1
Q
What was one of the first AI projects?
A
- Organised by John McCarthy, the Dartmouth workshop in 1956 studied 10 people for a duration of 2 months.
- It gave rise to the idea that every aspect of learning or any other feature of intelligence could be stimulated by a machine.
2
Q
Why was AI challenged in its early days?
A
It was believed that a machine would never develop the capabilities to reason logically, problem solve or play games - that these features were reserved for humans.
3
Q
Give an example of when the challenges associated with AI were overcome.
A
- Arthur Samuel created program which learned to play draughts (checkers) better than he could, disproving idea that computers can only do what they are told to do.
- Showed that computers could actually be intelligent.
4
Q
What was one the weaknesses associated with automatic translation in its initial development?
A
- Poor translations deriving from the automatic translators (e.g. in the Cold War - a time when US didn’t have many Russian speakers so drove need for translation) caused funding to dry up.
- These initial automatic translators gave illusion of intelligence but didn’t really know anything about the subject, they were just manipulating syntax.
5
Q
Give an example of an early chatbot.
A
- Eliza created in 1964-66.
- Early computer program
- Supposed to be a phycologist, offering intelligent advice and support.
- Not a reality, illusion of intelligence broken as picks up mere words and phrases.
- However, still considered intelligent as signified potential of computers.
6
Q
What does combinatorial explosion refer to?
A
- Need for immense computer power to effectively deal with large number of combinations.
- This is prevalent through the game of chess where there can be multiple combinations of different moves.
7
Q
What is a knowledge-based system?
A
- Emerged in 1969-1979.
- Generally a computer program that reasons and uses a knowledge base to solve complex problems.
8
Q
What are some problems associated with knowledge-based systems?
A
- Difficult for computer to gain the knowledge.
- The creation of multiple expert systems also creates problems as unable to share knowledge.
- No way of making a generally smart computer
9
Q
What is Artificial Intelligence?
A
The theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.
BUT is it really the goal to create a computer which acts as a human? More interest in creating something which can perform differently?
10
Q
What is one of the main reasons why AI is taken more seriously now?
A
- More data exists than ever before (trillions of words and billions of images on web) therefore, more data demands more computing power which demands better algorithms which demand more investment which all allows AI to function more effectively and successfully.
11
Q
Give an example which shows that AI has improved.
A
Google translate now provides much more accurate and reliable results that the automatic translators used during the Cold War.
12
Q
Give examples the emerging AI of today which challenges the morality and ethics of AI.
A
- Always been a debate about ethics and morality behind creating AI.
- Development of self-driving cars. (E.g. Google has developed Waymo by collecting data of driving around streets)
- How safe do they need to be?
- They actually highlight how most crashes are caused by human error. E.g. self-driving shuttle bus crashed on 1st day because the human wouldn’t stop even though the bus did.
- Disrupting the job market in numerous sectors such as military, pharmacy and farming.
- Research suggests that some AI programs exhibit racial and gender biases, reinforcing them.
13
Q
What are some other aims of AI?
A
- Developing a more global brain.
- Building a computer which can pass the Turing Test.
- Work cohesively with humans to produce a collective/combined intelligence.
14
Q
What does supervised learning refer to?
A
The machine learning task of inferring a function from labeled training data. (I.e. a teacher is required in supervised learning/human provides some data)
15
Q
What does unsupervised learning refer to?
A
- Unsupervised learning is a type of machine learning algorithm used to draw inferences from datasets consisting of input data without labeled responses. (I.e. not specified what data is more important/machine learns from patterns etc.)
- Can help us gain knowledge by organising data and examining the patterns that emerge.
16
Q
What does machine learning refer to?
A
- Another branch of AI.
- Allows software applications to become more accurate in predicting outcomes without being explicitly programmed.
- The basic premise of machine learning is to build algorithms that can receive input data and use statistical analysis to predict an output value within an acceptable range.
17
Q
What is generative design?
A
- Mimics nature’s evolutionary approach to design.
- Designers or engineers input design goals into generative design software, along with parameters such as materials, manufacturing methods, and cost constraints.
- Software explores all the possible permutations of a solution, quickly generating design alternatives.
- It tests and learns from each iteration what works and what doesn’t.
18
Q
What are some benefits of generative design?
A
- Brings designs that would otherwise never have been considered to light
- Provides many designs that all fit the criteria
- Creator and software working co-operatively
- Example of optimisation
19
Q
What is reinforcement learning?
