Please enable JavaScript.
Coggle requires JavaScript to display documents.
Cognitive Science (Vision, Object Recognition and Attention (Feature…
Cognitive Science
Vision, Object Recognition and Attention
Human Visual System
Consists of 32 separate systems, each specializing in a different visual characteristic such as colour, motion, location, and other visual properties
Binocular Vision and 2D vs 3D
Binocular Vision: vision wherein both eyes aim simultaneously at the same visual target and work together to generate a coherent visual image in the mind
Human eyes generates a 2-D photo of the world, but the brain needs a 3-D understanding in order for you to act in the world - also known as Stereopsis or binocular depth perception
Lot of processing is done by the human mind to do this conversion and build a comprehensive 3-D world view
Object/Pattern Recognition
The ability to identify objects in the environment
Once you 'see' and 'understand' it, you will automatically see it the way after that - humans have the power to filter out parts of what they see and also fill in missing pieces
What we see is interpreted by our brain using conceived knowledge
Template Matching Theory
An image generated from a stimulus is matched to an internal representation called a template
Feature Detection Theory
A distributed approach - explained using a group of demons in Pandemonium model
Different components in the brain identifies different features and announces their findings to others
A higher level cognitive component listens and gathers information about all the features identified so far and recognizes a higher level feature/object
If an object is recognized, the higher level component announces its findings and the level of match found
A decision component then makes the final decision about what the object can be recognized as
Recognition by Components Theory
Biederman proposes a 3D feature theory for 3D objects from a limited set of 3D featured called Geons - properties include
Invariant: identifiable from different perspectives
Discriminable: differs from others from any angle
Noise resistant: can be recognized even if partially obscured
Feature Integration Theroy
Developed by Treisman and Gelade
Used to explain visual search, in which we attempt to locate a target object hidden among distractors
During pre-attentive stage, features pop out effortlessly
Attention not required
Search occurs in parallel
During the focused attention stage features are combined together to create object representations
Attention is required
Search is serial
Attention
A form of mental activity or energy that can be distributed to perform different taks
Attention is:
Selective: not to get distracted in the face of distracting or competing stimuli
Shiftable: ability to shift focus of attention and move between taks
Sustainable: maintain attention for a consistent behavioural response during continuous and repetitive activity
Divisible: ability to divide and distribute it to respond to multiple tasks simultaneously
Voluntary and Reflexive Attention
Attention plays an important role in visual perception
Two types of attentions continuously compete to be the momentary foci of attention
Voluntary attention: we have control, enabling us to act in a goal-directed manner
Reflexive attention: driven by exogenous sensory stimuli redirecting our current focus of attention to a new stimulus, bottom-up attention
Models
Treisman's Attenuation Model
Unattended message is not blocked completely but attenuated
Problem with Broadbent's model: Cocktail Party Effect - important info get through
If a stimuli passes the threshold, it will leak through the filter and can be attended to
Deutsch-Normal Selection Model
More recent and supported by other studies
Selection happens in two stages in the attention processing sequence
At the first stage based on structure
Later at a second stage based on the semantic content when meaningful or relevant info is selected
Selected info is stored in short term memory for further processing
Multimodal Model of Attention
Selection can be based on physical or semantic characteristics
Allows for selection to take place early or late
The filter is 'moveable' and can take place at various stages of processing based on the observer's needs
Kahneman's Capacity Model
Mental effort is needed to engage in any mental gas
The greater the complexity, the greater the mental effort needed to solve a mental task
Effort - Arousal level - determines capacity
Attention has a maximum capacity which can change based on mental arousal level
Inverted U-shaped model: at high and low arousals humans are either too anxious or too tired
Medium levels of arousal are assumed to produce the greatest amount of available capacity
In order to direct attention appropriately
Attend to relevant information
Neglect irrelevant information simultaneously
By combining the following, a person exerts enough mental effort to complete mental tasks
Total attentional capacity
Momentary mental effort
Appropriate allocation policy
Theories of Attention
Bottlenck theories explain the narrowing of attention that enters conscious awareness
Capacity theories explain how attention is distributed to different informational sources
AI and Computer Vision
Marr specifies the steps a computer would go through to recognize an object
Image - a raw primal 2 1/2D sketch + distribution of intensity values - feature identification -3D sketch with linked object parts
Recent approaches
Segment the image into regions (context included)
Using knowledge of the world, interpret the regions to determine the content of the image
Detect the edges
Memory
Humans are born free of any knowledge which is built later from experiences
Memory is the capacity to retain information over time which allows learning from previous experiences
Memory systems can be characterized by coding capacity, storage duration, and recall
Later: Humans have two types of memory
The natural memory (inborn one)
The artificial memory (acquired through learning and practice of a variety of mnemonic techniques)
Process
Perception of Stimuli - Engram - Consolidation - Long term Potentiation - Storage - Recognition / Inference / Reconstruction - Recall / Retrieval
Distributed Memory
Semon : Memory is encode as an engram, a physical trace of experience, left on specific webs of neurons in the brain
