π Lecture 20 β Attention & Pattern Recognition
π Overview: This lecture explores the critical role of attention in combining features to perceive patterns, and how context and prior knowledge influence pattern recognition. It explains the difference between single feature detection and feature conjunction, introduces top-down processing, and discusses the Word Superiority Effect, demonstrating that word context can enhance letter perception.
ποΈ Topics Covered
The lecture covers the role of attention in feature integration through Treisman & Gelade's experiment on detecting a T among distractors, demonstrating that feature conjunction takes longer than single feature detection. It then introduces context and pattern recognition using ambiguous figures, explains top-down processing as high-level knowledge guiding perception, and concludes with the Word Superiority Effect and its explanation by Rumelhart & Siple regarding inferential perception.
π Lecture Summary
ATTENTION & PATTERN RECOGNITION
Attention is required to combine features to perceive patterns. Treisman & Gelade (1980) conducted an experiment where subjects tried to detect a T in an array of 30 I's and Y's. Subjects could detect this by looking for the single cross-bar feature of the T, taking about 800 milliseconds. When subjects had to detect a T in an array of 30 I's and Z's, they could not use a single feature; they had to look for the conjunction of features (both the vertical and horizontal bars), taking more than 1200 milliseconds. This difference of about 400 milliseconds shows that feature conjunction requires more attention than single feature detection. The researchers varied display size: subjects showed little difference between conditions for displays with fewer than five letters, but with more distracters, attention became substantially overloaded. This demonstrates that for familiar letters, the deficit in perceiving feature conjunctions only appears with large displays. π‘ Why this matters: This explains why we can automatically recognize single features but need focused attention to combine features in complex visual scenes.
Context & Pattern Recognition
The lecture presents an example of ambiguous figures where the same picture is perceived as an animal when placed in a line of animal pictures (hen, rabbit, dog, cat) and as a human when placed in a line of human pictures (man, woman, child, girl). This demonstrates that context influences how we perceive patternsβthe identical image is interpreted differently based on surrounding stimuli.
Top-Down Processing
When context or general world knowledge guides perception, this is called top-down processing, because high-level general knowledge determines the interpretation of low-level perceptual units. An example is the phrase "THE CAT" where the middle letter in the word can be seen as either A or H depending on the surrounding letters. This shows that the surrounding words have an effect on our perception of individual letters.
π Definition β Top-Down Processing: When high-level general knowledge and context guide the interpretation of low-level perceptual units.
π Example: In "THE CAT", the middle letter is ambiguousβit can be read as 'A' in "CAT" or 'H' in "THE"βshowing that the word context determines which letter we perceive.
Word Superiority Effect
The Word Superiority Effect (WSE) was demonstrated by Reicher and Wheeler (1970). They presented a brief display of either a letter (e.g., D) or a word (e.g., WORD). Immediately afterward, subjects were given a pair of alternatives and instructed to report which they had seen. If shown D, alternatives might be D or K; if shown WORD, alternatives might be WORD or WORK. Subjects were about 10% better in the word conditionβthey more accurately discriminated between D and K in the context of a word than as letters alone.
Why WSE?
Rumelhart & Siple (1974) explained why subjects are more accurate in the word condition. Suppose subjects identify the first three letters as WOR. There are several four-letter words consistent with WOR beginning: WORK, WORD, WORM, WORN, WORT. If subjects detect only the bottom curve in the fourth letter, in the word context they need only detect one feature to perceive the fourth letter (e.g., the curve of D distinguishes it from K). However, when the letter is presented alone and subjects detect the curve, they will not know whether the letter was B, D, O, or Q, since each is consistent with the curve feature. When the letter is presented alone, they must identify a number of features. This analysis implies that perception is inferentialβlike the curve in D helps in recognition only when combined with word context.
π Definition β Word Superiority Effect (WSE): The phenomenon where letters are more accurately recognized when presented in the context of a word than when presented alone.
π Formula/Mechanism: Word context β fewer features needed to identify a letter (e.g., one feature in WORD) vs. letter alone β multiple features needed (e.g., curve in D could be B, D, O, or Q).
π Example: Subjects correctly identified D vs. K about 10% more often when D appeared in "WORD" (as opposed to "WORK") than when D was presented alone, because the "WOR" context limited possibilities.