A
- Branch of AI that allows machines & software agents to automatically determine the ideal behaviour within a specific context, in order to maximize its performance.
- Can produce novel (potentially surprising, worrying, disturbing) behaviour.
- E.g. Chatbot Tay released by Microsoft was taken offline 16 hours later for tweeting racist and sexually explicit messages. However, Tay was designed to learn from its users so internet trolls deliberately taught it offensive behaviour i.e. taking advantage.
20
Q
What are some significant examples of machine learning that show progress?
A
- Google DeepMind analyses huge amounts of data to detect eye diseases.
- LipNet learns better that humans how to read lips. (e.g. professional lip-readers can determine ca. 20-60% of what person is saying but LipNet can detect 94.4% with accuracy.)
- Recommender systems
21
Q
What are recommendation systems?
A
Nowadays, commonly use machine learning algorithms to provide users with product or service recommendations.
22
Q
Why is there a need for recommendation systems?
A
- Drives businesses promoting custom. (e.g. Amazon, YouTube, Spotify, Netflix)
- Aids in identifying trends and personalising recommendations and results, increasing relevance of info shown to user.
- Can help user find new things (novelty i.e. things you’re likely to be interested in but haven’t seen)
- Helps user find unexpected things (serendipity i.e. things wouldn’t normally be interested in but people with similar interests are i.e. from similar activity on items) and a wider scope of things (diversity)
23
Q
What kind of information does recommendation systems collect?
A
- Users likes and dislikes (web facilitates this by providing like/dislike buttons, rating and review features)
- Can then identify interests of a user e.g. if user interested in historical documentary, they will be more likely to be interested in another historical documentary.
- The more ratings a user leaves, the easier it is to make stronger predictions about the future behaviour of user.
24
Q
What does cohorts refer to?
A
- The collective buying or rating behaviour of various users can be utilised to create cohorts of similar users that are interested in similar products.
- These cohorts can then be used to make recommendations & find similar items.
- The more the users interact, the better the recommendations.
25
Give some everyday examples of recommendation systems in action.
- Many things driven by recommendation systems nowadays. (Amazon, Journals, YouTube etc.)
- Amazon was one of first retailers who saw the potential of technology and its usefulness. Thus, enabled users to rate products on 5 point scale to then be able to identify user purchase or browsing behaviour which provides strong indication of user interests.
- Google News utilises a unary system to recommend news to users based on their history of clicks. (i.e. when users can only indicate when they might like something, unable to dislike so unary = neutral or positive opinions only)
26
What are the primary methods of creating recommendation systems?
- Collaborative filtering (used the most)
- Content- based (improving as we understand content more e.g. image recognition)
- Knowledge-based
27
What is collaborative filtering in recommendation systems?
- Uses the collaborative power of the ratings provided by multiple users to make recommendations.
- It uses specific ratings and correlates them across various users and items highlighting that similar users give similar ratings.
- Can be memory-based (user or item i.e. distinguished by items not just users.) or model-based (decision trees, Bayes, latent factor)
28
What are some challenges involved with collaborative filtering?
The underlying specified ratings are sparse. (I.e. millions of items exist but very few ratings so hard to define similarity leaning to less accurate recommendations?)
29
What is a content-based recommendation system?
- The descriptive attributes (metadata or content) of items are used to make recommendations.
- E.g. consider situation where a user rated the movie The Martian highlight but no access to ratings of other users. This rules out collaborative filtering methods.
- Uses supervised learning or regressing modelling.
30
What are some challenges involved in content-based recommendation systems?
Difficult to make serendipitous, novel or diverse recommendations.
31
What is a knowledge-based recommendation system?
- Descriptive attributes of items are used to make recommendations.
- Useful in context of items that are not purchased often (e.g. houses, cars, insurance and mortgages are not bought frequently so unable to decipher user behaviour over time as not large number of push cases so cant learn through supervised learning or collaborative filtering)
- Based on similarities between customer requirements & item descriptions.
- Constraint-based (i.e. items with specific descriptions are retrieved e.g. databases of different houses with different metadata attached to it are clustered so when look at one you can see other similar properties)
32
What are the 5 minimal steps for collaborative filtering?
1. Collect preferences
2. Finding similar users
3. Ranking (& weighting users)
4. Recommending items (scoring items)
5. Finding similar products
33
How can similarity be measured in collaborative filtering?
