Acquire information about the world around us (perception)
Integrate that information with prior knowledge from our stored memory
Store that knowledge in our memory so we can use it later to help us achieve our goals
First step in this process of acquiring knowledge about the world involves sensation and perception
Sensation: process by which our sensory receptors and nervous system receive stimulus energies from the environment and transduce them into neural impulses (transduction). The inherent stimuli. Objective
Perception: process of interpreting and organizing sensory information through use of previous knowledge. What gives stimuli meaning. Subjective.
The image is then transformed into a 3-D sketch in which the the axes of symmetry and elongation link the object parts
Symmetry axis: line that divides an object into mirror image halves
Elongation axis: line defining direction along which main bulk or mass of a shape is distributed
The 3-D sketch is object-centered – the object’s parts are described relative to one another and are linked on the basis of shared properties and axes
This solves the object constancy problem, allowing recognition of an object presented in different orientations and under different conditions, e.g., lighting changes
Alternative modes of perception: Mindfulness is largely about seeing the “suchness” of things, that is, seeing things directly without conceptual filters
What assumptions might you make about this woman if you were told she is from New England? from California?
Our preconceived notions prevent us from seeing the real person in front of us
Stereotyping
If the doors of perception were cleansed, everything would appear to man as it is, infinite.
To see a World in a Grain of Sand, And a Heaven in a Wildflower, Hold Infinity in the palm of your hand, And Eternity in an hour. – Blake
Many different neurons connect to the dendrites of each neuron
Some produce excitatory effect; others produce inhibitory effect
There are also different levels of intensity of these effects
Around a thousand connections are connected to each neurons
If the activation of the neuron reaches a certain minimum threshold, the neuron will fire
16A Notes
Because circuit analysis translates to a wide range of fields, we can model many physical systems as electrical circuits, often gaining insight about the system. You may have heard of neural networks, an important machine learning tool that can be used to “learn” tasks such as image and voice recognition from examples instead of explicit programming. Neural networks are modeled after biological neural networks, which are fundamentally circuits operating on electrical signals within a brain:
In a general sense, studying circuits provides you with the conceptual and mathematical tools needed to analyze such networks. More broadly, circuit concepts are relevant to understanding network analysis and signal flows in systems, which can be applied to areas ranging from transportation analysis to social network analysis. (
from EECS16A Note0)
Artificial neural networks (ANN)
The nodes or neurons are organized into layers in much the same way that human neural networks are
The weights attached to the connections between pairs of units in adjacent layers determine the overall behavior of the network
This is similar to the way in which excitatory and inhibitory neurons of various strengths connect to a particular neuron in human neural networks
The bias term indicates what the weighted sum needs to be before the node/neuron will activate
This is similar to the threshold necessary for activation of a neuron in human neural networks
An artificial neural network is an interconnected group of nodes, inspired by a simplification of neurons in a brain. Here, each circular node represents an
artificial neuron and an arrow represents a connection from the output of one artificial neuron to the input of another.
Ex: How might a computer recognize a “9” using neural networks?
