Artificial Intelligence Flashcards
What is the Turing Test?
behaviour-based
- computer must be able to convince person in another room that it can answer questions like a human would
- should convince people that any computer that could pass this test fair-and-square would be intelligent
DOES NOT care whether machine is intelligent or not
DOES care whether machine acts/behaves like it’s intelligent
What is the Searle Chinese Room Test?
internal-state
- computer that can translate languages is NOT intelligent because it doesn’t attach meaning to words
- algorithm being followed ≠ thinking
What is artificial general intelligence (AGI)?
understand or learn any intellectual task that a human being can
What is narrow AI?
don’t perform outside of the single task that they are designed to perform
What is natural language processing (NLP)?
automatic processing of human language (ie. by computers)
What is the traditional NLP approach?
long list of rules for processing language, formulated by people and programmed into computers
What is the modern NLP approach?
machines learn from text examples using artificial neural networks and similar approaches (like babies)
- statistical methods allows to compare different interpretations
What are the 3 advantages of modern machine-learning NLP methods vs. traditional hand-coded rules?
- relevance: learning automatically from bodies of text means that common cases are trained more carefully
- robust: statistical algorithms can make guesses about unfamiliar and erroneous input (ie. misspelled words)
- training: simply provide more text to further train modern algorithms
What are the traditional NLP steps?
- Recognize speech (Watson skipped this – it received ASCII versions of questions)
- Syntax analysis, or parsing
- Semantic analysis
- Pragmatics
What is parsing (syntax analysis)?
inferring parts of speech and sentence structure, using lexicon and grammar
takes sentence and shows how words are assigned parts of speech and build up to form a sentence
What is semantic analysis?
inferring meaning using syntax and semantic rules
What are pragmatics?
inferring meaning from contextual information
How is parsing the basis for computer programming?
- computer has to ‘understand’ programs in order to execute them
- programming languages are designed so that they can be parsed unambiguously
- grammar specifies all possible programs that can be written in a language
- designing programming languages (and their grammars) is fun and important part of CS
Why is parsing hard?
can be tricky – natural languages are ambiguous
Which type of NLP has trouble with ambiguities?
traditional, rule-based NLP
Syntax vs. Semantics
syntax:
- knowing if word is noun or verb (syntax) tells us something about its meaning (semantics)
- describes sentence’s structure
semantic analysis:
- provide additional information – word categories, relationship between words
- adds limited meaning that can be figured out using simple rules that don’t require much context
Pragmatics
What do most techniques do to find semantic meaning of words?
look for clues in surrounding text to disambiguate word meaning
When is pragmatics especially important?
when sentences contain pronouns
What is the traditional method to learning to translate between languages?
- Apply NLP to each sentence in each language
2. Follow a set of rules that define how to translate
What is the newer method to learning to translate between languages?
use machine learning techniques plus data on the web
What is machine learning?
process of using data to help systems ‘learn’ without being explicitly programmed how to do task
What are 3 limitations of traditional NLP?
- natural language is structurally ambiguous, so parsing alone cannot lead to understanding
- synonyms for words can’t be used interchangeably in every context
- natural languages have many exceptions to grammatical rules – no agreed-upon grammar for all uses of a language
What is the brain?
mass of highly interconnected neurons
- dendrites receive inputs
- axons carry output
- output is based on input
What neuron structure gives the brain many adaptive advantages?
many various connections among individual neurons is more important than individual neurons themselves
What are artificial neural networks (ANNs)?
many interconnected artificial neurons – outputs of some feed into others
What can simple ANNs carry out?
What can simple ANNs carry out?
basic logic functions
How can ANNs learn?
- with life experience, depending on particular person’s trials and errors, synaptic connections among pairs of neurons get stronger or weaker
- ANN could do something similar, by gradually altering (on guided trial-and-error basis) numerical relationships among artificial neurons
- wouldn’t need to be pre-programmed with fixed rules – would instead rewire itself to reflect patterns in data it absorbed
Can ANNs learn a specific operation?
yes
How do ANNs work?
neurons in ANN can solve one tiny function, then pass the result on to another neuron
What is deep learning?
using deep neural networks
neurons in ANN solve one tiny function, and pass the result on to another neuron
contains many layers of neurons – if one neuron makes mistake, it does not have profound outcome on overall effect
What are word embeddings?
multi-dimensional maps of language
words correspond to points in multi-dimensional space
- words that are “similar” should be close in this space
What are word embeddings created by?
training neural networks
What are word embeddings the key to?
new and improved translation software
How does word embedding work?
- need huge volume of text
- train neural network to associate nearby words in texts
- enter word into input layer, and network should produce nearby word in output layer
- if it can do this with small hidden layer, it is somehow embedding meaning of these words
What are some ideas to improve language translation?
- train artificial neural networks to recognize patterns across languages
- use word embeddings as input to networks
results almost as good as human translators
What are 3 types of machine learning?
- supervised
- unsupervised
- reinforcement
What is supervised machine learning?
data is labelled to instruct machine which patterns to look for
What is unsupervised machine learning?
data is unlabelled, machine looks for any patterns
What is reinforcement machine learning?
learn by trial and error to achieve a clear objective
What is machine learning?
looking for patterns in massive amounts of data
What are adversarial attacks?
inputs to machine learning models that attacker has intentionally designed to cause model to make mistake – like optical illusions for machines