- Euclidean distance
- Pearson correlation
- Cosine similarity
34
What is Pearson correlation?
- If number produced is positive then = similar
- If number is negative = similar in opposite way (i.e. one user really likes item another really doesn't/high score & low score)
- If number is 0 = no correlation.
35
Describe the taxonomies of artificial intelligence.
Taxonomy of AI definitions:
- Think humanly and rationally
- Act humans and rationally
36
What is the definition of autonomous and how is this relevant to AI?
- Definition: having the freedom to act independently
| - We increasingly want robots to become autonomous, able to think and act independently.
37
What are some tasks that robots aim to do?
- Navigate
- Search and retrieve
- Rescue
- Care/companionship
- Clean
- Scientific exploration
38
Which factors have to be taken into consideration when building a robot?
- How it will receive signals (e.g. through camera/sensor/microphone), process signals, signal to us (e.g. communication through movement/speech).
- Movement & manipulation (e.g. opening doors, grabbing cups)
- Energy source (e.g. power charger)
- Body
39
How can robots be built/constraints of building robots?
- Mobility can be fixed (e.g. robots that manufacture cars) or mobile (e.g. mars rover)
- Autonomy can be tele-operated or completely autonomous.
- Shape can vary = anthropomorphic (having human characteristics), zoomorphic (having or representing animal form) or mechanoid (designed to look and act like a human).
- Interactivity can be isolated or interactive.
- Learning can be fixed or adaptive.
- Application can be industrial (e.g. in lab with chemicals) or service related (e.g. cleaning)
- Environment can be controlled or unpredictable (need for robot to be more adaptable if in unpredictable environment).
40
What are some examples of different robots?
- Self-sustaining = EcoBot (finds energy source through eating pollution) & SlugBot (eats flies or slugs).
- Swarms = Swarm-bot, Swarmanoid (refer to smaller robots which work together, commonly used in rescues)
41
Give more details about the EcoBot.
- Project by Ioannis Irepoulos, John Greenman and Chris Melhuish.
- Utilises an artificial digestive system to self-sustain.
- The first EcoBot sustained itself for 2 weeks on 8 dead flies.
42
What are some benefits of humanoid robots?
- Can hold human tools and share workspace.
- Communicate naturally with humans.
- Shape what makes 'human-like' not anatomy.
- Companions
Benefits can also be considered as cons.
43
Give examples of humanoid robots.
- Kismet (made in the late 1990s at Massachusetts Institute of Technology by Dr. Cynthia Breazeal - a machine that can recognize and simulate emotions.)
- Cronos (developed by Owen Holland and colleagues, able to be pushed and its body will flex in response - moving away from grid robot design, moves in complex way, designed to explore machine awareness, works out how to move from its own computer simulation of itself within itself)
44
Give some benefits of swarm robots.
- Control is decentralised and distributed to the robots = autonomy.
- Smaller so can get into difficult spaces/areas, local sensing and communications.
- Strong response against failure (if one goes down, the other carry on so no central breakdown)
- Spatially distributed work (can be used for industrial clean up or agricultural needs)
45
Give some negatives of swarm robots.
- Difficult to design to achieve collective task.
| - Autonomy makes difficult to design human interface.
46
Give an example of when Swarm-bots could be used.
- Search & rescue situations.
- Need more than one & must cooperate and coordinate.
- E.g. 1 or 2 bots find child, determined child too heavy to drag to safety, signal for other bots to join, only room for 4 bots to grab child, child too heavy for 4 bots, bots form 4 chains of robots until there are enough bots to drag child to safety.
47
Give a specific example of a swarm-bot.
- ALICE
- Created by Markus Weibel, Dario Floreano & Laurent Keller.
- Sugar-cube sized robots evolves to be cooperative and altruistic (unselfish).
- Constructed upon an Artificial Neuron Network controller with multiple inputs and single output.
- Not simulated, real robot test for fitness.
48
Give an example of a hybrid robot.
- Hybrid robot combines deliberative and reactive construction.
- Symbrion by Paul Levi Serge Kernbach
- Takes swarm idea but each part has different functions but can also come together to perform functions so more flexible. (alternate between swarm mode and organism mode)
- Self-assembly into 3D organisms.
49
What is the Turing Test?
The Turing test, developed by Alan Turing in 1950, is a test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.
