CGS 001 UCD Midterm 1

Chomsky

—

Church-Turing

—

Donald Broadbent

—

George Miller

—

Lashley

—

Marr (Vision)

—

Shepherd and Metzler

—

Terry Winograd

—

Tolman

—

Turing

—

How are the rules of transformational grammar examples of algorithms?

"the transformational principles of transformational grammar are examples of algorithms because they specify a set of procedures that operate upon a string of symbols to convert it into a different string of symbols" (17). - It uses the hierarchical structure to which Lashley drew our attention. This is characteristic of languages in general. They are hierarchically organized (19).

What is the hypothesis of task analysis?

(definition) The idea that we can understand a complex task (and the cognitive system performing it) by breaking it down into a hierarchy of more basic sub-tasks. - The hypothesis has proved a powerful tool for understanding many different aspects of mind and cognition. We can think about a particular cognitive system as carrying out a particular task—the task of allowing an organism to exploit previously acquired information. We can think about that task as involving a number of simpler, sub-tasks that can be carried out by even simpler sub-sub-tasks etc. as per the hierarchy.

What is the ventral route? What is the dorsal route?

(definition) The ventral route is the route that information relevant to recognizing and identifying objects takes, which goes from primary visual cortex to the temporal lobe which runs along the bottom of the brain. (definition) The dorsal route is the route that information travels when it is used to locate objects in space and it runs from the primary visual cortex to the posterior parietal lobe along the top of the brain.

What is computational modeling? Why is it important for cognitive science?

- Computational modeling is one of the principle tools that cognitive scientists have for studying the mind. One of the best ways to understand particular cognitive abilities and how they fit together is by constructing models that "fit" the data. - This is important for cognitive science because it thinks of cognition in terms of the neural pathways and channels along which information travels.

Describe each: the primal sketch, the 2.5-D sketch, and the 3-D sketch.

- The primal sketch makes explicit some basic types of information implicitly present in the retinal image. These include distributions of light intensity across the retinal image - areas of relative brightness or darkness, for example. It also aims to represent the basic geometry of the field of view. - The 2.5D sketch represents certain basic information for every point in the field of the viewer. It represents the point's distance from the observer. The 2.5D sketch is viewer-centered. It depends upon the viewer's particular vantage point. - The 3D sketch gives an observer independent representation of object shape and size.

What is a micro-world?

A micro-world is a virtual world with depictions of how a program can carry out various tasks within the world. - SHRDLU is capable of various tasks-through a robot arm-for which it can pick up blocks and pyramids and move them around. But, it only has the tools to talk about what is going on in the micro-world it "lives in." - The separate processing systems collaborate in solving information-processing problems. There is cross-talk between them, because the programs for each processing system allow it to consult other processing systems at particular moments in the computation.

What are some reasons why someone would think that cognitive systems are functional systems?

Although the actual physical structure of our brains are different from many other animals, the functions that the brains carry out like perception of pain and simple reasoning are similar. The brain has plasticity, which means it's always changing and so at any given point the physical structure that performs a certain function is not likely to stay that same structure in the future

What is a channel capacity?

Channel Capacity: where the channel capacity of an information channel is given by the amount of information it I can reliably transmit - The perceiver's capacity to make absolute judgements is an index of the channel capacity of the information channel it's using. - Absolute judgment: e.g.: ... naming a colour. Human subjects are limited in the absolute judgments they can make. - Relative judgment: e.g.: ... identifying which of the two colours is darker, or which of the two tones is higher in pitch. -

How is a connectionist network different to human brain?

Different: connectionist networks are different when compared to human brains because they utilize the idea of the back-propagation learning algorithm. Some forms of pattern recognition that are different to that of the human brain are predicting patterns in the movements of prices on the stock markets, valuing bonds, and forecasting demand for commodities. Brains learn the way they do because of how they are constructed—in particular because of the patterns of connectivity existing at each level of neural organization (between neurons, populations of neurons, neural systems, neural columns, and so forth)— in short, brains deteriorate.

What is it to represent knowledge procedurally?

Each component system is essentially made up of a vast number of procedures that work algorithmically to solve very specific problems. The system as a whole works because of how these procedures are lined-up and embedded within each other.