Storage and recall in memory is a brain-wide process that happens in several different areas of the brain in a collaborative manner
Each elect of a memory (sight, sounds, words, emotions) are encoded in the same part of the brain that originally created that fragment (visual cortex, motor cortex, language area)
The recall of a memory effectively reactivates the neural patterns generated during the original encoding
Sensory Memory
Information from the different sensory modalities is stored in separate sensory memories
Iconic memory is a visual sensory store with a short duration of less than 1 second
Echoic memory is an auditory sensory store with a duration several seconds long
Experimental Study
Sperling presented letter arrays
In the whole report condition, participants attempted to recall the entire array but could only remember 4 or 5 letters
This shows iconic memory has a high capacity, capable of storing most of the information seen in the visual field
In the partial-report condition, the rows were cued and the participants were told that they would be required to report a row when they hear the same sound used for cue
Next when the sound was played the participants could remember all the letters
Decay
Iconic memory is short-lived and fades rapidly
Loss of information over time in memory is called decay
Instead of a cue if a blank inter-stimulus interval (ISI) is used then experiments showed that it erases the icon formed for the first row which is called masking
Other experiments showed that blinking interrupts binding of object to its position in iconic memory - results in error and remembering the location of the letters - a scenario known as cognitive blink suppression
Short term / working memory
Has a limited capacity less than iconic memory
Can be increased by chunking
Coding can be acoustic, visual or semantic
Has a limited durations. Information can decay in seconds - better retention possible by rehearsal (repeating items)
Long term memory
Consists of several distinct subtypes
Explicit or declarative memory holds memory for facts or events
Demonstrated by saying, and occurs with conscious recall
Implicit or procedural memory holds knowledge for skills such as riding a bicycle
Demonstrated by actions and occurs without conscious recall
Declarative long term memory
Tulving first proposed two distinct kinds of declarative long term memory
Episodic memory contains personally experienced events and is organized temporally or spatially
Semantic memory contains factual knowledge in general and can be organized hierarchically or in context with other knowledge
Model
The Modal Memory Model
Shows how information is transferred between the major memory types
The first major process model
Formulated by Atkinson and Shiffrin
The Working Memory Model
From Baddeley
Shows interactions between components of working memory
Brain and Memory
Cerebellum - plays an important role in processing procedural memory
Basal Ganglia - formation and retrieval of procedural memory
Thalamus - evolved to help relay information from the brain stem and spinal cord to the cerebral cortex
Hippocampus - transfer between short and long term memory
Cerebral cortex plays a key role in memory - consists of 4 lobes that perform many functions including processing short term memories and retaining longer term memories
Amygdala - primary role in the processing and memory of emotional reactions
Mental Model
Key points
People are rational in principle but fallible in practice
There are three main classes of theory about the process of deduction
Formal logic
Content-specific rules
Mental models
Formal logic is psychologically implausible because people are affected by the content of deductions
Content-specific rules view ignores the fact that people are able to make valid deductions based solely on logical propositions and rules of reasoning
Mental models form the basis for various kinds of reasoning
Language
Language can be interpreted very differently based on individual perception
A mental model is a structure that stands for something in the the world which is used in mental processes
Humans form mental models from the semantic interpretation of statements - what we hear and understand
Mental models very based on the communication and the language
We pay more attention to whether the model makes sense than whether its logical
Examples
Abstract: from statements based on perception
General: features, knowledge
Visual: maps, images
Theory
Johnson-Laird proposes that humans solve logic problems using subjective models of the situation, rather than a set of rules
Johnson-Laird and Byrne argue that deductive reasoning is carried out neither by formal logic nor by content-specific rules or schemas but by mental models, which are mental representations that correspond in structure to the situations that they represent
Deductive Reasoning
First generate an example to conform with argument (reject is invalid if cannot be done)
Generate an alternative example and continue until all possibilities are explored
Hold all examples in memory
Rule vs Model
Rules always work and are easier for large tasks but rules are difficult to learn and remember, and get complex with 'and-or' and chaining of if-then
Intuitively prefer models but wrong models can lead to error
Rule-based or proof theoretic inference uses rules of reasoning to draw valid conclusions
Model-based or model theoretic inference draws conclusion by generating examples to explore all logical possibilities
The Wason Selection Task
Inderdisciplinaries
Neuroscience
Study of nervous system anatomy and physiology in human and other species
Cognitive neuroscience studies neural structure and processes underlying human cognitive function
Methods
Case Study
Lesion Study
Directions
Anterior 'front'
Posterior 'back'
Ventral 'bottom'
Medial 'middle'
Dorsal 'top'
Lateral 'side'
Major Cortical Features
Two cerebral hemispheres on each side, left and right
Corpus callosum
nerve fibres join the two hemispheres of the brain
Gyrus
fold of cortical tissue
Fissure
deep cleft or separation between gyri
Hippocampus
One in each side of brain
Consolidation of information from short-term to long-term memory and spatial navigation
Split brain
Information from one side of body is mapped onto the contra-lateral side
