Cog Sci 1- Midterm 2

Front 1

• Parallel constraint satisfaction

Back 1

Process by which a problem is solved by using a parallel algorithym to find the best assignment of values to interconnected aspects of the problem. Example: Necker cube

Front 2

Connectionist models: sparse

Back 2

Information is represented by a few neurons within a largely populated neuronal network. Visual cortex processing representation.

Front 3

• Neuroscience basics

Back 3

o Neuron, synapses, action potential...

Front 4

• Connectionist models

Back 4

o Distributed, sparse, local representations

Front 5

Two types of connectionist networks

Back 5

ANNs (Artificial neural networks) or PDS (Parallel Distributed System)

Front 6

Connectionist models: local representations

Back 6

grandmother neuron. One particular neuron that fires to achieve a result.

Front 7

Connectionst models: distributed

Back 7

multiple neurons working simultaneously and in conjunction to achieve a result. Direction of movement as example.

Front 8

• Connectionism

Back 8

style of modeling based upon networks of interconnected simple processing units

Front 9

• Connectionism o History

Back 9

Ramon y Cajal- showed that neurons were physically separated from other neurons. In 1943, McCulloch & Pitts published "A Logical Calculus of Ideas Immanent in Nervous Activity“ neurons with a binary threshold activation function were analogous to first order logic sentences. Perceptrons.

Front 10

• Connectionism o Single-layer perceptron

Back 10

cannot XOR function

Front 11

• Connectionism o Basic elements

Back 11

1. A set of processing units. 2. A state of activation 3. An output function 4. Connectivity patternamong units 5. Activation rule 6. Learning rule 7.Environment

Front 12

• Connectionism o Activation functions

Back 12

step funciton; on or off. Segmoidal is gradual

Front 13

• Connectionism o Linear separability

Back 13

ANNs learn decision boundaries

Front 14

• Connectionism o Multi-layer perceptron

Back 14

has hidden layer- resolves XOR function issue- no decision boundary

Front 15

• Connectionism o ANN topologies

Back 15

3 layer, feed forward and recurrent

Front 16

• Connectionism Learning paradigms- supervised

Back 16

Compute desired outputs from given inputs–Criterion: Similarity between desired and computed outputs –Applications: Classification, prediction- reads handwriting for post office (predicts what a letter is based on trained net experience.) parallel to concept learning in humans

Front 17

• Connectionism o Learning paradigms- reinforcement

Back 17

–Select actions to maximize rewards • Requires a task and a policy–Criterion: Reward accumulated over time –Applications: Mobile robots, making plans

Front 18

• Connectionism Learning paradigms- unsupervised

Back 18

–Find structure in given inputs–Criterion: Ability to predict where new inputs are likely to be –Applications: Clustering, outlier detection. Same as supervised but without given outputs to train with.

Front 19

• Perceptron learning rule

Front 20

• Backpropagation

Back 20

Supervised learning algorithm – “Backwards propagation of errors”: the errors propagate backwards from the output nodes to the inner nodes.• Used in feed-forward networks • Requires differentiable transfer function • Algorithm: gradient descent – Finding the weights that minimize the error

Front 21

o Gradient descent

Back 21

Present a training sample to the neural network• Compare the network's output to the desired output. Calculate the error in each output neuron. • For each neuron, calculate what the output should have been, and a scaling factor, how much lower or higher the output must be adjusted to match the desired output. This is the local error. • Adjust the weights of each neuron to lower the local error. • Assign "blame" for the local error to neurons at the previous level, giving greater responsibility to neurons connected by stronger weights. • Repeat the steps above on the neurons at the previous level, using each one's "blame" as its error. how to get to global minimum in backpropogation in errors vs constraints in supervised networks.

Front 22

• Hebbian learning

Back 22

“When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one of the cells firing B, is increased.” –Unsupervised – Synaptic strength (weight) is increased if both source and target neurons are active at the same time –Otherwise, weight is decreased

Front 23

• Hopfield network

Back 23

• Performs like an associative memory–content addressable memory • All connections are symmetric (wij = wji)• Binary threshold units • Network dynamics convergence is guaranteed –i.e. local mini- mark will converge to a "remembered" state if it is given only part of the state. (handwriting) Limitations –If patterns too close network can’t reconstruct them –Net structure is rigid: only 1 layer, and there has to be the same number of neurons as inputs

Front 24

• Graceful degradation

Back 24

Aka fault-tolerance • Property that enables a system to continue operating properly in the event of the failure of some of its components. • Each unit participates in the storage of many patterns, and each pattern involves many units The loss of a few components will degrade but not loosethe information