#
There is huge variety of ways in which people write 9’s
To simplify things, we can represent the “9” by decomposing it to a grid of 28 x 28 pixels of varying shades of gray (between 0 and 1)
First (input) layer of network
Starts with bunch of neurons or nodes corresponding to an array of 28 x 28 pixels in the image
Each node holds a number that represents the grayscale value of the corresponding pixel, ranging from 0 for black to 1 for white
This is the neuron’s activation level
Activations in one layer bring about activations in the next layer, which in turn bring about activations in the next layer…
This is loosely analogous to how, in biological networks of neuron, some groups of neurons cause other neurons to fire
Second layer (or first “hidden layer”)
Each neuron in the second layer might pick up on whether there is an edge in one particular region
You assign a weight to each one of the connections between a particular neuron in the second layer and the neurons in the first layer
Then you take all the activations from the first layer and compute their weighted sum according to the weights
Could make the weights associated with almost all of the pixels 0 except for some positive weights in target region
To really pick up on whether there is an edge here, could also have some negative weights associated with the surrounding pixels
Sum is largest when those middle pixels are bright but surrounding pixels are darker
But maybe you don’t want the neuron to light up anytime the sum is bigger than zero – maybe you only want it to be active when the sum is bigger than say 10
So you add in some other number (the bias), like -10, to the weighted sum
The bias tells you how high the weighted sum needs to be before the neuron starts getting meaningfully active
The connections between the other layers also have weights and biases associated with them
Third layer (or second “hidden layer”)
When we recognize digits, we piece together various components
e.x: A “9” has a loop near the top and a line on the right whereas an “8” has a loop on the top and one below
Each neuron in the third layer corresponds to one of these subcomponents
e.x: A particular neuron in the third layer might be activated by any generally loopy pattern toward the top
These subcomponents are made up of the various edges from the second layer
Last (output) layer
Has 10 neurons, each representing one of the digits
The activation in these neurons – some number between 0 and 1 – represents how much the system thinks a given image corresponds with a given digit
Learning is about getting the computer to find a setting for all of the different weights and biases so that it will actually solve the problem at hand
ANNs can compute any function that can be computed by a digital computer
However, it was not until the emergence of backpropagation learning algorithm that it became possible to train multilayer neural networks
The strength or weight of the connections between neurons in adjacent layers varies: neural networks learn by modifying these weights
Learning algorithms that are programmed into the ANN change the weights of the connections between pairs of neurons in adjacent layers in order to reduce the “mistakes” that the network makes
The basic idea is that each hidden unit connected to an output unit bears a degree of “responsibility” for the error of that output unit
If the activation level of an output unit is too low, then the weight between the output unit and each hidden unit connected to it is increased to decrease the error
The network then assigns error levels to the next layer of hidden units, so the error is propagated back down through the network until the input layer is reached
tl;dr: neural networks have to ’learn’ by adjusting stimuli weights. When a network gets an answer wrong, it has to recursively pop back each layer and adjust the corresponding weights (increasing correct value weights, decreasing activated incorrect value weights)
Q: How many neurons should there be in each hidden layer? A: There are a number of empirically-derived rules-of-thumb. Of these, the most commonly relied on is “the optimal size of the hidden layer is usually between the size of the input and size of the output layers”
Q: How many hidden layers are needed? Are more layers better? A: No. Situations in which performance improves with additional hidden layers are very few. One hidden layer is sufficient most of the time.
Q: Why are more hidden layers not necessarily better? A: Increasing the number of hidden layers much more than the sufficient number will cause the network to overfit the training set. It will learn the training data, but it won’t be able to generalize to new unseen data.
Limitations of models of object perception discussed above: assumes, in theory, perception is objective and accurate; in real life, that is often not the case…
What we perceive, the way we perceive, is not always what would be predicted by these models
Our concepts, expectations, and beliefs play a much bigger role in perception than we usually realize
Perception engages both top-down and bottom-up processing
Bottom-up processing: analysis of information coming from stimuli through sensory receptors
Object perception as combination of stimulus information from sensory receptors
Emphasizes the importance of information coming from the outside world
Top-down processing: information processing guided by higher-level processes, such as our beliefs, expectations, and memories
Our knowledge, beliefs about the world inform our perceptions
Emphasizes the importance of information coming from our minds
Note: we use both throughout everyday situations, rarely exclusively either or
Models can almost be sort of categorized/grouped by which of the two processing model they put emphasis on
E.x. you see a water bottle on your desk
You know it’s a water bottle since it’s physically there. You can see, and perhaps touch it
You know it’s a water bottle because of it’s features which have meaning to you through living in the modern world (perhaps owning a water bottle yourself!)
If you were an alien, or a homosapien from a very long time ago, you wouldn’t know that the object was a water bottle (among other things…)
Optional: Deductive versus Inductive Reasoning
Deductive reasoning works from the more general to the more specific, i.e., “top-down” approach.
In deductive reasoning there is usually a first premise, then a second premise (both of which are proven through observations), and finally an inference.
Ex: All men are mortal. Smerdley is a man. Therefore, Smerdley is mortal.
Inductive reasoning works the other way, moving from specific observations to broader generalizations and theories, i.e., “bottom-up” approach
Inductive reasoning extracts a likely (but not certain) premise from specific and limited observations.
