Artificial Intelligence Flashcards
What is AI?
A machine that uses any kind of algorithm to perform perceptual, cognitive, or conversational functions typical of the human mind, such as visual and speech recognition, learning, reasoning, and problem solving
What are the 3 types of AI?
- Artificual Narrow Intelligence: performs specific intelligent tasks but not others.
ex: Playing chess, driving a car, but cant do both - Artificial General Intelligence: hypothetical, human-level intelligence in every sense.
- Artificial Super Intelligence: hypothetical, super-human intelligence.
What are the 4 main subfields of AI?
Machine learning = Algorithms that can find patterns in data and then use those patterns to make predictions (most useful in biz)
Natural language processing (NLP) = Recognizing, understanding, producing human languages (ex: Alexa, Siri) ** use machine learning to get better
Computer vision = sees the world with cameras sensors etc. (uses ML to break down into components, get betterw and understand wtf goin on)
Robotics = physically embodied AI (manipulators like factory arms, mobile like drones/cars, humanoid like those creepy ones)
How does machine learning work? WHat are hte main steps? Example?
Human programmers dont tell it instructions for a certain task, instead it gets trained to an algorithm so it learns how. Uses crap ton of data to train it bc need to learn patterns
1) training
2) prediction
Example: give thousands of photos and humans tag them as cat or not cat, then algorithm knows which ones mean cat and picks up patterns that apply to all cats, then learns what makes a cat and can pick out cats from nontagged photos
What are some other regular examples of machine learning?
- Email spam filters (starts with a standard then learns what you like/dont like)
- Recommendation systems (YouTube, Spotify, etc.)
- Targeted advertising algorithms (Facebook etc.)
- Voice recognition / NLP
- Image recognition / computer vision (i.e. driverless cars, cancer diagnosis –> see already confirmed skin cancers, then learns correlations, then predicts)
What is deep learning?
More complex form of machine learning; multiple “layers” - separates data into stages
More complex so usually more accurate, but also harder to interpret for humans.
What is the ‘black box’ with AI? Issues?
Not always clear why AI makes a certain decision/prediction (bc many steps from input to output, so path hard to follow bc complex af)
Issues:
- scientists/programmers cant explain how it got there bc crap ton of deep learning involved, so cant see if its doing the right thing/regulate/etc. (Ex: could be big issue if used in criminal justice or somethin bc could be doing it wrong but we wouldnt know
- people dont want to trust it if they dont know the reasining/why/how behind the AI choices
- black box can make it easy for companies to hide their algorithms and therefore can be used for nefarious uses/personal gain (ex: predict commiting crime again after release (recidivism), but company has financial incentive to have more people coming in bc being paid to use the AI for these cases)
Tradeoff between accuracy and interpretability?
Not always! sometimes the simple ones can also be very accurate
What are some machine bias examples?
- AI that crops photos for twitter newsfeed automatically selected white face
- AI for video call background did not pick out black persons face and instead just made them a shirt in front of a cool background
- AI predicting recidivism labelled African americans higher risk more often, even when adjusting for ones that actually did reoffend (more whites assigned low risk but did reoffend, and more afamer assigned high but didnt)
- AI designed to determine if patients get extra medical attention by assigning risk scores, was disproportionaltely assigning lower scores for same level of health conditions as white person :. less likely to get extra help when needed
- how? algorithm based partially off of healthcare costs, but afr amer use healhtcare less probably because of distrust, so less overall but still just as unhealthy etc.
What are the impacts of automation and robots on unemployment?
- will take lots of the menial repetitive jobs (bc can be written into a formula)
- soft skill based work probably okay (creativity, emotion, social skills)
- lots will be partially automated so potential decrease overall demand for humans, but also help shift ones that stay into the more interesting parts of the job
What has been the actual evidence of AI being used to replace jobs?
- more hiring people who know how to work with and use AI, less hiring people who have same skills as the AI
- those who have the same skill as AI receiving lower wages
What is algorithm aversion? What are some examples and impacts of it?
= Tendency to prefer relying on humans than on algorithms for recommendations and advice
- Doctors who get advice from an algorithm (but not another human doctor) seem less capable and professional
- Managers who hire people using an algorithm are seen as less fair and effective (but algo is more fair bc recruiters BIASED)
- People prefer to get medical advice from a human because they think AI ignores their unique characteristics
Is algorithm aversion ever less significant?
Yes, when it is an objective/quantitative task (ex: analyzing data, predicting weather, etc.)
VS people more averse to AI for subjective tasks (ex: hiring employees, composing a song, dating recommendations, etc.)
How can one overcome algorithm aversion?
- different language and framing
- –> frame something as more utilitarian (more objective) will make people more cool with the AI recommendation
- allow people slightly change the algorithms output = feel better about using it because have a tiny bit of control
What is the law of accelerating returns? What is the singularity?
human progress is growing at an exponential pace, because the more advanced we are, the faster we can progress = even more advanced = even faster rate ….
Singularity = the asymptote on the curve –> the point where progress is happening at an infinite pace
