What AI is and trends around AI11%（当サイト設定） · 6 topics

Q1. Consider the four-level classification of AI. A mechanism that uses human-designed features (such as word counts) and learns spam-detection rules from data falls into which level?

1. AI incorporating deep learning
2. AI incorporating machine learning
3. Simple control program
4. Classical AI

Q2. Consider the four-level classification of AI. A mechanism that also learns from data the very features useful for judging images, and tells apart objects in photos, falls into which level?

1. Classical AI
2. AI incorporating machine learning
3. AI incorporating deep learning
4. Simple control program

Q1. Consider issues discussed in AI debates. A thought experiment says that even if a person who doesn't know Chinese rearranges symbols per a rulebook and replies perfectly, they cannot be said to understand the meaning. What is it called?

1. Frame problem
2. Symbol grounding problem
3. Knowledge acquisition bottleneck
4. Chinese room

Q2. Consider issues discussed in AI debates. Interviewing experts and turning their knowledge into rules took great effort, and much of it was tacit, so expert systems did not spread. What is this problem called?

1. Frame problem
2. Knowledge acquisition bottleneck
3. Symbol grounding problem
4. Chinese room

Q1. You are choosing a search method. You use the Mini-Max method on a game tree and want to cut branches that need not be evaluated to reduce computation. Which method fits best?

1. Depth-first search
2. Alpha-beta pruning
3. Monte Carlo method
4. Breadth-first search

Q2. Which best describes IBM's Watson?

1. AI from a Japanese research project developed with the goal of passing Japanese university entrance exams
2. A question-answering system that beat human champions on a US quiz show
3. A system combining deep learning and reinforcement learning that beat a professional Go player
4. A system that beat the world chess champion by searching with dedicated hardware

Q1. Which early automated planning method describes actions as preconditions, actions and effects, and finds an action sequence from an initial state to a goal?

1. STRIPS
2. Cyc project
3. MYCIN
4. SHRDLU

Q2. In ontology research, which best describes the idea of a “lightweight ontology”?

1. An idea that puts top priority on reducing the amount of ontology knowledge to save storage capacity
2. An idea that it is best to build prediction models only from data statistics, using no knowledge at all
3. An idea of automatically finding relationships between concepts through data analysis without insisting on exactness
4. An idea of having people strictly define relationships between concepts through philosophical reflection to build a correct knowledge system

Q1. Which best describes the background to machine learning drawing attention in the third AI boom?

1. Because theory proved machine learning always gives correct answers even with little data
2. Because collecting experts' knowledge as rules could now be done automatically
3. Because the spread of the internet made big data available and computer performance also improved
4. Because computing every branch of a search tree became possible in a short time for any problem

Q2. Which best describes the “curse of dimensionality”?

1. As feature dimensions increase, the data needed for training grows sharply, making performance hard to improve
2. As layers deepen, gradients shrink from the output side and layers near the input stop learning
3. Fitting training data too closely so prediction accuracy on unseen data drops
4. Learning a new task causes knowledge of previously learned tasks to be lost

Q1. Which best describes the difference between statistical and rule-based machine translation?

1. Statistical MT translates using grammar rules written by people; rule-based learns translations from large parallel data
2. Statistical MT is a kind of deep learning, and rule-based is a newer improved version of it
3. Statistical MT chooses translations probabilistically from parallel data; rule-based translates with rules written by people
4. Statistical MT is only for speech translation, and rule-based is only for text translation

Q2. Which statement about debates around AI is NOT appropriate?

1. The term “deep learning” was first used at the Dartmouth workshop (1956)
2. Embodiment debates hold that interaction with the outside world through a body is important for intelligence
3. The Loebner Prize is a competition for conversation programs based on the Turing test
4. The term “artificial intelligence” was first used at the Dartmouth workshop