebook, and Amazon trillion dollar companies. This explains why the top AI talent work for Google, Facebook and Amazon. By leveraging their AI capabilities, Amazon won the market place contest (50% of ecommerce), Google and Facebook won the attention battle (85% of ad revenue), and Google won search (90% of search).
While there is alot of AI hype around diagnosing disease, autonomous vehicles and winning games, most of the impact of AI is behind the scenes. This is how Jeffrey Bezos described this to his sharholders. "Much of what we do with machine learning happens beneath the surface. Machine learning drives our algorithms for demand forecasting, product search ranking, product and deals recommendations, merchandising placements, fraud detection, translations, and much more. Though less visible, much of the impact of machine learning will be of this type – quietly but meaningfully improving core operations."
Predictions of companies using AI
4. What are the primary roles of AI? Cognitive Assistance (Decisions) vs. Autonomy (Intelligent Automation)
5. Should it be called AI if Siri, Alexa, Google or Cortana helped?
Everything is AI now. Uses NLP, OCR, language translation, Siri,
If uncertain decision are made without human interaction.
If curation is well done without asking
6. AI doesn't make decisons, people do - Hybrid (Humans & AI)
AI doesn't make decisions. It can either assist a human in a decison or drive automation (a bot or triggers an action - w/d $100)
It also doesn't decide the best response. It is a hybrid (humans decide the parameters, AI decides based on the data)
Harvard study on hiring. While the study
Book: Weapons of Math Destruction
Biases
7. What can humans do better than AI? Context and Creativity AI just doesn't get it. Understand. Context - AI doesn't know what a car is. It doesn't have a mental model of it like a human. It also is limited to its data set. It doesn't know .... Creativity - 8. What can humans do that AI cannot? - Goals, Aspirations and Missions
Goals - What problems are we going to solve?
Aspirations -
Missions -
Deep Blue won Jeapordy, another missreport - doesn't understand context. It was just a massive hyperlink attached to search and NLP.
NEED Story of DARPA and self-driving cars - false predictions
9. What can AI do better than humans? Machine Learning Superpowers
Regression
Association
Correlation
Self-play
10. What Human-like skills is AI good at? Machine Learning Models
Use example of chatbots
Conversational UX, Machine Learning, NLP, Sentiment Analysis, Multilingual, Analytics & Administration, RPA, voice bots, cognitive abstraction
Vision, Sound Processing, Sound transmission, NLP, Robotic Physical Automation, RPA, Geospatial, Predicting, Pattern Recognition, Decision Making, Associations, Explainability
Mental health tools - showing success
11. How does AI learn from data? ML, DL & RL
12. What are the limitations of data? Passive vs. Active Data sets
Passive is collection of data (such as Facebook, Google, Amazon, Netflix) and active would be to ask questions to improve the data sets. An example is long covid sufferers who have 51 possible symptoms. An active data collection would ask the person which symptoms they have to improve the ML model.
13. Why AI is biased - humans and data
14. Why AI Requires Quality Data Sets - Large for Deep Learning
See blog, some public data sets (ImageNet) have 3% errors.
15. AI based on incomplete data will be misleading
Small sample sizes - coronary artery disease - high quality & high efficiency 10. Isn't AI going to pass human? Artificial General Intelligence vs Billions of Narrow AI
16. What is Intelligent Automation - Robotics (Physical) vs. Process (Digital)
AI doesn't do anything - it needs a human or bot or technology to take the action
17. Why are Autonomous Vehicles taking so long?
AI Powered Systems - autonomous vehicles, GANs (GPT-3, BERT), facial recognition, language translation, recommender systems, navigation systems, chatbots/virtual assistants,
18. Why do most AI project fail?
19. Headlines of AI success are often followed by disappointment
It's hard to replicate in every day practice
20. How will managers and staff leverage AI?
Citizen Developers
AutoML
Citizen Data Scientists
21. When will AI become mainstream (other than the overused buzzword) like the internet?
The progress of AI is beginning to shift from the concentrated AI talent working for tech giants and doing research to almost every sector of the economy. To understand how you and your company can ride this wave and avoid the undertow, the following are the "AI must knows" to leverage the collaboration of AI and humans.
22. Will AI drive Workforce Disruption and Reduction - Yes. It will also drive job creation.
23. Isn't AI going to pass human? Artificial General Intelligence vs Billions of Narrow AI
24. Should we be afraid of AI? - Elon Musk
25. What are the major downsides of AI?
26. AI Can't Accurately Predict an Uncertain Future - though it can help
27. Why AI is a Must in Managing Complexity - 65 complex challenges
28. Why AI Gives Us a chance to Keep Up With Complexity - complexity is growing
29. How AI could power Humankind's second remake - Assisted Strategic Reasoning
