Midterm 3 Review PDF

Preview

Summary

Midterm study guide...

Description

COGS 1 Midterm 3 Review (Based on Professor Barrera’s Lecture 26 Review Slides) Adrian Lam --------------------------------------------------------------------------------------------------------------------Topic Overview Material from Past Midterms Lecture 18 - Dr. Taylor Scott - Distributed Cognition and Cognitive Ethnography Lecture 19 - Dr. Jim Hollan - A Glimpse of Human-Computer Interaction Lecture 20 - Dr. Brad Voytek - Data Science Lecture 21 - Dr. Virginia de Sa - Learning to See Lecture 22 - Dr. Gary Cottrell - How Does the Brain Make the Mind? Lecture 23 - Dr. Philip Guo - HCI/UX/Design Jobs for New College Grads Lecture 24 - Dr. Steven Dow - Advancing Collective Innovation Lecture 25 - Dr. Steve Barrera - So, what is Cognitive Science? --------------------------------------------------------------------------------------------------------------------Material from Past Midterms David Marr’s Levels of Analysis 1. Computational Level (Most Abstract) What problems does the system solve (and why)? What is it for? What does it do? 2. Algorithmic Level How does it solve these problems? What steps does it take? What representations does it use? 3. Implementation Level (Most Concrete) How is this instantiated in the brain or system? What are the ACTUAL pieces inside the black box? Major Theme in Cognitive Science A major theme in Cognitive Science at UCSD is exploring how cognitive activity can be distributed across multiple agents and representational media/devices. Understanding distributed cognition requires the consideration of a number of factors. Specific Goals for COGS 1 Understand some of the larger questions being asked by cognitive scientists [at UCSD]. 19 different speakers, from various parts of Cognitive Science. Have an appreciation for the tools and methods being used to answer these questions. Behavior, cognitive ethnography, computational models, electrophysiology (EEG/ECoG/single-cell), MRI, fMRI, DTI, text/image/video analysis, etc. Begin to see how these same tools and methods can be used to understand the typical instances of cognition in everyday life.

HCI, technology/product development, organizational behavior, medicine, education, navigation, fear conditioning, etc. Focus Area #1: Idea: Cognition is not something that one can do on his or her own. Instead, cognition often involves our surroundings, and it can also be limited by our surroundings and abilities. Lec. 18 Dr. Scott discussed distributed cognition. Understanding how people think requires looking beyond the brain and including their interactions with the objects, people, and systems around them. Lec. 19 and 23 Dr. Hollan (and Dr. Guo) demonstrated how technology can be used to study, explain, and augment cognitive activity. If our interactions with objects and systems are important to our cognitive activity, then using good design principles should augment this. Lec. 20 Dr. Voytek explored the emerging discipline of Data Science. Applying to issues as diverse as optimal paths in a city to deciphering brain wave patterns, this is a new set of techniques for understanding information beyond the capability of a single mind. Focus Area #2: Computational Modeling and Applications Idea: Different regions of the brain are specialized to perform or execute specific behavioral functions. This can be manipulated, modeled, and enhanced by artificial systems. Lec. 21 Dr. de Sa noted the difficulty in understanding a process like vision. Because of the variability and plasticity of the function, multiples approaches are necessary to understand and model the system. Lec. 22 Dr. Cottrell further descried how deep neural networks could model and predict complex human behavior. Lec. 24 Dr. Dow descried how groups can facilitate innovation. These principles can be used to design systems and artifacts to optimize creative processes. --------------------------------------------------------------------------------------------------------------------Lecture 18 - Dr. Taylor Scott - Distributed Cognition and Cognitive Ethnography Big idea: Dr. Scott discussed distributed cognition and cognitive ethnographies. Understanding how people think requires looking beyond the brain and including their interactions with the objects, people, and systems around them.

Ideas on Cognition Over Time: Behaviorism → Cognitivism → Post-Cognitivism Period 1: Behaviorism Behaviorism is interested in stimulus-response interactions. Key People: John Watson and B. F. Skinner. Downfall: Behaviorism cannot be the primary explanation for high-level human cognition because there has to be something that happens inside the “black box.”

Period 2: Cognitivism Cognitivism equates the human brain as the center of cognition and uses or describing cognition. Key People: George Miller, Noam Chomsky, and Allen Newell. Downfall: Cognitivism cannot be the primary explanation for high-level human cognition either because cognition is not isolated inside the brain.

Period 3: Post-Cognitivism Post-Cognitivism explains that human cognition when looking at cognition. Key People: Don Norman, Edwin Hutchins, and Jim Hollan.

Distributed cognition theory attempts to describ . In other words, . If you only examine the human agent, some of the cognitive activity will be lost. The social environment provides practices and constraints for cognition.

