TASK 1 - ARTIFICIAL INTELLIGENCE Flashcards
Artificial Intelligence
= system that displays intelligent behaviour by analysing environment + taking actions
= set of algorithms and techniques that try to mimic human intelligence
- generates predictions
goals for AI
- ability to reason: games, surgery
- knowledge representation: language, objects translated into programming
- planning: navigate how to get from A to B
- natural language processing: understand language and its context
- perception: how do we see, hear, feel… things
- general intelligence: emotional intelligence, creativity, intuition
= autonomous thinking robot, almost indistinguishable from human
types of AI
- rational AI: do not adapt behaviour over time
- learning rational AI: evaluates actions to adapt reasoning rules + decision-making methods
- general/strong AI system: can perform most activities that humans can do
- narrow/weak AI system: can perform one or few specific tasks
- black-box AI: accurate but can’t trace back the reason for certain decisions
learning mechanisms of AI
- machine learning
= computers programmed to learn from past experience and example data
- search and optimisation, constraint satisfaction, logical reasoning, probabilistic reasoning, control theory
machine learning
- supervised learning
= provide it with examples of INPUT-OUTPUT BEHAVIOUR –> generalise from examples + behave well also in non-similar situations (prediction)
- instead of giving behavioural rules to the system (machine learning)
- parallels concept learning in human and animal psychology
- exemplify function approximation: infers function from labeled training data consisting of a set of training examples; training data take the form of a collection of (x, y) pairs
- -> goal is to produce a prediction y* in response to a query x*
- form predictions via learned mapping f(x) which produces an output y for each input x
machine learning
- unsupervised learning
= type of SELF-ORGANISED Hebbian learning that helps find previously unknown patterns in data set WITHOUT pre-existing labels
- allows the modelling of probability densities of given inputs
machine learning
- reinforcement learning
= concerned with how software agents OUGHT TO take actions in an environment in order to maximise some notion of cumulative reward
- intermediate between supervised and unsupervised learning
- finding a balance between exploration (of uncharted territory) and exploitation (of current knowledge)
- does not need labelled input and output pairs to be presented
- does not need sub-optimal actions to be explicitly corrected
learning mechanisms of AI
- deep learning
= data-up
- feed data structure modelled on the human brain a bunch of data
- algorithms help computer learn based on that data
- neural network: mimic the organisation of the human brain (neurones connected to neurones) with algorithms and data
- multiple layers inside these networks + small processing units with heaps of weighted connections among them
predictions of AI
- types
- prediction = ability to take info you have and generate info you didn’t previously have
≠ automation
1. timelines + outcome predictions = when AI milestone will be achieved
2. scenarios = if conditions met, certain outcome will happen
3. plans = if certain plan is implemented, then certain goal will be achieved
4. issues + meta statements
predictions of AI
- methods
- causal models = making conclusion based on fact concerning the ultimate outcome
- non-causal models = if you don’t know what influences what, you can only make hypotheses about the future
- outside view = gather examples and claim they follow a trend
- philosophical arguments = pin-point problems that should be solved to get to AI
- expert authority = rely on expertise
- non-expert authority = rely on non-experts; no reason to believe them
predictions of AI
- assessment methods of predictions
- extracting verifiable predictions: deriving empirical predictions from arguments that have been made (testing model from inside)
x method increases uncertainty as it narrows the consequences of the prediction (testing model from inside) - clarifying and revealing assumptions: make the prediction as thorough as possible; assess the assumptions and the logical structure behind the argument
- enthymematic gaps (hidden assumptions): should be revealed as they clarify where true disagreements lie and where we need to focus investigation in order to find out truth of prediction - model testing and counterfactual resiliency: testing the strength of model from the outside; imagining the world history had happened slightly differently and checking whether the model would have stood up in those circumstances
- find nodes of disagreement, illustrate tension between the given model and other models of history –> not to rule out certain models
predictions of AI
- mistakes of AI predictions
1) overestimate the effect in the short tun and underestimate the effect in the long run (AI winters)
2) if technology is far enough from technology nowadays, we do not know its limitations (MAGIC) –> if it becomes magical, anything one says about it is no longer falsifiable (POWERFUL)
3) we overestimate competence of its predictions
4) AI still needs a lot of human input/guidance
5) exponential growth can collapse when physical limit is hit + when there is no more economic rational/incentive
6) there won’t be sudden developments but we will evolve with the technologies
7) innovations will take longer to develop, as infrastructure needs to be advanced first
progress of AI
- historical level: natural language processing (NLP) = understanding and modelling human language
- contemporary level:
- text-based agent = CBT techniques in conversation-like interactions
- virtual reality - near future: AGI = reach many goals and complete tasks in almost superior way to humans
- far/unknown future: super intelligence = high-level AI that far surpasses human intelligence
- singularity
applications of AI
- counselling: as a supplement may benefit clients
- clinical treatment + training: adaptive training; more willing to seek help
- clinical decision making: fuzzy expert systems
- managers: judgment will become most valuable workforce skill
progress of AI
- counselling profession
- counselling = forming a professional relationship + empowerment + accomplishment of goals
1. historical: does not fulfil any goal
2. contemporary: possibly helps accomplish goals
3. AGI: fulfils empowerment + goal achievement, professional relationship raises ethical questions
4. superintelligence: fulfils all requirements
