PSC 100 - L6 - Object Recognition

apperceptive agnosia

Recognizing Objects

The ability to perceive an object’s features but not the object in its entirety

associative agnosia

Recognizing Objects

The ability to see an object but cannot name it

• Cannot link the visual input to visual knowledge

• Can draw well from memory

variations in stimulus input, contextual influences

Recognition: Some Early Considerations

The process of object recognition is complex, why?

bottom up processing

Recognition: Some Early Considerations

Processes that are directly shaped by the stimulus; data driven

top down processing

Recognition: Some Early Considerations

Processes shaped by knowledge; concept driven

input pattern, feature detectors, smaller features

Recognition: Some Early Considerations

The Importance of Features

Recognition begins with identifying visual features in the blank

• Evidence for blank in the visual system

• Larger units can then be detected by assembling the blank

visual search tasks, simple feature, combo of features

Recognition: Some Early Considerations

The Importance of Features

Tasks in which participants examine a display and judge whether a particular target is present

• Efficient when the target is defined by a blank - detected first

• Slow when a target is defined by a blank

feature detection

Word Recognition

Evidence suggests that object recognition begins with what?

tachistoscope

Word Recognition

Factors Influencing Recognition

A device used to present stimuli for precise amounts of time

repetition priming effect

Word Recognition

Factors Influencing Recognition

An effect that describes how words that were recently seen are better recognized. How does this relate to the results of a study that presented a brief presentation of stimuli, and used a post-stimulus mask that interrupted processing?

word superiority effect

Word Recognition

Where it’s easier to perceive & recognize letters in context than if they appear in isolation

well formedness

Word Recognition

How closely a letter sequence conforms to the typical patterns of spelling in the language

• E.g. HZYQ > FIKE > HIKE

easier, context effects

Word Recognition

Degree of Well Formedness

The more well-formed a letter sequence is, the blank it is to recognize the sequence and the greater the blank are produced by the sequence on recognition

regular

Word Recognition

Degree of Well Formedness

One problem of wellformedness is that people can perceive stimuli as being more blank than they actually are, producing errors

• E.g. DPUM is likely to be misread as DRUM

feature nets

Propose that recognition depends on a network of “detectors,” organized in layers of increasing granularity.

• Lowest layers focus on simple features (individual lines and angles)

• Bottom up flow across layers through combo of earlier-layer detectors allows recognition of more complex objects

• Feature detectors > letter detectors > bigram detectors > word detectors

activation level

Feature Nets & Word Recognition

Design of a Feature Net

Each detector in the network has a blank

• W/ input, this increases

• Some detectors will be easier to activate than others

response threshold

Feature Nets & Word Recognition

Design of a Feature Net

Detectors fire when their blank is reached

false, complex assemblies of neural tissue

Feature Nets & Word Recognition

Design of a Feature Net

True or false? Individual detectors are individual neurons/groups of neurons

recency

Feature Nets & Word Recognition

Design of a Feature Net

Starting activation levels depend on this. Detectors that have fired recently will have higher activation levels; a warm up effect

frequency

Feature Nets & Word Recognition

Design of a Feature Net

Starting activation levels depend on this. Detectors that have fired frequently will have higher activation levels; an exercise effect

well-primed detector

Feature Nets & Word Recognition

Ambiguous Inputs

A weak signal will likely be enough to trigger only a what?

frequency, recency

Feature Nets & Word Recognition

Ambiguous Inputs

Priming depends on blank and blank - network is biased to respond to inputs similarly to how it has responded previously

helps more than it hurts

Feature Nets & Word Recognition

Ambiguous Inputs

The bias to recognize frequent or primed words results in errors AND correct recognition. It (helps more than it hurts/hurts more than it helps)

distributed knowledge

Feature Nets & Word Recognition

Feature nets contain knowledge that is not locally represented. It is in a network that is reflected by r-ships across detectors. What is this called?

efficiency > accuracy

Feature Nets & Word Recognition

Efficiency vs Accuracy

The same mechanisms that enable the network to resolve ambiguous inputs and recover from errors can also result in recognition errors. What does this tell about the relationship between efficiency and accuracy?

McClelland & Rumelhart model

Descendants of the Feature Net

States that info flows bottom-up, top-down, and within the same level

• Higher level word detectors can influence lower-level detectors (top-down flow of info)

• Includes excitatory connections and inhibitory connections

recognition by components model

Descendants of the Feature Net

Applies the feature net model to recognition of 3D objects; includes an intermediate level of detectors that is sensitive to geons

• Feature detectors (curves, edges) > geon detectors > geon assemblies representing relations between geons > object model

geons

Descendants of the Feature Net

Recognition by Components Model

Basic shapes proposed as the building blocks for all 3D forms

• Can be identified form virtually any angle

• Most objects can be recognized from just a few of these (partial occlusion doesn’t necessarily prevent recognition)

object model

Descendants of the Feature Net

Recognition by Components Model

A representation of the complete, recognized object

IT cortex

Descendants of the Feature Net

Object Recognition & the Brain

Cells in the blank might be the biological foundation for word/object detectors

• Viewpoint-dependent cells may trigger viewpoint-independent cells

faces

Recognizing blank specifically seems to involve specialized neural structures

prosopagnosia

Face Recognition

Faces Are Special

An inability to recognize individual faces (including their own) despite otherwise normal vision

inversion

Face Recognition

Faces Are Special

Faces show a powerful blank effect (more errors are made the more it is oriented towards this)

fusiform face area, subtle distinctions

Face Recognition

Faces Are Special

This area is particularly responsive to faces, but high levels of activation can also be produced by tasks requiring blank

holistic perception

Face Recognition

In contrast to other object recognition, face recognition seems to depend more on blank, which is the perception of the overall configuration rather than an assemblage of parts

configurations, external knowledge

Top-Down Influences on Object Recognition

What are the limits of feature nets? (Explain)

sentence, context & expectations

Top-Down Influences on Object Recognition

The Benefits of Larger Contexts

While some top-down effects can be explained by feature nets (e.g. word superiority effect), others require more explanation. What are they? (Explain)