In patients, the corpus callosum is severed
Brain recording techniques
Brain electrical activity measurements
Single-cell recording
an electrode is inserted into or adjacent to a neuron
Multiple-unit recording
a larger electrode is used to measure the activity of a group of neurons
EEG
PET scan
fMRI
CAT scan
Magneto-encephalogram
Neurons
Black reaction
Discovered by Camillo Golgi
Method of staining nerve tissue
Stark black deposit on the soma as well as on the axon and all dendrites, providing an exceedingly clear and well-contrasted picture of neuron against a yellow background
Parts
Nucleus
Dendrites
Soma
Myelin sheath
Node of Ranvier
Schwaan cell
Axon terminal
Synapse
Electrical potential
Have potassium and sodium ions which cause electrical flow inside our brain
Thin outer membrane of the neuron is made of proteins, which have channels that may allow the passage of ions
By maintaining an uneven match of positive and negative ions, the neuron has an electric potential across the outer membrane
Potential
Resting potential
Difference between the voltage inside and outside the neuron at a stable non-firing state
Incoming signals have the effect of locally altering the potential at the dendrites where the signal arrives
If the incoming signal is maintained it will very quickly change the potential of the whole cell body
Slow potential
Transfer point
Cumulative at the axon base
The affect of different incoming signals having different effects on the local potential
Thresholds for transmission
Ion pumps
Effect of the potential at the axon base has little or no influence on the potential at the site of contact without the mechanism of ion pumps
Action potential
If the potential at a point along the axon rise beyond a threshold, ion pumps at that point will operate, and cause a sudden large local increase
Frequency
Encodes the signal from the neuron
Because of resting period between action potentials, maximum firing rate is about 200 per second
Transmission through synapse
The same signal transmitted to all of recipient neurons, can cause different effects on the recipient neurons depending on type and ions at the synapses
When an action potential arrives at the end of an axon it causes the release of many tiny neurotransmitters
Dopamine
Hormone and neurotransmitter
Reinforces happiness
ADHD
Alcohol and other drugs reduce dopamine and inhibits the inhibitory signal
Linguistics
The study of language
Variety of theoretical approaches and methodologies
Grammatical rules, animal language, development, and computer speech recognition
Characteristics
Communicative
Allows for communication
Arbitrary
Consists of symbols that refer to and stand for something
Generative
Symbolic elements can be combined to generate a large number of meanings
Structured
Symbols are ordered, governed by rules to form a structure of a sentence or expression
Dynamic
Incorporate changes, addition of new words and rules
Grammatical rules
Phonology
Rules governing sounds
Phoneme is smallest unit of sound in the sound system of language
Semantics
Rules for understanding meaning
Morphology
Rules governing word structure
Morpheme is the sound of the smallest unit of spoken language that has a meaning
Syntax
Rules for arranging words in sentences
Language
Acquisition
Babbling
stage
Utter a smaller set of phonemic sounds
One-word stage
Speak out words and morpheme
Cooing stage
Begin to utter a wide range of sounds
Two-word stage
Production of two-word sentences
Understand past tense
Development
Follows a U-shaped pattern of development
Children start with copying or imitating
Proceed to understanding the rules of grammar
Progress to learning the exception to rules
Deprivation
Critical period
If linguistic experience is missing in critical period, language ability is impaired
Language and evolution
Evidence
Universal characteristic of language
The development stage of a child
Defined brain areas
Aphasia
disorder caused by damage to the parts of the brain that control language
Broca's Aphasia
Understand well but have problems producing speech
Generate short, broken sentences with many pauses
damage to the lower left frontal lobe
Wernicke's Aphasia
damage to posterior portion of the left hemisphere
Problems comprehending speech; generates irrelevant replies
Produces rapid, fluent and seemingly automatic speech with little to no meaning
Wernicke-Geschwind Model
Wernicke created model of how we process language in our brain
Geschwind later expanded this idea
Over simplified explanation of the cognitive process behind language processing by the neural system
Criticism
Other brain areas are involved in language processing such as memory
Symptoms are not always consistent with problems in those brain areas
Psychology
Analyzing results
Scores on dependent variable for the two groups are compared
If the test scores in the control group are significantly higher, then the hypothesis is supported
Experiments
Theory
Hypothesis
Experiment
Accept / reject
Variables
Independent variable
Dependent variable
Groups
Control group
Experimental group
Investigates internal mental processes such as reasoning, visual imagery, language, memory and observable external behaviours such as problem solving, decision making, talking, response time, and accuracy of recall
Seeks answers such as what makes up the mind, how do the parts interact and generate behaviour
The study of mind and behaviour using scientific experimental methods to gain knowledge and validate theories
View and Approaches
Gestalt Psychology
Explores mind as a conscious integrated whole, Gestalt, which is more than the components
Method: phenomenology
Insight learning
Preparation
Understanding the problem and preliminary attempts
Incubation
Problem is put aside for some time
Subconscious thoughts
Verification
Confirm the solution works
Illumination
'Aha' stage, becomes aware of the solution
Law of Perceptual Organization
Ways visual parts group together to form objects
Similarity in lightness, colour, shape
Closure
Proximity in visual field
'Pragnanz' or good figure tend to go together
Functionalism
Explains what the mind does, mental activity, rather than elements
Voluntarism
Elements of the mind form higher order elements voluntarily by the power of will
Method: introspection
Structuralism
Structure of mind can be understood from its elements which combine based on a chemical law