Front 25

• Neurological plausibility

Back 25

"Natural neural networks –Billions of neurons, trillions of connections –Each neuron has dozens of neurotransmitters • Brain as electro-chemical machine–Pre-and post synaptic plasticity –Single neurons do not have local representations • e.g. conceptual/propositional interpretations –Except for grandmother neuron!–Synapses are one-way (vs. symmetric links in ANNs) –Connections with other neurons are either excitatoryor inhibitory, but not a mixture –Networks are considerably larger than current ANNs"

Hint 25

None

Front 26

• Artificial Intelligence

Back 26

Artificial Intelligence:“making a machine behave in ways that would be called intelligent if a human were so behaving”. (John McCarthy 1955). • Robot. Slavic word for “worker”. Used for the first time in a theatre play by Kapex, 1921. – Robot Inst. of America (1985):A programmable multifunction manipulator designed to move material, parts or specialized devices through variable programmed motions for the performance of a variety of tasks. • GOFAI: “Good and Old Fashioned AI” – Symbolic reasoning ignore connections to the world• Robotics was a small branch of AI – Mostly assembly– Few mobile (Shakey, CART)

Front 27

Turing test

Back 27

Paper: Computing machinery and intelligence (Alan Turing, 1950) –test of a machine’s capability to perform human-like conversation • It goes like this: –A human judge engages in a natural language conversation with two other parties, one a human and the other a machine. –If the judge cannot reliably tell which is which, then the machine is said to pass the test. –It is assumed that both the human and the machine try to appear human.

Front 28

• Classic AI

Back 28

mind needs real world interaction: Examples of problems with classic AI • Computer Vision –Successful in constrained environments • e.g. factories with constant lighting conditions, geometry, etc –In the real world these conditions never hold • e.g. objects partially occluded, distance of objects from the eyes changes, lighting conditions, etc • Locomotion, object manipulation... –Humans and animals perform it naturally and elegantly!

Front 29

• Embodiment

Back 29

“Of all of these fields, the learning of languageswould be the most impressive, since it is the most human of these activities. This field, however, seems to depend rather too much on the sense organs and locomotionto be feasible.” embodied metaphor- affection as warmth. embodied robotics, feldman event structure.

Front 30

• Applications

Back 30

–Classical approach to AI: Many applications (clever algorithms, expert systems, etc) –Embodied AI: more limited (educationaland entertainmentareas)

Front 31

• Simulations vs. real robots

Back 31

simulations cannot predict all variables and threfore cannot accurately completely represent the real world.

Front 32

• Classic paradigm (horizontal decomposition)

Back 32

sense-think-act “A chain is only as strong as its weakest link”•Cognitive process is not dependentonthe hardware (i.e. actuators)•Decisions are mediated by knowledge•Knowledge should be explicitly present

Front 33

o Sense-think-act, examples

Back 33

Shakey, assembly robots

Front 34

• Brooks argument (Intelligence without representation)

Back 34

Brooks argued that human intelligence is too complexand poorly understood –How do we know what are the right subpiecesof the problem? • Approach: Creating AI by –Building up in complexity, having complete systems at each step –Test the systems in the real world (i.e. with sensing and acting) • Conclusion –It’s better to use the world as its own model• Hypothesis –Representation is the wrong unit of abstraction

Front 35

• Behavior-based paradigm

Back 35

Ghengis-Situatedness refers to the robot’s ability to sense its current surroundingsand avoid the use of abstract representations. -Embodimentinsists that robots be physical creatures and thus experience the world directly rather than through simulation. “Simulations are doomed to succeed!”(R.A. Brooks) insect like robot that senses environment and adapts without programming.

Front 36

• Subsumption architecture o Layers

Back 36

Layers of behaviors –SubsumptionArchitecture is built in layers –Each layer gives the system a set of pre-wired behaviors –Higher levels build (subsume) upon lower levels to create more complex behaviors –All the layers operate asynchronously –The behavior of the system as a whole is the result of many interacting simple behaviors Attila (MIT, 1989)

Front 37

• Subsumption architecture o Augmented finite state machines

Front 38

• Examples in robot implementations: Herbert, Genghis (Attila)

Back 38

Herbert- collects soda cans, Ghengis- insect like, follows humans with heat sensors, walks,

Front 39

• Discussion o Who has the representations?

