Symbolic vs Subsymbolic AI

Question 1

What is symbolic AI?

Answer 1

Good old fashioned AI, knowledge and reasoning are seperate, knowledge represented as explicit symbols, reasoning through application of predefined rules and logic, symbols manipulated by rules to solve problems

Question 2

What is the symbolic grounding problem?

Answer 2

Problem within symbolic AI wherein all knowledge is given not acquired - means representations aren’t grounded to nay real world knowledge/system doesn’t understand what things are, and means that for these systems to get better we have to keep giving them more knowledge

Question 3

What is the Chinese room?

Answer 3

Thought experiment demonstrating the symbolic grounding problem - a guy could sit in a room and be able to “speak” chinese chinese by following step by step rules to produce output symbols from input symbols without understanding them

Question 4

What are some examples of symbolic AI?

Answer 4

SHRDLU AI, General problem solver, semantic networks, deep blue

Question 5

What is SHRDLU?

Answer 5

symbolic AI system that takes inputs of size, shape and position of blocks in a tabletop world and instructions on how to interact with the blocks, able to follow and combine instructions in novel ways through conversation with user

Question 6

What is the General problem solver?

Answer 6

symbolic AI system able to solve problems through means-end reasoning, can solve restricted search problems such as Tower of Hanoi

Question 7

What are semantic networks?

Answer 7

symbolic AI systems that are knowledge structures that efficiently represent knowledge about the world, can be used to make inferences about novel examplars e.g., if graph specifies that canary is a bird and birds have wings can infer that canaries have wings or robins have wings

Question 8

What is Deep Blue?

Answer 8

symbolic AI system using supercomputer (greater computational power than symbolic systems before), able to beat world chess champion by searching through million of moves per second (brute force)

Question 9

Explain how symbolic AI works as if it is in a game such as chess?

Answer 9

Symbolic AI tries to concieve world as set of game states and actions able to take as moves in the game.
knows current game state and goal state, know operations/ possible moves and how these change game state

Question 10

What are the strengths of symbolic AI?

Answer 10

Transparent, explicit, expert systems, well-suited for logic-based problems

Question 11

What are the weaknesses of symbolic AI?

Answer 11

narrow not general intelligence, not suited for ambiguity and certainty of real world, struggle with high dimensional data, cannot learn themselves, not good model of biological brains

Question 12

What is Moravec’s paradox?

Answer 12

idea that everything with AI is “upside down” - it finds things humans find difficult stuff easy but things humans do without thinking really hard, only really applies to symbolic AI as sub-symbolic AI means don’t need to fully understand something before computer can replicate it

Question 13

What is sub-symbolic/connectionist AI?

Answer 13

learn from scratch, work on level “below” explicit representations and instead learn patterns by processesing data through interconnected nodes/neurons typically organized in layers, knowledge and reasoning are not seperate - implicit knowledge stored in the connection weights, and computations involve adjusting these weights based on learning from data

Question 14

What is ImageNet?

Answer 14

challenge to build a neural network to classify over a million labelled images of natural scenes into their correct categories from 1000 possibilities, most substantial drop in error rate came in 2012 with introduction of CNNs e.g., AlexNet, error rate has now dropped below human level (about 5%), shows strengths of sub-symbolic AI, especially when take inspiration from biological brains

Question 15

What are some examples of sub-symbolic AI?

Answer 15

Neural networks e.g., CNN, RNN, NPCs in video games, RL

Question 16

What is a CNN?

Answer 16

type of feed-forward neural network specialised for processing visual data e.g, images through heirarchal feature extraction using filters with local receptive fields, similar to primate visual system

Question 17

What is an RNN?

Answer 17

tyoe of feedforward network where hidden units receive additional input from themselves in the previous timestep

Question 18

What are the strengths of subsymbolic AI?

Answer 18

can learn to do things we don’t fully understand, good at working with large datasets (paralell processing), can generalise to understand new inputs, better as models of biological brains

Question 19

What are the weaknesses of subsymbolic AI?

Answer 19

black box, data dependency, difficulty with logic based reasoning, limited domain transfer

Question 20

What are hybrid AI systems?

Answer 20

combine elements of symbolic and subsymbolic AI, mechanisms for integrating e.g, combining symbolic reasoning with the learned representations from neural networks.

Question 21

What is Alpha Go?

Answer 21

hyrbid AI for playing board game Go, combines tree search for decision-making and a deep neural network for evaluating board positions and guiding the search process, able to beat world champion

Question 22

What is a feedforward neural network?

Answer 22

data flows in one direction from input to output through “hidden” internal layers comprised of individual units or “neurons”, each connected to everything in the previous layer via weighted connections

Question 23

What is AlphaGo Zero?

Answer 23

sub symbolic AI system, able to beat humans at Go and other games through neural network trained using reinforcement learning from self play