Ex: I have a bag of many coins, and I’ve pulled 10 at random and they’ve all been pennies, therefore this is probably a bag full of pennies.
“Objective reality” is often not as objective as we think…
#
A fool sees not the same tree that a wise man sees. — William Blake
Self-fulfilling prophecies: People generally think that it is our experiences and perceptions that create our beliefs, but often, it is actually our beliefs that create our experiences and perceptions
Our beliefs and expectations influence others’ behavior
The Pygmalion effect: study found that students who were (randomly) labeled intellectual “spurters” showed significantly greater gains in IQ and academic performance after 8 months than controls
Follow-up: If teacher believed that girls learn to read faster than boys, they did
Children who were told they were neat and tidy became more neat and tidy than those who were told they should be neat and tidy
Follow-up: children who are told that they are good at math showed greater improvements in math scores than those who were told that they should try to become good at math
Those who over-idealize romantic partners as having many virtues and few faults tend to have happier and longer-lasting relationships (Miller, Niehuis, & Huston, 2006)
Moreover, the partners who are over-idealized tended to develop those traits over time! (Sandra Murray)
People live up to their expectations – we tend to bring out what we focus on
Our beliefs and expectations influence our own behavior
Study by Mark Snyder found that when a man was led to believe that a woman found him attractive, she was more likely to act as if she did
Perceptual constancy: perceiving objects as unchanging (having consistent lightness, color, shape, and size) even as illumination and retinal images change
Many visual illusions result from the overuse of strategies employed to achieve perceptual constancy
Is Tile A or Tile B darker or are they the same color?
Illusion results from visual system’s attempt to maintain lightness constancy: we perceive an object as having a constant color, even if changing illumination alters the wavelengths reflected by the object
Shape constancy: we perceive the form of familiar objects as constant even while our retinal images of them change
A door casts an increasingly trapezoidal image on our retinas as it opens, yet we still perceive it as rectangular
MĂĽller-Lyer illusion:
Is line AB or line BC longer?
Size-distance constancy: Our brains are used to perceiving angles as corners that are near or far away and sees the inward-facing corners as more distant and therefore smaller
Are the two parallelograms the same size and shape?
Ponzo illusion:
Which line is longer?
Moon illusion: Does the moon appear larger near the horizon or when it is high in the sky?
When the moon is near the horizon we perceive it to be farther away from us than when it is high in the sky, but since the moon is actually the same size, our minds make it look bigger when it is near the horizon to compensate for the increased distance
The Magical Kingdom of Salt
In the Salar de Uyuni of Bolivia, the world’s largest salt flat, with no other objects in sights, the human eye loses its ability to establish a proper field of depth. The result is some bizarre pictures.
You enter a store and notice that the store doesn’t carry all the brands you may be familiar with, so you’re going to have to make your choice based on the product package alone
You are really interested in finding a brand that is considered
Durable, strong, and well built (“rugged” condition) OR
Classy, attractive, and refined (“sophisticated” condition)
Single-cell recordings of monkeys show activation of particular cells in lower temporal only when full-face photos of other monkeys are presented
Recognition accuracy for faces and houses: parts vs. whole
Study (Tanaka and Farah, 1993) in which participants were shown series of faces with person’s name and series of houses with owner’s name
Later on recognition test, they showed greater recall of
Parts of houses but
Whole faces
Do people tend to perceive men or women more in “parts”? Women (Gervais, Vescio , Forster et al., 2012)
Prosopagnosia: failure to recognize particular people by the sight of their faces
After stroke, sheep rancher could not recognize people but could recognize sheep
Someone would walk in the room and he wouldn’t be able to tell if it was the wife/neighbor/robber/etc
Note: the eyes also play a special role in perception
70-90% of famous portrait paintings sampled from the last five centuries have an eye at or within 5% of the painting’s exact centerline (Christopher W. Tyler)
Every man indicates in his eye the exact indication of his rank. – Emerson
Visual illusions suggest that the mind is at least in part modular (Jerry Fodor)
That is, it is not solely organized in terms of faculties, such as memory and attention, that can process any type of information
Rather, there are specialized information-processing modules that
Respond automatically
Cannot be “switched off”
Modular processes are usually characterized by…
Fixed neural architecture
It is sometimes possible to identify determinate regions of the brain that are associated with particular types of modular processing
e.x: fusiform face area for face recognition
Specific breakdown patterns
Modules can fail in highly determinate ways, which provide clues on the form and structure of processing
e.x: prosopagnosia
Other Neurological Disorders Related to Visual Perception
#
Visual agnosia: inability to recognize/identify visual objects despite relatively good visual perception
Usually due to damage in occipital or temporal lobes
“Mr. P” in Oliver Sacks’ Man Who Mistook His Wife for a Hat
Man with agnosia puzzling over a picture of a cow suddenly found himself making alternating up-and-down movements with fists. He looked down at his hands and said, “Oh, a cow!”