Method: introspection
Three states of consciousness
Preconscious
Conscious
Subconscious
Psychoanalytical Psychology
Idealistic principle
Reality principle
Pleasure principle
Types of Learning
Classical conditioning
Applies only to involuntary reflexive behaviours - no conscious though
Unconditioned stimulus (US)
Conditioned response (CR)
Unconditioned response (UR)
Conditioned stimulus (CS)
Operant conditioning
Applies to any voluntary motor act and is more general
Punishment
Reinforcement
Anthropology
The study of human origin, social and cultural influence on human upbringing and behaviour, and evolution of Homo sapiens
Approach
Comparative Cognition
Similarities and differences in cognitive abilities between different species
Evolutionary psychology
Evolution of the mind
The selection forces that may have given rise to our current mental state
Behavioural economics
How evolution may have shaped the way we make decisions involving money
Comparative
Comparative Cognition
Animal memory
Object permanence
Ability to know that an object exists even though it cannot be seen
Problem solving
Kohler's chimps and insight learning
Transitive inference
Involves knowing that if A is bigger then B and B is bigger then C, then A is bigger then C
Each species has adapted to its ecological niche
Cross-species study of cognitive ability
Comparative Neuroscience
Involves comparing the brains of different animal species
Are animals with bigger brains smaller then animals with smaller brains
We must adjust for body size
Cephalization index
Proportion of brain size to body size
Should ideally only consider cognitive size
Factors other than intelligence influence the brain size
Birds have small brains because they need to be light to fly
Dolphins have larger brains because water can support a larger body weight
Not every part of the brain does cognition
Hindbrain mostly regulates basic physiological function
Cortex, especially neocortex, part of closely linked cognition
Neocortex is 80% of total brain volume for humans In primates its 50% (monkeys) and rodents its 30%
Only mammals have neocortex
Evolution
Natural selection
Variation in traits within a species
Selection, which is a change in the environment that favours one trait over another - survival of the fittest
The passing of genetic material from one generation to the next, Inheritance
Evolved psychological mechanisms
Cosmides and Tooby
Each device evolved under selection pressures to solve a specific problem and can be considered as a module
Get activated under special environmental context
Categorization
We form concepts in a graded continuous fashion, all or none
Natural categories are continuous, the mind mimics this organization
Allows us to generalize our knowledge from one category to another
Concepts are organized around representative members of a class, the typicality effect
Memory
Better memory for information we are exposed to more often because it is more relevant or important for our survival
Recall for words is proportional to their frequency for occurrence
Logical Reasoning
From the EP perspective, under certain conditions such as 'cheater detection' which is related to survival, humans are good in solving logical problems
Wason Selection Task
If both answers were correct it means either lucky or good logically
Most people get 2nd correct because it's related to survival and cheater detection
Judgement under Uncertainty
Uncertain judgments: when we make decisions without complete information
Most everyday decisions in life
Heuristics
Past records or experiences
We often rely on this
Fallicies
Heuristics can often lead us to commit fallacies
An argument that uses poor reasoning
Misunderstanding of statistical rules
3 Types
Base-rate fallacy
Ignoring base rates
Ex) Jack the lawyer / engineer
Gambler's fallacy
Ignoring independent outcomes
Ex) gambling
Conjunction fallacy
Ignoring the conjunction rule
Ex) Linda the bank teller
Language
Evolved to promote social bonding
Allows for complex coordinated social behaviour
May have been a key factor as to why neocortex have grown more in humans
Most individuals are left-hemisphere dominant with respect to language with specific localized areas devoted to language comprehension
Implies that language capacity is pre-sepcified and present at birth and it genetically coded, and shaped by selection forces
Sex Differences
Attributed to a sexual division of labour in which men hunted and women gathered
Hunting may have fostered increased spacial ability in men
Gathering may have promoted increased verbal abilities in women
Distinction is too broad, women are better at locating objects from memory
Approaches in AI
Study of manmade systems that behave in ways characteristic of natural living systems
Computer 'creatures' are created through evolutionary rules. They navigate, seek out prey, and avoid predators in virtual environment
Complex adaptive behaviours emerge from such systems
Genetic algorithm
Procedure in machine learning, applies the concepts of fitness for survival and reproduction to generate a variation of possible solutions
Behavioural Economics
The study of economic decision making and the factors that affect it
Classical view is that people are rational and always make choices that increase their money or value
In reality the way we think about money is affected by heuristics and evolution
Examples
The Ultimatum Game
Loss Aversion
Loss Aversion and Framing Effect
Mental Accounting
Endowment Effect
Sunk-Cost Fallacy
Philosophy
Search for knowledge
Methods
Inductive
Generalization of observations
Abductive
Explanation: drawn from facts (not definitive)
Deductive
Assertion: given the facts (definitive)
Approach
Define problems, raise questions, seek answers by observing and arguing about facts or data by reasoning, and acquire knowledge
Two branches
Metaphysics
Traditional branch, explores mind-body problem
Epistemology
Study of knowledge, explores acquisition and representation of knowledge
Common problems addressed
Knowledge to Acquisition Problem
Declarative knowledge
Facts that are acquired by observation (probably not innate)
Procedural knowledge
Can be both acquired and innate
Smell preference and reflex is innate
How to fix a bike or computer programming skills are learned through experience and observation