Front 40

• Discussion o What subsumption architecture is not

Back 40

It is not connectionism/ANNs – Connectionists expect representations to arise from networks• Representations are not necessary! Appear on the observer’s eye/mind – No biological significance between ASFMs& neurons (network nodes)• It is not production rules – Analogy with IF-THEN rules... • ...but layers run in parallel– There is no matching with rules in a database • It is not a blackboard – Analogy of SA being a blackboard architecture• Metaphor for group of experts gathering around a blackboard to collaboratively solve a complex problem – But processes are already “hardwired”to the correct place (i.e. no flexibility on how to gather knowledge)• It is not German (Heidegger) philosophy– Behavior-based approach has similarities– But SA is based on engineering principles

Front 41

• Discussion o Limits to growth

Back 41

How many layers can be built in the subsumptionarchitecture before the interactions between layers become too complex to continue? –Not many (<10), scalability problem!• How complex can the behaviors be that are developed without the aid of central representations? –From Herbert (soda can collector) to Cog• Can higher-level functions such as learning occur in these fixed topology networks of simple finite state machines? –Examples from nature on “learning by instinct”(pre-wired) • Honey bees: distinguish among flowers, routes from/to hive, etc• Butterflies: total amount of information learned is constant –Current technology (~1987) is (was) not flexible enough

Front 42

• Collective intelligence o Swarm intelligence: Stigmergy

Back 42

Stigmergy: method of indirect communication in a self-organizing emergent system where its individual parts communicate with one another by modifying their local environment. They drop pheromones. Termite hills

Front 43

• Collective intelligence summaryl

Back 43

–No direct communication between individuals –Information communicated through the state or changes in the local environment –Individuals decides action based on a simple algorithm –Critical number of individuals needed for intelligent behavior to arise

Front 44

• Body and World: different views from Psychology

Back 44

Gibson. We perceive in order to operate on the environment: Perception is not about inferences, it is “designed”for action. • Information is directly conveyed to the brain without computations on representations –How? Affordances • Gibson called the perceivable possibilities for action affordances • Affordanceis an action that individuals can potentially perform in their environment. –e.g. surfaces for walking, handles for pulling, space for navigation, tools for manipulating, etc. • Main idea: Our whole evolution has been geared toward perceiving useful possibilities for action.

Front 45

• Body and World: different views from Psychology, Philosophy, Linguistics...

Back 45

–Being-in-the-world argument • “We function in the world because we are part of it”(Heidegger,1962) • i.e. there is no division between the subject and the world –Dreyfus (1991) pushed this further • CRUM is hopeless because our intelligence is inherently non-representational –Smith (1991) proposed “embedded computing” • Emphasis on the interaction with the world rather than internal processing

Front 46

• Body and World: different views from Psychology, Philosophy, Linguistics...

Back 46

"• Situated Action –CRUM neglected the role that situations and contextplay in human problem solving & learning –e.g. Antonio Torralba’scontext challenge • How far can you go without running an object detector?

Front 47

• Body and World: different views from Psychology, Philosophy, Linguistics...

Back 47

• Situated Action (Schuman, Lave & Wenger) –Problem solving in realistic contexts depends on direct interaction with the world and other people. • e.g. Using a complex machine like a computer by learning how to interact with it, not by having an abstract representation of it –Main point: CRUM is incapable of appreciating the contextual ways in which people deal with the world

Front 48

o Affordances, intentionality, situated action... all what is in slides 16-26

Front 49

• Thagard’s responses

Back 49

• Thagard’sresponses –Denial (of the body and the world) • Only for CRUM fanatics...–Expand CRUM • Pointless? e.g. house on top of the hill – on (house, hill) does not capture all the knowledge given by the perceptual experience –Supplement CRUM • Combination of nonrepresentational operations of brain and body with computational procedures that seem crucial for high-level cognition–Abandon CRUM • Thagard’sway out... dynamics

Front 50

• Computer models, strong AI argument, functionalism (Chomsky), brain and language (Feldman)

Back 50

functionalism- mental states exist only for causal reasons- a type of materialism. Chomsky brougth down behaviorism with universal grammar.

Front 51

• Framing, neuron binding, metaphors and the brain (Lakoff)

Front 52

• Intentionality, Chinese room argument, computers don’t have semantics... (Searle)

Front 53

Perceptron history

Back 53

Initial excitation: Rosenblatt (1957) Simplest kind of feedforwardneural network: a linear classifier –Strong interest in modeling how neural networks might contribute to thought • Dark winter: Minsky& Papert(1969) –Single-layer perceptroncan not learn XOR function–AI & Psych shifted to rule and concept-based systems • 1980s: ANN renaissance –Connectionism is born (Hinton & Anderson (1981); Rumelhart& McClelland (1986)) • Multi-layer perceptron

Front 54

monoism (contrast to dualism)

Back 54

has behaviorism, materialism and idealism

Front 55

Mars 3 levels of analysis

Back 55

computational (goal), algorithymic (strategy), implementation