What does it mean to say that cognitive abilities degrade gracefully?

The factors listed below derive directly from the fact that minds are realized in brains. They exhibit gradual deterioration in performance over time. As we get older reaction times increase, motor responses slow down, and recall starts to become more problematic. The deterioration is gradual and usually imperceptible within small time frames. This is a function of how brains are wired, and of the biochemistry of individual neurons. The same holds true for how cognitive abilities emerge and develop. Brains learn the way they do because of how they are constructed-and in particular because of the patterns of connectivity existing at each level of neural organization. - Cognitive abilities and skills themselves evolve over time, developing out of more primitive abilities and giving rise to further cognitive abilities. Eventually they deteriorate and gradually fade out of existence. In some unfortunate cases they are drastically altered as a result of traumatic damage. This means that an account of the mind must be compatible with plausible accounts of how cognitive abilities emerge. It must be compatible with what we know about how cognitive abilities deteriorate. It must be compatible with what we know about the relation between damage to the brain and cognitive impairment.

Why might abstracting away from neural implementation to study cognitive science be a bad idea?

It might be a bad idea to abstract away from neural implementation because (1) temporal dimension of cognition. Cognitive activity needs to be coordinated with behaviour and adjusted on-line in response to perceptual input. The control of action and responsiveness to the environment requires cognitive systems with an exquisite sense of timing. The right answer is no use if it comes at the wrong time.

What is latent learning and what does it show about behaviorism (if anything)?

Latent Learning: is the idea that reinforcement is not necessary for learning and that it is a type of learning which is not apparent in the learner's behavior at the time of learning, but which manifests later when a suitable motivation and circumstances appear. (cancels out reinforcement theory as only reliable source of learning)

What do the Broadbent dichotic listening experiments show?

The Broadbent dichotic listening experiments show that when presented with three different stimuli (3 letters or digits in one ear and a different string of 3 letters or digits in the other ear) the subjects were asked to report the stimuli in any order and what they found was that they performed best when they reported the stimuli ear-by-ear. - This shows us that information comes through the senses and passes through a short-term store before passing a selective filter. The filter screens out a large portion of the incoming information, selecting some of it for further processing. Only information that makes it through the selective filter is semantically interpreted. - This shows that we can only attend to a single information channel at a time (assuming each ear is a separate information channel) and that section between information channels is based purely on physical characteristics of the signal.

What does (very generally) the Church-Turing thesis state?

The idea that anything that can be done in mathematics by any algorithm can be done by a Turing machine (Turing machines are computers that can compute anything that can be algorithmically computed).

What is the hypothesis of subconscious information processing?

The idea that much of what we do is under the control of planning and information-processing mechanisms that operate below the threshold of awareness. Even though we are often conscious of our high-level plans and goal (what goes on top of the hierarchy), we tend not to be aware of the information processing that translates to those plans and goals into action.

Describe the implementational level of analysis in cognitive science and give at least one example.

The task is to find a physical realization for the algorithm. - ie: To identify physical structures that will realize the representation states over which the algorithm is defined and to find mechanisms at the neural level that can properly be described as computing the algorithm in question. - e.g.: identify the neural structure on a general level in relation to the algorithm (the neural networks and neurons) - e.g.: identify the neural mechanisms in play in context of algorithm (the firing of neurons and the signals that occur)

What is chunking?

There is a way of working around this bottleneck... chunking. This is where we can reliable sequences of numbers with single numbers [ie: we pick out the same number in two different ways—with binary expression 1100100, or with the decimal expression 100 (where 0 represents the number 100 and thus chunking it from 7 to 3 numbers) if we use the decimal system we are within the limits where the binary systems allow us to be at capacity. - Natural language is the ultimate chunking tool.

Describe the computational level of analysis in cognitive science and give at least one example.

The computational level is the highest level. This is were scientists analyze in very general terms the particular type of task that system performs. It utilizes the "top-down" approach. ... A computational analysis identifies the information with which the cognitive system has to begin (the input to that system) and the information with which it needs to end up (the output from that system). ... - Tasks of analysis at the computational level are: to (1) translate a general description of the cognitive system into a specific account of the particular information-processing problem the the system is configured to solve, and to (2) identify the constrains that uphold any solution to that information-processing task. - e.g.: in vision, the basic task of the visual system is to derive a representation of the 3d shape and spatial arrangement of an object that will allow that object to be recognized.