Association
Simple ideas are acquired unconsciously and complex ideas are learned by reflection
Empiricism
Knowledge is acquired through experience
Rationalism
We have both innate knowledge and use reasoning to build new knowledge
Nativism
Born with knowledge
Consciousness
Definition
Consciousness as one's own individual private mental life (opposite subconscious)
Being conscious of various aspects of one's existence, actions, and environment as awareness influences behaviour
Conscious as opposed to being asleep or in a coma (unconscious)
Consciousness does not equal awareness
Consciousness: able to foresee and apply intelligence
Awareness: knowledgable about surrounding
Awareness is considered as a prerequisite for consciousness
Qualia
The subjective quality of experience
Knowledge cannot generate the subject experience
Reductionist
Way to try to define the whole as a sum of the components
John Searle defines consciousness as emergent property of the brain. Based on the concept of emergence, mind can be considered as an emergent property of the brain
Cartesian Theatre
Consciousness is generated at a single center point of the brain where all information funnels in
All information about an event is not received at once
No single point in the brain that receives input and generates motor reaction as the central processing unit in a computer
The Mind Body Problem
Nature of Mind (Philosophical Positions)
Dualism
Substance Dualism
Mind and body are made of two separate substances
Property Dualism
Mind and body are made up of the same stuff but have different properties
Classical Dualism
Mind controls the body
Functionalism
Mental states are not only physical states but the functioning or operation of those physical states
Monism
Idealism
Everything is mental
Physicalism
Everything is physical
Identity Theory
Mind is the brain
Physical substance (brain) and non-physical substance (soul, consciousness, thought, desire, beliefs)
Relationship between mind and body
A Classical Problem
Artificial Intelligence
The goal of work in artificial intelligence is to build machines that perform tasks normally requiring human intelligence
Machines are programmed to demonstrate human intelligence in
Problem solving and decision making
AI and CogSci
Explore how humans do such intelligent tasks and develop algorithms to build machine intelligence
What makes us intelligent
Ability to learn from experience
Apply the best techniques to meet our goals
One limitation: memory
What makes us wise
Ability to predict the unseen and find the best path to follow given our goals and present situation
Techniques
Search
Find all links to the answers
Many different search algorithms exist. The right one is chosen based on the problem, data, constraints, and available resources
Learn
Get feedback from previous work and try to reduce error
Machine learning methods
Knowledge Acquisition/Retrieval
Generate usable knowledge from data by analytics
Statistical and machine learning methods
Decision Making
How to quickly find the best decision
Methods applied
Rule based systems
Case based systems
Hybrid systems
Apply multiple constraint satisfaction techniques
Cognitive Science
Implement mechanical mind - machine intelligence - strong AI
Cognitive simulation: simulate processes of the mind
Intelligent Agents
AI approach implements multiple intelligent agents (AI), small autonomic or self-managing systems computer software or system, capable of learning and executing a specific intelligent process
Approaches to designing IA's
Turing's Finite State Machine (TFSM)
Rule-based decision making
Craik's model of cognitive/IA
Decision Support Systems
Rule-based
Decisions are guided by rules
Ex) expert systems
Employs knowledge (facts); logic (rules, primarily in the form of an if-then structure) and algorithms or procedure for searching and choosing rules
WebMD is a popular Expert System that is freely available on the internet
Challenges to Expert Systems: difficulty encoding rules, static, representation, and search
To overcome some limitations, Expert Systems may include statistical information
Case-based
Decisions are guided by example cases or heuristics
Previous use-cases serve as guide for decision making and reasoning applied to choose the case
Ex) A case is defined by disease symptoms and paired with treatments, a case is the characteristics of a problem on a device and the corresponding solution
Search
Blind Search
Contain no information about the layout of the problem space
An exhaustive search inspects each state to find a goal state
Breadth-first search
Check all the states you can reach in one step from the current position and then select one in a sequence and explore all the states you can get to next. If a goal state is found, stop, otherwise continue
Helps to find a better solution since all options are explored at early stage
Depth First
Follows a search path to the end before moving on to another path
A limitation with depth-first solutions is that an optimum solution may not be found
Other Learning Techniques
Statistical modeling
Applies different statistical modelling approaches
Evolutionary computing
Applies concepts such as the most fit solution survives and new solutions are generated by reproduction algorithms
Artificial Neural Network
Modeled after the brain neural system where information comes from many sources and their interconnections and importance decides the final outcome
Natural Language Processing
Application
Natural languages have evolved and are used buy humans
Cross disciplinary research on AI and Linguistics focuses on using computational approaches to understand and generate speech in natural language
Four Stages
Speech recognition
Steps in an automated speech recognition process
Spoken language is converted to a speech spectrogram
Phonemes are extracted from the speech stream
The phonemes are assembled to form words
Syntactic analysis
Analyze individual words in the order they occurred
Use grammar analysis techniques
Analyze the phrase structure by breaking down the sentence to its hierarchical constituents
The results is understanding how the words are grammatically related in the sentence structure
This phase strictly focuses on the structure and not the meaning
Grammar
The hierarchical relationship between the parts of a sentence are known as its phrase structure