Due to some error between vision and verbal communication
Visual neglect syndrome or unilateral spatial neglect:
Tendency to ignore – or to be unaware of – information on one half of visual field, usually the left side
Typically occurs after damage (e.g., stroke) to right hemisphere, particularly damage to the parietal and frontal lobes
Relatively common, easy to test for
Patients are asked to bisect each line. Their markings are typically skewed to the right, as if they do not see the leftmost segment
Patients are asked to draw from memory or to copy an illustration (Driver & Vuilleumier, 2001)
House
Experimenter: Are the two houses the same or different? Patient: The same. Experimenter: Which house would you prefer to live in? Patient: The left house.
Capgras syndrome: characterized by belief that family and/or friends are imposters
Damage to pathway between visual cortex and amygdala, which regulates emotions
Emotional “glow” that we normally feel around people we are close to is missing
Ramachandran argues that this emotional “glow” is, to a large extent, what gives us a sense of continuity in our relationships
Classified as some kind of schizophrenia
Functional blindness (conversion disorder): unexplained vision loss with no organic basis
Cambodian women who had witnessed horrible war atrocities became either partially or wholly blind
Impairs primarily body functions / processes
Psychological defense mechanism
Blindsight: vision without awareness
Blindness resulting from damage to visual cortex
When presented with various shapes like circles and square, or photos of faces of men and women, patient could not tell (or guess) what his eyes were gazing at
However, when shown pictures of people with angry or happy faces, he was able to guess the emotions expressed, at a rate far better than chance
Patients are also able to correctly “guess” the identity or location of particular objects
Patients report that they get a “gut” feeling that allows them to perform these tasks
Blindsight patient was able to meander around all the clutter in a hallway that he was told was empty (Weiskrantz)
A second pathway of visual perception may account for this phenomenon
Adults who were born blind and later gained vision through newly-developed surgical interventions (e.g., cataract surgery) usually have some difficulty recognizing objects
At age 3, Mike May lost his vision in an explosion. Decades later, a new cornea restored vision to his right eye.
Unfortunately, although signals were now reaching his visual cortex, it lacked the experience to interpret them
May could not recognize expression, or faces, apart from features such as hair
Yet he can see an object in motion
Ended up committing suicide because he found himself in a world he didn’t (couldn’t) understand or comprehend
There is a critical period for normal sensory and perceptual development
Kittens reared in a cylinder with only vertical black and white stripes later had difficulty perceiving horizontal bars
Kitten would play with rod only when it was held upright
Cup/pail composed of cylinder and curved tube geons in a particular arrangement
An artificial neural network is an interconnected group of nodes, inspired by a simplification of neurons in a brain. Here, each circular node represents an artificial neuron and an arrow represents a connection from the output of one artificial neuron to the input of another.
#
Is Tile A or Tile B darker or are they the same color?
70-90% of famous portrait paintings sampled from the last five centuries have an eye at or within 5% of the painting’s exact centerline (Christopher W. Tyler)
Patients are asked to bisect each line. Their markings are typically skewed to the right, as if they do not see the leftmost segment
Patients are asked to draw from memory or to copy an illustration (Driver & Vuilleumier, 2001)
Blindsight patient was able to meander around all the clutter in a hallway that he was told was empty (Weiskrantz)
