PHI3681 Midterm Vocabulary

Turing Test

• interrogator questions 2 witnesses and analyzes their written responses to determine who the bot and human are

• bot is considered intelligent/correct if the interrogator guesses wrong 70% of the time

Turing Machine

mathematical model of possible machine using rules to make output for a specific input

expert systems

• computer emulates human’s decision-making skills

• ex. diagnosis AI

heuristic search

• finding the most efficient path from start to end

• ex. game playing

machine learning

• use algorithms and statistical models to make inferences from patterns in data

• ex. decision trees

backpropagation learning

• training method with forward pass to calculate output and backward pass using weights to calculate gradient

• ex. image or text processing with DNNs or NLPs

overfitting

when the model performs well on the training set but poorly on new data

underfitting

when the model performs poorly on both the training set and new data

deep convolutional neural network (DCNN)

• input, convolutional, pooling, fully connected, and output layers to determine important features for classification

• ex. object recognition

generative adversarial network (GAN)

• generator makes data to test and fool the discriminator

• the discriminator tries to discern real and fake data until generator produces highly realistic data

• ex. realistic image generation

transformer

• has self-attention mechanism, positional encoding, and encoder-decoder architecture

• ex. translation

nuisance parameter

• unknown constraint not of primary interest but needed to complete the distribution

• ex. translation

abstraction-as-subtraction

look at examples and remove their differences to find their abstract properties

abstraction-as-composition

• the process of building a more abstract idea out of slightly less abstract ideas

• ex. constructing the idea of a triangle out of the ideas of angle, side, and 3

abstraction-as-representation

• the process where people use specific exemplars of a category to stand in for the rest of the category members

• ex. mammal abstracted as a dog 

adversarial example

• inputs meant to skew a model’s predictions to inaccuracy

• ex. poisoning attack - mislabeling an image

.vector-space semantics

• representation of words and sentences as vectors in multidimensional space

• related meanings grouped together in an NLP

Blockhead thought experiment

• program with finite number of every possible correct sentence could pass the Turing Test but isn’t consciously understanding what it’s saying. 

• ex. ChatGPT as therapist?

anthropomorphism

• the tendency of humans to attribute human-like characteristics on the basis of superficial or irrelevant similarities

• ex. ChatGPT as an attractive man because it sugarcoats its points and is agreeable

anthropofabulation

• anthropocentrism + confabulation

• criteria for intelligence based on an exaggerated degree of humans possessing that trait

• ex. thinking a Roomba is “dramatic” for being stuck in a certain spot

activity maximization

• XAI technique to optimize input to activate certain neuron and output

• ex. DCNN neuron activation to recognize parts of a dog’s face

salience mapping

• highlighting most relevant regions for machine learning models

• ex. recognizing ears of a dog vs. ears of a human

counterfactual explanation

• motivated by causal justice and discrimination

• what would have happened if the bias didn’t exist

• ex. what if ChatGPT didn’t differentiate qualifications based on ethnicity and gender? would be representative of equality in hiring candidates for a job

Rashomon set

• set of reasonably accurate predictive models

• ex. blood sugar, age, BMI, and exercise levels determining risk of diabetes

algorithmic bias (statistical)

• systematic deviation output, performance, or impact, relative to some norm or standard

• ex. regression model based on skewed data underpredicts housing prices in certain areas. 

algorithmic bias (ethical)

• AI unfairly discriminates against certain groups of people

• ex. COMPAS algorithm reporting that recidivism rates are higher among black people, ATS hiring more white men over anyone else

standard debiasing strategy

• mathematical quantification of bias to reduce and eventually eliminate it