Transformational Grammar
Phrase structure grammar doesn't tell us how we can rearrange a sentence to express new meanings
Semantic analysis
Prior phonemic analysis can produce the meaning of some words by comparing the strings of phonemes against a database of sounds
A match helps to retrieve the word meaning
If a perfect match is not found then ambiguity arises
Knowing the type of word (noun, verb) from syntactic analysis helps disambiguate and recover word meanings
Construct the overall meaning of the sentence from the meaning of the words or phrases
Pragmatic analysis
Pragmatics are the social rules of language use that enable speakers to make themselves clear from the contexts and type of speech
Five types of speech (Seattle)
Assertive
Assertion of a belief
It is hot in here
Directive
Instructions
Turn down the radio
Expressive
Describes psychological states
I apologize for yelling at you
Declarative
The utterance itself is an action
You are fired
Commissive
Commit speaker to a later action
I will take out the garbage later
Challenges
Update the word vocabulary with new words and new phrases created everyday
Efficient storage and search of the enormous vocabulary to produce real time response
Grammar rules are often not followed in many chats or has different rules in different languages
Pragmatic analysis is very challenging as computers do not have complete contextual information as humans do
AI Application
Voice recognition systems and Speech-to-text API
Siri for iPhone limited Q/A capabilities
Cortana is a similar application for windows based systems
Google STT and Watson STT Application Programming Interfaces - converts voice to text for dictation
GPS navigational systems provide directions using voice communication
Ongoing work
Research focus now on conversational systems and text composition
Introduce learning capabilities, advanced search and narrative capabilities in systems
Contextual processing, multi-lingual systems voice and image recognition systems for intelligent Q/A
Representations
Rules
Mental representation of the from IF (condition) THEN (action)
As logic, rule is a form of knowledge representation in AI
Forward chaining: Used in production systems - similar to Turing's FSM
Rules are searched for that matches a given state
One rule is selected and executed that modifies state according to the rule
This moves the system step by step towards the goal state
Different algorithms exist to backtrack and change route so that the goal state can be reached if it exists
Backwards chaining: work backwards from goal to start
Allows one to track possible causes for an action by inspecting the current state
Examples
Problem solving
Planning
Decision making
Computational Power
Learn new rules by:
Trial and error when doing some task and experiencing cause and effects
Example of learning rule by rats in Psychology experiment - rat presses lever for food bu frequency decreases when pressing lever gives a mild electric shock
From external knowledge sources (reading, viewing)
Breaking down other rules by chunking
Rules are a big part of our knowledge
Example of rule: If CAUSE then EFFECT
Using observation of effect to find cause: Effect so maybe CAUSE
Some rules may be innate: Reflexive actions, feeding, universal grammar
Rules are modular, easy to add to
Apply to problem solving and learning
Limitations
Inflexible
Difficult to express probability using simple rules
Rules have fixed structures
Difficult to control, process and follow
Can get very complicated with many branches of IF-THEN when there are many criteria that control a decision
Conditions may be complex and consist of multiple clauses that are combined with 'and', 'or', 'not'
If a condition is not satisfied, alternative route must be defined which may be another IF-THEN rule
If there are many nested rules it becomes very difficult to follow the rules
Knowledge acquisition is difficult
Tacit knowledge - difficult to transfer
Expert systems such as solving a computer problem are based on rules but it is often very difficult to extract the rules from the experts
Involves tacit knowledge
Asking the relevant questions to get the knowledge from an expert is difficult for a non-expert
Rules are generally static and require manual update every time rules change
Linguistic Rules
Language learning and use requires rules and do does NLP
There are Phonetic rules for pronunciation
Syntactic rules
Tenses, derivation of words, transformation of sentences
Innateness - we learn words and syntax by reinforcing correct speech
Formal Specification of Rules
Precondition -> action
Multiple rules can have the same condition: after the action is executed, condition is no more satisfied and another rules that matches the new state has to be searched
Find a rule using search and select one when multiple rules fit the condition
Selecting a Rule
The process by which a single selection is made by applying a set of prioritizing factors
Specificity: more specific match
Operational priority
Recency - use the more recent rule
Dependency - if the goal state depends on some sequence of state transition, follow rules that enable that sequence
Concept
Lock and Hume - concepts are learned trough sensory experience
Recent cognitive science - concepts are learned from experience and from other concepts
Concept = mental representation of a class of objects or events
Researchers introduced the following terms to describe new views of concepts
Schema
Frame
Script
Schema
Rumelhart: a structure of the related object/idea showing the constituent typical properties or features
Includes kind (categories) and part hierarchies and other associations
A structured system capable of organizing and perceiving new information
Many involve implications/rules to separate concepts
Cognitively it is useful to have a package of information that can be applied as a whole
Frame
Minsky argued that thinking should be understood as a frame of whole rather than by the structure or parts
We think of related concepts which belong to the same frame
Represents stereotyped situations - used widely in AI
Script
Schank and Abelson explained scripts as concepts that show psychological computational collaboration
Typical sequential occurrences - processes
Procedural knowledge
In AI
Use different data structures to store and organize information
