Revise Artificial Intelligence: here is a selection of questions of every level, with the answer and an explanation. To test yourself without seeing the answers, play the quizzes.
Play the Artificial Intelligence quizzesAnswer : Techniques that enable a machine to perform tasks requiring intelligence
AI groups together techniques that enable a machine to perform tasks that normally require intelligence, such as understanding a sentence or recognizing a face.
Answer : Siri
Siri, Apple's voice assistant, understands a spoken request and responds to it. Excel, Photoshop and Git are not voice assistants.
Answer : A program that converses with a user in natural language
A chatbot is a program that converses with a person in natural language, in writing or by voice. The other options describe other computing objects.
Answer : 1950
In his paper “Computing Machinery and Intelligence”, published in 1950, Alan Turing proposed the imitation game, which became the Turing test: a judge must tell a human from a machine based on written conversations.
Answer : John McCarthy
John McCarthy, one of the organizers of the 1956 Dartmouth conference, is credited with proposing the term “artificial intelligence”.
Answer : 1997
Deep Blue, developed by IBM, won the match against Garry Kasparov, the reigning world champion, in May 1997, by 3.5 to 2.5.
Answer : Jeopardy
Watson, developed by IBM, beat two former human champions of the quiz show Jeopardy, which is broadcast in the United States, in 2011.
Answer : Go
AlphaGo won 4 games to 1 against Lee Sedol in March 2016, at the game of Go.
Answer : A label, that is, the expected answer
In supervised learning, each example is associated with a label, for instance “cat” or “dog” for a photo, which the model learns to predict.
Answer : Finding clusters or hidden structure in the data, without labels
Without labels, the model looks for groups of similar examples by itself, for instance customers with similar buying habits.
Answer : A reward obtained after each of its actions
The agent learns by trial and error: it seeks to maximize the cumulative reward that its actions earn it in an environment.
Answer : Regression
A regression predicts a continuous numerical value, such as a price or a temperature, from examples whose value is known.
Answer : Classification
Classification assigns one category among several possible ones, here “spam” or “not spam”, based on examples that are already labeled.
Answer : To evaluate the model on data it has never seen during training
A model can seem very good on the data it has learned from. Only a test set, kept aside, can measure its performance on new cases.
Answer : A model that sticks too closely to its training data and generalizes poorly to new data
Overfitting occurs when the model memorizes the particularities of its training set, noise included, instead of finding general patterns.
Answer : The artificial neuron
An artificial neural network is made of neurons connected to each other and organized in layers. Each neuron computes a weighted sum of its inputs.
Answer : The instruction or question text sent to the model to get an answer
The prompt is the input the user gives the model: a question, an instruction or an example. The quality of the answer depends a great deal on how it is worded.
Answer : A unit of text, often a word, part of a word or a punctuation mark
Text is split into tokens before being processed by the model. A token can be a whole word, a fragment of a word or a punctuation mark.
Answer : The maximum amount of text, in tokens, that the model can process in a single request
The context window limits how much the model can take into account at once: beyond it, the oldest passages are no longer visible to it.
Answer : A wrong or invented answer, presented confidently as if it were accurate
An LLM produces plausible text without checking the facts. It can thus invent a source, a date or a quotation, which is why it is worth checking its answers.
Answer : Llama
Llama is a family of large language models developed by Meta. Gemini is designed by Google, Claude by Anthropic and GPT by OpenAI.
Answer : Attention Is All You Need
Published in 2017, the paper “Attention Is All You Need” introduced the Transformer, an architecture based on the attention mechanism, which underpins most large language models.
Answer : Continuing its training on a targeted dataset to adapt it to a task
Fine-tuning specializes a model that is already trained by adjusting its weights on examples specific to a domain or a response style.
Answer : A numerical representation, in the form of a vector, that captures the meaning of content
Texts with similar meanings have similar embeddings, which makes it possible to search documents by similarity rather than by identical words.
Answer : AI winters
AI winters are periods, notably in the 1970s and the late 1980s, when unfulfilled promises caused funding to collapse.
Answer : Frank Rosenblatt
Frank Rosenblatt introduced the perceptron in 1958, a model with a single layer of neurons that learns to classify inputs by adjusting its weights.
Answer : AlexNet
AlexNet, trained on graphics processors, sharply reduced the error rate of the 2012 ImageNet competition: it was a turning point for deep learning.
Answer : Geoffrey Hinton, Yann LeCun and Yoshua Bengio
In 2018, the ACM awarded the Turing Award, often described as the Nobel Prize of computing, to these three pioneers of deep neural networks.
Answer : The ability to explain to a human the reasons for a decision produced by the model
Explainability (XAI) aims to make a model's decisions understandable, which is necessary when a decision affects a person, such as a credit refusal.
Answer : Modifying the weights in small steps to reduce the loss function
The model adjusts its weights in the direction that reduces the error measured by the loss function, repeating this step over many examples.
Answer : Gradient backpropagation
Backpropagation applies the chain rule for composite functions layer by layer, from output to input, to provide these gradients to gradient descent.
Answer : Convolutional neural networks (CNNs)
The filters detect local patterns, such as edges or textures, then combine them layer after layer. This approach is at the heart of computer vision.
Play the Artificial Intelligence quizzes