β Key Takeaways
The critical takeaway is that attention is essential for combining single features into complex patterns, as shown by Treisman & Gelade's experiment where feature conjunction detection took 400ms longer and was more affected by display size than single feature detection. Context powerfully shapes perception, as demonstrated by ambiguous figures and the top-down processing principle, where high-level knowledge determines interpretation of low-level units. The Word Superiority Effect shows that word context enhances letter recognition by about 10%, because perception is inferentialβa word context reduces the number of features needed for identification compared to isolated presentation. Students must understand that these effects reveal perception as an active, constructive process rather than a passive registration of sensory input.
π§ Quick Revision Questions
- What was the difference in reaction time between detecting a T by a single feature (among I's and Y's) versus detecting a T by feature conjunction (among I's and Z's) in Treisman & Gelade's experiment?
- Why did display size affect subjects' performance more in the conjunction condition than in the single feature condition?
- What is top-down processing, and how does the "THE CAT" example illustrate it?
- What is the Word Superiority Effect, and how much better were subjects in the word condition compared to the letter-alone condition?
- According to Rumelhart & Siple, why does word context make letter recognition easier than recognizing a letter alone?
π Lecture 21 β Pattern Recognition (Continued)
π Overview: This lecture examines neural networks as a model for pattern recognition, focusing on McClelland and Rumelhart's interactive activation model. It explains how parallel distributed processing (PDP) resolves the paradox of bottom-up versus top-down processing, demonstrating how features, letters, and words interact through excitatory and inhibitory connections to facilitate word recognition.
ποΈ Topics Covered
The lecture covers the interactive activation model by McClelland and Rumelhart, which solves the bottom-up versus top-down paradox in pattern recognition. It explains neural network components including nodes, links, excitatory and inhibitory connections, weights, activation rules, and learning rules. The lecture also discusses how PDP models have improved computer functioning and how word superiority effect emerges from network interactions.
π Lecture Summary
Pattern Recognition (Continued)
This model is also called PDP's or Parallel Distributed Processing. McClelland and Rumelhart (1981) made a pattern recognition network that solves the paradox of bottom-up versus top-down processing. They implemented this network to model our use of word structure to facilitate recognition of individual letters. In this model, individual features are combined to form letters and individual letters are combined to form words. This is a connectionist model that depends heavily on excitatory and inhibitory activation processes. Activation spreads from the features to excite the letters, and from the letters to excite the words. Alternative letters and words inhibit each other. Activation can also spread down from the words to excite the component letters, so a word can support the activation of a letter and promote its recognition. In such a system, activation will tend to accumulate at one word and repress the activations of other words through inhibition. The dominant word will support the activation of its component letters, and these letters will repress the activation of alternative letters. The word superiority effect is due to the support a word gives to its component letters. The computation proposed by McClelland and Rumelhart's interactive activation model is extremely complex, as is the computation of any model that stimulates neural processing. This process helps us in understanding how neural processing underlies pattern recognition.
The network diagram shows a simpler version of a neural network displaying the word WORK. At one level there are feature detectors (not shown). At another level there are letter detectors. We recognize W because of features; O, R, and K are recognized similarly. All words activate because of letters. Four words of WORK are activated β these three words are competitors in word recognition. WORK is recognized by four letters. This is a parallel distributor processing model, also called a neural network model of pattern recognition. It is called a neural network because there is an abstract concept or quantity called nodes.
π Definition β Connectionist model: A model of pattern recognition where individual features combine to form letters, and letters combine to form words, using excitatory and inhibitory activation processes. π Definition β Word superiority effect: The phenomenon where a word supports the activation of its component letters, making letter recognition easier within a word context than in isolation. π Example: In the word WORK, the word-level activation supports recognition of each individual letter (W, O, R, K) while inhibiting alternative letters at the same positions.
Neural Networks
Neural networks consist of: Nodes, Links (both excitatory and inhibitory), Weights, and Learning which consists of re-adjustment of weights.
Nodes are a set of processing units. Nodes should not be confused with neurons. Nodes are represented by features, letters, and words in the interactive activation model. They can acquire different levels of activation. All boxes in the diagram are nodes; lines are links. Nodes are connected through these lines.