Schema: Graph structure with nodes and links and algorithms to process information in this structure
Frame: A set structure used to divide knowledge in to substructures to represent stereotyped situations
Script: Array or list structure with each defining a step of the process in sequence
Mental Processes
Process of inference
Find matching or relevant concepts
Draw inference based on knowledge about those concepts
Exceptions or special properties are remembered with respect to the typical properties
Three ideas about how concepts are formed
Prototype
Exemplar
Theory
Concept hierarchy: Semantic Network
Recording all details about every different instance requires high storage and retrieval time
How much detail to store as concept
Semantic Network in AI
An economical approach
The basic concept is formed by inheriting properties from more general parent classes
Then locally stored exceptions are used to override the inheritance
Neurological Plausibility
Ashby and Waldron show prefrontal cortex and basal ganglia contribute to concept learning
Damage to different areas of brain results in specific deficits
Learning
Conceptual combination
Analogy
Using the knowledge of one object or event or experience (source) to solve another one (target) due to some similarity between the two
The similarity may be in physical feature, characteristic properties, functionality, application or the underlying concept
Often there are no established rules and concepts available to solve the target problem
Rules are available but selection and application of the right combination of rules given the constraints may be too complex
Procedures
Determine the most relevant matching criteria
Retrieve a source from memory that best matches the target
Get target
If a matching source is found
Construct a solution
Adapt the solution to suit the target better
Learn the adaptation schema
AI Technique
Case-based Reasoning
Reuse it to formulate a solution
Review the solution to verify that it suits the target
Retrieve matching cases
Retain the new solution as a case in the central repository
Computational Power
Analogy is used in decision making, reasoning, problem solving, thought, and perception
Learning: Adapt old solutions
Language perception
Limitations
Lack of previous experience
Adaptation may be complex
Analogies may be misleading and need analogy therapy
Analogy in Education
Use familiar source for analogy
Make the mapping clear
Use deep systemic analogies
Describe the mismatches
Use multiple analogies
Perform analogy therapy
Image
Psychological experiments suggest that mind thinks with both words and images
Mental image is an exact representation of our experience
Image: Mental representation that is similar to what it represents - not only visual representation but also other types of images
The mind perceives and interprets information using multiple representations but the strongest one creates the link to the information in the memory
Visual Image and the Mind's Eye
Human vision is a very complex process
Images seen by the eye are processed and interpreted by the mind to make us see what we see, which may not be the reality
Evidences support the hypothesis of mind's eye
Illusion shows what we see is unreal, therefore, what we see is actually what our mind shows us
Computer vision
Computer vision has advanced a lot in recent years but it is still crude compared to human vision
Modern computer vision research takes into account the steps involved in biological vision
Images vs Words
Images have more representational power
Images cannot replace verbal representation
Can be visual rules, concepts, and analogies
In Computation
Theories exist about the actual form of representation of image in our mind
Some say that mind uses array like structure to perform visual task
Others think that connectionist models are used in image interpretation
Visual thinking is useful where solution depends on visual appearance or spatial relationship
Different visual image processing operators help in problem solving, drawing analogies and learning language
Images may be fetched from memory for processing when objects are not in sight
Visual Image Processing Operations
Inspect
Find or Scan
Zoom
Rotate
Transform
Psychological Plausibility
Psychological experiments compute reaction times
Computation
Planning
Planning with visual representations involves creating visual maps or steps of what needs to be done based on the goal -> visual rules
Depends on individual characteristic - some do it visually and others are more comfortable in verbal or written planning with rules
Others
Decision Making
Problem solving
Learning
Language
Neurological Evidence for Imagery
Kosslyn presents evidence that parts of the brain used in visual perception are also used in visual mental imagery
Patients with brain damage (parietal lobe) that causes deficit in visual perceptions sometimes have similar imagery deficit
Mental imagery activates visual areas of the brain and largely increases blood flow to the visual cortex and hence it is used in treatment
Connection
Connectionist Approach
Information is represented as a graph with nodes storing the information and links showing the dependencies of the information pieces and how that affects the overall behaviour
Used to model the distributed nature of the mental processes that occur by the firing of the brain neurons
In AI, this representation and processing is modelled using Artificial Neural Network
Key differences from other representations
Execution is parallel unlike logic, rule, concept, image, analogy
Each node contains limited information
Can be organized in multiple layers
Types of Links
Links can be excitatory or inhibitory
Signals coming via an excitatory link act positively and those coming via inhibitory link act negatively towards generating the effect
Activation of a node is affected by the signals from the connected nodes and the types of the links
Two types of connectionist representations
Local representation
Each node is used to represent one concept/idea
The one-way excitatory connections make activation flow from observed behaviour to inferred traits
Easy to infer how each node value affects the linked node
Symmetric inhibitory link reflects that it is hard for both to be true
Distributed representation
Semantic Network