What does transformational grammar tell us about the organization of cognitive ability?

This is characteristic of languages in general. They are hierarchically organized (19). When it comes to the organization of cognitive ability, transformational grammar tells us 1st, that a sophisticated, hierarchically organized, cognitive ability, such as speaking and understanding a language, involves stored bodies of information (about phase structures and transformation rules). And 2nd, that these bodies of information can be manipulated algorithmically (19).

Describe the algorithmic level of analysis in cognitive science and give at least one example.

This is one level below the computational level. This level tells us how the cognitive system solves the specific information-processing task identified at the computational level and how the input information is transformed into the output information. - e.g.: Information from the sensory systems about the distribution of light in the visual field is transformed into a representation of the three-dimensional environment around the perceiver. - The three famous examples are the 3 sketch levels. The Primary Sketch, 2.5D Sketch, and the 3D sketch. This is because it describes how basic input of light and other things like that eventually transform into our vision where can have a spatial sense of the object.

What is the main conclusion from Tolman's spatial learning experiments?

Tolman Concluded that place learning from his spatial learning experiments was the correct and evidenced experiment as seen with the rats in the 1st group turning right when entering from the south and turning left when entering from the north. Where the 2nd group of rats would enter from the south and make a right to find a reward and then be released from the north to find a reward on the west (turn right, but never left). He concluded that this was because the rats were recognizing their surroundings and not that of the sequence of movements they were making.

What is the distinction between a top-down and a bottom-up approach to cognitive science?

Top-Down Approach: starting out with general theories about the nature of through and the nature of cognition and working downwards to investigate how corresponding mechanisms might be instantiated in the brain. (Marr's theory of vision) ... Bottom-Up Approach: beginning with individual neurons and populations of neurons, or perhaps even lower down, with molecular pathways whose activities generate action potentials in. Individual neurons, and then trying to build up from that by a process of reverse engineering. (Ungerleider & Mishkin: which lobe processes the field of vision. Two visual systems hypothesis)

Joseph Weizenbaum

creates ELIZA - one of earliest "chatterbots"

Distinguish between the deep structure and surface structure of a sentence.

(definition) Deep Structure: this is the grammatical build-up of the sentence. how it is built up from basic constituents (syntactic categories. ie: noun, verb, adj., etc.). Called phase structure grammar. (definition) Surface Structure: this is literally what we see on the surface or the way the sentence is presented to us. The actual organization of words in a sentence, derived from the deep structure according to the principles of transformational grammar.

Is there anything correct about the behaviorist paradigm?

Something that is correct about the behaviourist paradigm is to reject the idea of mentalism. As well, they were also right that we can see conditioned learning in observed measurable behaviors (Tolman's rats and Skinner's box)

What is multiple realizability and why is it important to cognitive science? Provide at least one example.

(definition) Multiple Realizability: the same mental property, state, or event can be realized by different physical properties, states, or events. In the sense of functional systems, they function the same way and can be realized by multiple different physical structures. - It is important to cognitive science because it originally made scientists believe that studying the mind through the brain would be misleading because there are multiple physical structures that would perform the same function. - Eg. What makes something a heart? Hearts are organs that pumps blood around the body in particular, they collect deoxygenated blood and pump it towards the lungs where it becomes re-oxygenated. The actual physical structure of the heart is not particularly important and an artificial heart will do the job just as well and so it still counts as a heart.

What is syntax?

Syntax is the aspects of language use that have to do with how words can be legitimately put together to form sentences.

What are the major arguments/results that lead us to believe that behaviorism is false?