Patterns of connections: Nodes are connected to each other by excitatory or inhibitory connections that differ in strength. Another important concept is activation rules, which specify how a node combines its excitatory and inhibitory inputs with its current state of activation.
π Definition β Nodes: A set of processing units in a neural network that can acquire different levels of activation; they are represented by features, letters, and words in the interactive activation model.
Excitatory connections are those connections that make other nodes active. Nodes connected with excitatory connections are active or charged. Inhibitory connections are those connections that make other nodes relax and switch off. Because of these connections the neural network exists.
State of Activation: Nodes can be activated to various degrees. We become conscious of nodes that are activated above a threshold level of conscious awareness. We become aware of letter K in the word WORK when it receives enough excitatory influences from feature and word levels.
A Learning Rule: Learning generally occurs by changing the weights of the excitatory and inhibitory connections between the nodes. The Learning rule specifies how to make these changes in the weights. Two aspects are initial weights and re-adjustment of weights.
π Definition β Excitatory connections: Connections between nodes that make other nodes active or charged. π Definition β Inhibitory connections: Connections between nodes that make other nodes relax and switch off. π Definition β Activation rules: Rules that specify how a node combines its excitatory and inhibitory inputs with its current state of activation. π Definition β Learning rule: A rule that specifies how to change the weights of excitatory and inhibitory connections between nodes to enable network improvement. π‘ Why this matters: The learning component is the most important feature of a neural network model because it enables the network to improve its performance. In a lab in California, a computer learned how to speak by reading and re-reading simple English sentences β improving from its own mistakes.
PDP and its significance
Parallel processing models have improved computer functioning, leading to super computers, also called parallel computers. Multiple processors that communicate with each other work faster than serial processing computers. The paradoxes are resolved β like forests being seen at the same time as the trees, words are seen at the same time as the letters. Context helps in object perception; object perception helps in perception of context.
β Key Takeaways
The interactive activation model by McClelland and Rumelhart resolves the bottom-up versus top-down paradox by allowing activation to spread bidirectionally between features, letters, and words. Neural network components include nodes (processing units), excitatory connections (which activate other nodes), inhibitory connections (which suppress other nodes), and weights that determine connection strength. The learning rule enables networks to improve performance by re-adjusting weights based on experience. The word superiority effect demonstrates how word-level information supports letter recognition through top-down activation. PDP models have led to parallel computers that process information faster than serial processing systems.
π§ Quick Revision Questions
- What are the three levels of processing units in McClelland and Rumelhart's interactive activation model for word recognition?
- How do excitatory and inhibitory connections differ in their function within a neural network?
- What is the word superiority effect and how does the interactive activation model explain it?
- What is the role of the learning rule in neural network models, and what does it specify?
- How does parallel distributed processing resolve the paradox of bottom-up versus top-down processing?
π Lecture 22 β Pattern Recognition (Continued)
π Overview: This lecture explores how sentence context and speech context influence word recognition and pattern perception. Through classic experiments by Tulving, Mandler & Baumal (1964) and Warren (1970), it demonstrates that prior context reduces the amount of bottom-up information needed to identify words, and even allows listeners to "fill in" missing phonemes, showing that perception is actively shaped by top-down expectations.
ποΈ Topics Covered
The lecture covers the effect of sentence context on word identification through the Tulving, Mandler & Baumal (1964) experiment, the interaction between bottom-up information and context, the implications of context for filling in missing words, the phoneme restoration effect demonstrated by Warren (1970), and the Warren & Warren (1970) follow-up study showing how subsequent words determine identification of ambiguous speech sounds.
π Lecture Summary
Effects of Sentence Context
Cognitive psychologists study whether sentence context affects word recognition. An experiment by Tulving, Mandler & Baumal (1964) demonstrated this effect at the multiword level. They used sentences like "Countries in the United Nations form a military alliance" and "The huge slum was filled with dirt and disorder." Each sentence provided an eight-word context preceding a critical word. In various conditions, subjects saw 0, 4, or 8 words of context before the critical word (e.g., "disorder") was presented for a very brief duration ranging from 0 to 140 milliseconds. The experimenters manipulated both the amount of context and the exposure duration to study how bottom-up information interacted with context.