Semantic Networks for Understanding Contextual Relation
Semantic networks were originally devised as a means of explaining how word meanings are understood in language by contextual relation
Linguistic theories regard
Lexicon = inventory or words
Grammar = structural rules
Language + Lexicon + Grammar
Understanding words and grammar is often not enough to understand the language
Contextual relation among words is very important - cognitive grammar
Context-based Language Processing
Spreading Activation and Metaphors
A spreading activation model of semantic memory would start a small amount of activation at each word found in a sentence
The activation spreads to the connected concepts and total activation is higher in a more connected concept
If multiple concepts exist for the same word, the concepts with more activation is chosen to disambiguate
Proximity og concepts in a semantic network may be used to explain our use of some metaphors
Applications
Education
AI systems: Ontology, semantic web
Perception
Collins and William demonstrated that the time humans take to answer a true/false to statements about concepts is consistent with the organization of semantic network
Distributed Representation
A network of nodes and links generate the final output/behaviour from a set of input/perceptual data
Networks are trained using a set of training data by strengthening the excitatory links and weakening the inhibitory links until the network produces the expected outcome
Connection
Application and Limitation
Constraint Satisfaction: problem solving, planning and decision making often involve balancing multiple constraints
Limitations: hard to represent relations
Logic
Universal Turing machine
Implementing mental processes on computers allowed researchers to
Generate models of the mental processes which can be used to develop intelligent applications and robots
Validate their hypothesis about the mental processes by comparing the computer generated results in the human responses
The central hypothesis of COGS
Thinking = Representations of mental data + Mental Process
Can do any computation given a set of input data and an algorithm defined as a state machine
Categories of Representation
Logic Propositions
Rules
Concepts
Analogies
Images
Connections
Mental model
Rationality
Hum beings are mostly rational that is, human behaviour can be explained by reasoning based on belief and/or logic
Rationality implies that there was some sort of reason behind a behaviour - may not have a firm truth value
Logic is firm and conclusions are made based on premises that have definite truth value
Choice of a restaurant can be rational (not critical) vs choice of a route to reach destination on time (critical)
Planning to travel after Christmas is a logical decision
Logic is used in critical problem solving and decision makingl
Mathematics
To stimulate cognitive processes of problem solving and decision making, we need to represent logic in computer programs
Facts and knowledge
A formal representation of logic was defined later to avoid ambiguity inherent of natural language
History
Syllogism means inference which is a kind of logical argument that applies deductive reasoning to arrive at a conclusion based on two or more propositions that are asserted or assumed to be true
Fact 1: If you need to go out in the rain, then you should take an umbrella
Fact 2: It is raining now
Fact 3: I have to go to school
Conclusion: I should take an umbrella
Logic Operations and Truth Values
We can use logic propositions and rules to validate the truth value of the hypothesis/decision
Deductive reasoning
Uses facts and implications
Abductive reasoning
Logical inference 'may be' that is drawn backwards using logic implications
Inductive reasoning
Creates new knowledge from generalizing evidences
Predicate Logic
Propositional logic has a few limitations
Implication and Operators
Logic implications and relational operators work the same way in predicate logic as in propositional logic
Problem Solving, Planning and Decision Making
Problem Solving
Identify current state and the goal state
Explore problem space to find paths to move from current state to goal state
A problem space is a tree structure that shows all possible paths for state transition, states and operators
States are nodes in the tree
Operators (processes) transform a state and produce paths to other states
If the foal is far, define one or more sub-goals that can be reached from the current state
Strategies
Heuristics means a rule of thumb or a guide that assists us based on previous records or experience in making a decision
Does not guarantee optimal or best result
Difference Reduction or Means-end analysis
Other strategies: Brute-force, sequential or parallel search algorithms
Optimization strategies
Algorithms are used to find optimum or near optimum solutions
Decision Making
Heuristics in Making a Choice
Familiarity
Recognition
Equality
Best feature
Default
Imitation
Decision making may or may not include problem solving
Rational Decision Making
Apply rationality to search through the problem space based on facts, rules and heuristics
Unbounded Rationality: continue searching all possible paths - brute force search
Optimized search: apply optimization and other algorithms to narrow doesn't the search and avoid exploring unnecessary paths
Satisfying: Set a threshold of satisfactory level and stop when that level is reached
Ecological rationality: set the threshold of satisfaction or search criteria based on the situation
Planning
We do planning all the time for every single action we perform with consciously or unconsciously
Unconscious planning involves more heuristics and is based on experience
Conscious planning involves logic, rules, concepts, connections, heuristics and computation
Change Blindness
The effects are quite dramatic. A very brief time interval between two presentations of a picture make it almost impossible to notice a change unless you specifically pay attention to the part that changes
Simons examined our ability to recognize a change when we see a picture or a scene for a second
Experiment
A man stops a professor on campus and asks directions. Suddenly two men carrying a plywood sheet pass between the man and professor. After the men with the plywood pass, the professor continues with the conversation