(1) Tolman and Honzik: Initial experiment with rats and maze w/three groups. First group rewarded from the beginning, second group not rewarded ever, third group rewarded after the first ten days and results were that the third group learnt to navigate the maze much faster than the first group of rats did. Surprising because there was no reward for the first ten days so they shouldn't have been learning maze? BUT they did and the result was latent learning was born (the idea that learning happens even when there is no reward being given) but then the question was what type of learning and so the next experiment came along. ... (2) Tolman, Ritchie, and Kalish were trying to understand whether the rats in the navigating maze experiment were learning through place learning (learning the spatial layout of the maze) or response learning (learning what sequence of movements they had to do to get to the end of the maze). So they did a second experiment where they had a cross maze with only one turn and there was cheese and rats. There were two groups of rats and each were placed on alternating sides of the maze each time they ran the experiment. For group 1, their reward (the cheese) was always placed on the west side of the maze, so on alternate trials, the rats had to take different turns based on where they started. This represented the rats learning the spatial layout of the maze (place learning). The second group of rats had their reward placed on alternating sides of the maze so they had to make the same turn each time they ran the maze. This represented response learning where the rats would learn to make the same turn each time to get their reward. The results of this experiment was that the rats in the place learning situation learned quicker, showing that organisms learn spatial layouts easier than response learning. However this was only specific to the problem solving and learning behavior of an organism and there was no broad theory to explain all complex behaviors. ... (3) The third guy, named Lashley introduced two main ideas that explained all complex behaviors of organisms in a hierarchical way. In his paper, there were two main theories that were learned from it: the hypothesis of subconscious information processing and the hypothesis of task analysis.

Why is SHRDLU an important development in cognitive science?

(1st) 1st it gave a powerful illustration of how abstract rules and principles such as those in the sort of grammar that we might find in the theoretical linguistics could be practically implemented. If we assume that a speaker's understanding of language is best understood as a body of knowledge, then SHRDLU provided a model of how that knowledge could be represented by a Cognitive system and how it could be integrated with other forms of knowledge about the environment. (2nd) 2nd it illustrated the general approach of trying to understand the model cognitive systems by braking down into distance components, each of which carries out a specific information-processing task. (3rd) 3rd it is based on the fundamentals assumption that understanding language is an algorithmic process.

What are the two conclusions that Marr drew from Warrington's experiments?

(1st) that information about the shape of an object must be processed separately from the information about what the object is for and what its is called. (2nd) that the visual system can deliver a specification of the shape of an object even when that object is not in any sense recognized. - Here Marr describing how he used these neuropsychological data to work out the basic functional task that the visual system performs.

What replaced behaviorism (if anything)? Describe.

(A)Initially, behaviorism was very popular and because the key ideas of behaviorism conflicted with cognitive psychology: the study of the mind, cognitive science could not be born while behaviorism was still a popular belief. In the 1900's, as behaviorism began to be questioned, a new way of thinking came about where people focused on studying the mind as an information-processing mechanism (machine) instead of purely a stimulus-response linking mechanism (machine). (B)One theory that replaced behaviorism was the idea of Computational Theory of the Mind. Meaning the idea of the Universal Turing Machine, where the idea that we are central processing units and our brains are computers that follow algorithms to follow through with this processing.

What is a cognitive map and how is it related to representation?

(a) Cognitive Maps are what Tolman described as evidence that animals form high-level representations of how their environment is laid out. (b) It is related to representation (stored information about the environment) because representations are one of the fundamentally explanatory tools of cognitive science. Cognitive scientists regularly explain particular cognitive achievements by modeling how the organism is using representations of the environment (ie: navigational achievements of rats in mazes) (11).

What is the distinction between place learning and response learning?

(a) Place learning (presumed correct by Tolman): the idea that cognitively able organisms code spatial information in terms of places rather than in terms of a particular sequence of movements. (b) Response learning (presumed proved incorrect by Tolman): the idea that animals code in a sequence of movements rather than their spatial surroundings.

What is the basic goal of transformational grammar and what does it tell us about cognition?

(a) The basic goal of transformational grammar is to explain the connection between sentences of the first type and to explain the differences of the second type. This is done by giving principles that state the acceptable ways of transforming deep structures (17). Look at the language part of cognition as a set of algorithms. E.g.: - John hit the ball & the ball was hit by John. When the surface structure is not the same but the deep structure remains the same, the meanings can be the same but they "appear" different. - Jean is eager to please & Jean is easy to please. They "appear" the same, but the deep structure is different so they have different meanings.

What is an unconditioned stimulus? What is a conditioned stimulus?