π Definition β Bottom-up information: Information derived directly from sensory input (e.g., visual features of letters or acoustic features of speech). π Definition β Top-down processing: The use of prior knowledge, expectations, or context to influence perception.
π Formula: Probability of correct identification = f(context amount, exposure duration) β The chance of correctly identifying a word increases both when more context is provided and when the word is shown for longer.
π Example: Results showed: At 0ms flash duration, 0 context = 0% correct, 4 context = 10%, 8 context = 16%. At 140ms flash duration, 0 context = 70%, 4 context = 80%, 8 context = 98%. At 60ms flash duration, 0 context = 30%, 4 context = 60%, 8 context = 70%. The maximum effect of context was seen at 60ms exposure. The effect diminished somewhat between 60 and 140ms because subjects in the 8-word context condition performed almost perfectly and showed little benefit from further exposure, whereas subjects in the 0-word condition continued to benefit from longer exposure. These results indicate that subjects can take advantage of context to improve their identification of words.
Implications
This experiment shows that we can use sentence context to help identify words. With context, we need to extract less information from the word itself in order to identify it. We can also use context to fill in words that didn't even occur. We are able to fill in missing letters as we read the sentence, perhaps not even noticing they were missing. PDP (Parallel Distributed Processing) models can help explain this effect of context better than other models while also accounting for feature analysis.
π‘ Why this matters: Context reduces the sensory evidence required for recognition, meaning perception is not purely data-driven but also expectation-driven.
Context and Speech
Text works this way, but does it work with speech? Speech is experienced sequentially in a more linear fashion than text. The Phoneme Restoration Effect was demonstrated in an experiment by Warren (1970). He had subjects listen to the sentence: "The state governors met with their respective legislatures convening in the capital city." A 120 ms pure tone replaced the middle 's' in "legislatures." Only 1 in 20 subjects reported hearing the pure tone, and even that person wasn't able to locate it clearly.
π Definition β Phoneme Restoration Effect: A phenomenon in speech perception where listeners "fill in" a missing or replaced phoneme based on context, often not noticing the manipulation.
π Example: Warren & Warren (1970) presented subjects with sentences such as:
- "It was found that the *eel was on the axle" β subjects reported hearing "wheel"
- "It was found that the *eel was on the shoe" β subjects reported hearing "heel"
- "It was found that the *eel was on the orange" β subjects reported hearing "peel"
- "It was found that the *eel was on the table" β subjects reported hearing "meal" In each case, the asterisk denotes a phoneme replaced by non-speech. The identification of the critical word was determined by subsequent context.
Implications
The implications of the phoneme restoration experiments are: (1) Context fills in gaps and affects our perception just as in texts. (2) The identification of the critical word is determined by what comes after the critical word (e.g., "heel," "peel," "meal," "wheel" are critical words). (3) Thus, the identification of words can depend on the perception of subsequent words. In a nutshell, when you face a problem, you should grasp the context. When you grasp the context, you are able to understand and handle the problem.
β Key Takeaways
Context powerfully influences both visual word recognition and speech perception by reducing the amount of bottom-up sensory information needed for identification. The Tulving, Mandler & Baumal experiment shows that increasing sentence context and exposure duration both improve word identification, with the largest contextual benefit occurring at intermediate exposure durations. The phoneme restoration effect demonstrates that listeners use context to perceptually fill in missing or distorted speech sounds, often without noticing the manipulation. Crucially, word identification can depend on information from words that come after the target word, showing perception is actively constructed from both prior and subsequent context. PDP models provide a useful framework for explaining how top-down context and bottom-up feature analysis interact in pattern recognition.
π§ Quick Revision Questions
- What was the main independent variable manipulated in the Tulving, Mandler & Baumal (1964) experiment besides exposure duration?
- At which exposure duration was the maximum effect of sentence context observed on word identification?
- What is the phoneme restoration effect, and what does it reveal about speech perception?
- In the Warren & Warren (1970) experiment, what determined whether subjects heard "wheel," "heel," "peel," or "meal"?
- How do PDP models help explain the effect of context on pattern recognition compared to other models?