Exam - AI Flashcards
narrow vs general AI
narrow - specialized AI for a specific task/fxn
general - good/useful for wide range of roles
old-school applications of AI
logic interface
provide facts/info/data to the AI –> AI returns logical conclusion
used for
planning
optimization
finding solutions within constraints
playing games
how does AI store concepts/objects?
stored as high dimensional vector series
mathematically store values/properties of obj/concept
series of vectors influencing eachother –> forms neural net
3 types of AI learning + examples of each
supervised - give data to AI, tell it what the results should be. test using unknown inputs and see what the AI outputs (eg speech recognition
unsupervised - give data to AI, classifies data into categories based on shared traits –> data clustering + pattern recognition (eg anomoly detection, data mining)
reinforcement learning - data –> results –> modify algo based on results to improve. feedback loop learning to incrementally improve results (eg live route planning)
training AI against themselves
symmetric AI - AI competes against the exact same algo. A level of randomness is intentionally introduced as to prevent exact same behaviours
asymm AI - training AI against intentional opposites –> adversarial AI
As A beats B, B gets better. This causes a feedback loop where both AIs improve over time
how does image generation work
Nutshell: train the AI to denoise pure noise image into generating final image
start with pristine image –> train AI to recognize image
gradually add more noise to the image –> train AI to recognize + denoise –> denoise algo requires image restoration
eventually get to pure noise –> tell AI “Image is supposed to be X, denoise it” –> denoise algo used to “re”generate image from scratch
how does natural language processing (NLP) work
train AI to store words + concepts as indiv vector
diff combination of words in a sentence –> influences stored vector data of other words –> allows AI to generate more natural sentences that make sense –> AI can adapt predictive word generation algo based on existing sentence structure/context
each added word changes the vector values for other words –> generative (pre-trained) transformation (GPT)
pre-trained AI that generates text using transformative algorithm to allow NATURAL language processing
How does AI get better and better? whats teh difference
recall: concepts/words are stored as high dimensional vectors –> add even more dimensions to store even more data associated with the concept
therefore the AI has a more detailed grasp of the concept + how to use it
in chatGPT –> more higher dimensional vectors = better semantics, interpretation, finer nuanced meanings, etc