(a) Unconditioned stimulus: is a form of stimulus that is not neutral for the organism and typically provokes a behavioral response, such as salvation seen in the famous example of Pavlov's dogs. Would unconditionally produce the response even in the absence of the conditioned stimulus (food) (b) Conditioned stimulus (?): a previously neutral stimulus that, after becoming associated with the unconditioned stimulus, eventually comes to trigger a conditioned response. (bell)

What is a "Skinner Box" and how is it relevant to behaviorism?

(a)A Skinner Box is a box where the rat has a response lever controlling the delivery of food. The rat receives a reward for behaving in a particular way (pressing a lever or pushing a button). Each time the rat performs the relevant behaviour, it receives the reward. The reward reinforces the behaviour. This means that the association between the behaviour and the reward is strengthened and that the rat's performing the behavior again becomes more likely. (b)It is relevant to behaviorism because it utilizes this theory of association and classical conditioning. The rat becomes conditioned to perform the behaviour. This is an observable response and thus inherently a part of behaviourism (all behaviours are learned with the interaction with the environment) and that behaviours can be taught using reinforcement.

What is conditioning?

(definition) Conditioning: the association between the behavior and the reward is strengthened and the rat's performing the behavior again becomes more likely (sub)Classical Conditioning: the strengthening of the association between conditioned stimulus and unconditioned stimulus eventually leads the organism to produce the unconditioned response to the conditioned stimulus alone, without the presence of the conditioned stimulus.

What is a Turing machine? Be sure you are able to describe how it works and label a diagram.

A Turing machine is a machine that's behaviour is entirely determined by the table (can be represented as a sequence of numbers), its current state (M or q1), and the symbol in the cell (square) it is currently scanning there is no ambiguity and no room for the machine to exercise "intuition" or "judgment."

What is a Universal Turing machine?

A UTM is a Turing machine that can run any specialized Turing machine. It can take an input of a program specifying any given specialized Turing program. It is the theoretical precursor to modern-day general-purpose computers.

What is a connectionist or artificial neural network?

A connectionist or artificial neural network is proposed and pursued of a new set of abstract mathematical tools for modeling cognitive processes. It abstracts away from many biological details of neural functioning in the hope of capturing some of the crucial general principles governing the way the brain works. Most artificial neural networks are not biologically possible in anything but the most general sense. What makes them so significant is that they give cognitive scientists a bridge between algorithm and implementation. These algorithms work by changing the weights of the connections between pairs of neurons in adjacent layers in order to reduce the "mistakes" that the network makes.

What is an algorithm?

An algorithm is a finite set of rules that are unambiguous (not open to more than one interpretation) and that can be a[plied systemically to an object or set of objects to transform it or them in definite and circumscribed ways.

Describe a connectionist network generally (including descriptions of nodes and levels).

An artificial neural network contains a large number of units-nodes-(which might be thought of as artificial neurons). Each unit has a varying level of activation, typically represented by a real number between -1 & 1. The units are organized into layers with the activation value of a given layer determined by the activation values of all the individual units. The simultaneous activation of these units and the consequent spread of activation through the layers of the network, governs how information is processed within the network. The basic architecture of the network is clearly illustrated within the diagram (in (4)). The network is composed of a set of processing units organized into three different layers (input layer, hidden layer, and output layer). The first layer is made up of input units, which receive inputs from sources outside of the network. The third layer is made up of output units, which send signals outside the network. The middle layer is composed of what are called hidden units. Hidden units are distinctive by virtue of communicating only with units within the network. They are key to the computational power of ANN. Networks without hidden units are only capable of carrying out a limited variety of computational tasks.

What is an information channel?

An information channel is a medium that transmits information sender to a receiver. (20).

What is association?

Association: refers to a mental connection between concepts, events, or mental states that usually stems from specific experiences. A connection or relationship between two items with the result that experiencing the first item activates a representation of the second.

What is back-propagation?

Back-Propagation: this is the process of training. It begins with a random assignation of weights and then presents the network with a training series of input patterns of activation, each of which is associated with a target output pattern of activation. Differences between the actual output pattern and the target output pattern result in changes in weights. This process continues until errors have diminished to almost zero, resulting in a distinctive & stable pattern of weights across the network.

Why is the Two Visual Systems hypothesis important for cognitive science?

Important because of the tools that were used to arrive at it (e.g the study of brain damaged patients and experiments on monkeys) and because it illustrated a bottom-up, as opposed to top-down, way of studying the mind.

What is an information-processing bottleneck?

Information-Processing Bottleneck: where the human perceptual systems are information channels with built-in limits. They can only process around 7 limits at a time (idea founded by Miller).

Why would it be accurate to say that chatterbots do not understand language?

It be accurate to say that chatterbots do not understand language because chatterbots like ELIZA do not analyze the syntactic structure or the meaning of the sentences it encounters. It is simply programed to respond to certain cues by making one of a small set of responses and it cannot report on or navigate its environment.

What is the function (general) of the ventral pathway?

It carries information relevant to object identification (the "what" pathway)

What is the function (general) of the dorsal pathway?

It carries information relevant to object location (the "where" pathway)

What is a cross-lesion disconnection experiment?

It is a methodology that is designed to trace and uncover pathways through which information flows - Also addresses the problem about making inferences about a certain region's function through looking at what the organism is unable to do when that region is damaged - Because the information flows through a pathway and is not performed by a singular brain part it is not fair to assume that a particular damaged brain part is correlated with a damaged behavior because the behavior could be caused by a separate part that is functioning perfectly but cannot function properly because it is missing input from the damaged part of the brain

How is SHRDLU an advancement over a chatterbot?

It is an advancement to the chatterbot because it was one of the first attempts to write a program that is not just trying to simulate conversation, but is capable of using language to report on its environment, to plan actions, into reason about the implications of what is being said to it.

What is parallel processing? Give an example.

Parallel Processing: The processing is parallel because the flow of information through the network is determined by what happens in all of the units in a given layer—but none of those units are connected to each other. - E.g. cognitive model (derived primarily from experiments on normal subjects, rather than from studies of brain-damaged patients).

What is reinforcement?

Reinforcement: a reward for correct behaviour that refers to any stimulus which strengthens or increases the probability of a specific response.

What is reverse engineering?

Reverse engineering is the process by which one takes an object and tries to work backwards from its structure and function to its basic design principles.

What view does SHRDLU provide of linguistic understanding?

SHRDLU illustrates a view of linguistic understanding as resulting from the interaction of many, independently specifiable cognitive processes. It combines the syntactic system, semantic system, and cognitive-deductive system (read more 34).

What is selective attention? Describe one such example.

Selective hearing is the idea that we only attend to some of what we hear. - e.g.: when we attend a cocktail party, we can often hear many ongoing and unrelated conversations and we manage to focus on only the one we want to listen to.

What is serial processing? Give an example.

Serial Processing: Serial Processing: processing in physical symbol systems step-by-step. Information travels through a fixed series of information-processing "stations" in a fixed order. ??????? - E.g. neurological model (visual information about the word's appearance needs to be phonologically recoded before it can undergo further processing) - E.g. The Turing machine can only read one cell at a time

How is a connectionist network similar to human brain?

Similar: connectionist networks are similar to human brains because it is similar to the essence of a neural network... pattern recognition. Many different types of cognitive ability count as forms of pattern recognition and the tools provided by ANN have been used to model a range of cognitive processes. Because they can be trained, they can be used to model how cognitive abilities are acquired. And, like human brains, they are not "all-or-nothing" - even when damaged they can continue to perform, albeit in a limited way (unlike digital computers, which function either optimally or not at all).

What can cognitive scientists learn from behaviorism?

We can learn that one type of learning is caused by conditioning. This entailed both conditioned stimulus and unconditioned stimulus. (see A6 for explanation for both).

How are sensory modalities (like vision) akin to information channels?

We can think of perceptual systems as information channels. Vision, for example is a medium through which information is transmitted from the environment to he perceiver. So are audition (hearing) and olfaction (smell). Thinking about perceptual tools in this way gave Miller and other psychologist a a new set of tools for thinking about experiments on humor perception.

What is weighting in a connectionist network?

Weighting in a connectionist network: The strength of the connections between individual neurons varies and is modifiable through learning. This means that there can be several distinct neural networks each computing a different function even through each is composed of the same number of units organized into the same set of layer and with the same connections holding between those units.