Example: Determining the answer to 2,374 times 5,649 requires us to use our brain and our surroundings (in this case, a piece of paper and a pencil).

Cognitive artifacts are objects to help do some of the thinking for us. Example: Calculators, pencils, and paper can be cognitive artifacts. If you use a calculator to solve a math problem, cognition was not entirely in your head.

A cognitive ethnography is the systematic study of how members of a community live, interact, and communicate with each other and with their surroundings.

Methods to Conduct a Cognitive Ethnography Make Observations and Take Field Notes Freeze Actions with Photographs Record Audio and Transcribe Interviews Record Video of the Interactions of People Problems with Cognitive Ethnographies

Problems with We fill in gaps in visual scenes and ignore what happens in the background. Problems with We do not hear disfluencies in speech. Problems with We only remember the gist of what people say, not word for word. Solving the Problems with Cognitive Ethnographies by Using the The “Cognitio-Scope” represents the techniques that allow us to overcome these problems with cognitive ethnographies. Slow Down and Pay Attention Set aside time to look, reflect, and critically examine the details. Be Honest Describe things as they are and do not interject our own expectations and opinions. Think Small The smallest moment of human activity is loaded with interesting cognitive phenomena. --------------------------------------------------------------------------------------------------------------------Lecture 19 - Dr. Jim Hollan - A Glimpse of Human-Computer Interaction Big Idea: Dr. Hollan demonstrated . If our activity, then using

s are important to our cognitive should augment this.

Professor Jim Hollan suggested the following:

Thinking with Computers Thinking isn’t done exclusively in the head – it is a Computers are special because they are so flexible and allow different types of expression (i.e. new kinds of thought). As computers increase in use, we are expanding and changing the ways we interact with them. A new computer form (e.g. Apple Watch) enables

Data Revolution are revolutionizing data collection of human activities in real-world settings. Sensors are in phones, computers, wearables, etc. Tools for capturing and using behavioral data creates Case Study: We can measure how humans interact with a car as they drive. This driving study is an example of recording behavior to have an objective understanding of what we do.

Activity histories provide software can better organize and use this

. New tracking

ChronoViz is an example of software that helps you organize and navigate

Activity Trails was a project that observed how people worked with computers in a law office. s could be organized and used to trigger episodic memory for people when interrupted. --------------------------------------------------------------------------------------------------------------------Lecture 20 - Dr. Brad Voytek - Data Science Big Idea: Dr. Voytek explored the emerging discipline of Data Science. Applying to issues as diverse as optimal paths in a city to deciphering brain wave patterns, this is a new set of techniques for understanding information beyond the capability of a single mind. Data Science Data Science is the study of how the quantification of observable phenomena can lead to human understanding of the processes giving rise to those phenomena. Unique Aspect of Data Science #1: We can use data to predict future outcomes. Example: How likely is someone going to click on an ad? Unique Aspect of Data Science #2: Certain phenomena require more or less data to lead to human understanding and/or prediction accuracy. Example: Google Translate is not very accurate with a small amount of data, but it is very accurate with a large amount of data.

Professor Voytek’s Perspective on Cognitive Science and Cognition What is Cognitive Science? It is the study of intelligences. Natural and Artificial. What is Cognition? Cognition involves communication, computation, reasoning, inference, memory, planning, and decision making. What has Cognition? People and children. Chimps using tools (and observing) to crush nuts. Crows (with very different brains) problem solving and using tools. Octopuses (with a distributed nervous system) using tools. Watson, the computer Jeopardy champion. Slime molds that can optimally solve the traveling salesman problem. The growth pattern of slime molds is intelligent in that it can form an efficient network between food sources. Case Study: Analyzing City Dynamics for Uber in San Francisco Study Using data of the Estimated Time of Arrival (ETA) and whether one completed their ride, we can examine how likely one will use services like Uber. Results If the ETA is too early or too late, one is not likely to ride with Uber. One’s willingness to wait from 2013 to 2014 has decreased. Case Study: Signals and Noise Study Play a sentence that is filtered, then clean, then filtered again. “She writes to her brother everyday.” Results We are unable to understand the first filtered sentence, but we were able to understand the second unfiltered sentence after listening to the clean sentence. Electrocorticography (ECoG) Electrocorticography (ECoG) allows you to record from the surface of the cortex. The main advantage is isolating a particular patch of cortex associated with some function. This can’t be done at the scalp or via fMRI because they lack the spatial resolution. This procedure is typically done prior to surgery for epilepsy. The goal to identify areas that should not be removed during the procedure. This takes many iterations and allows for the opportunity to do some experiments. Deep Brain Stimulation (DBS) Deep brain stimulation (DBS) allows you to activate a region of neurons. Electrical signals are sent to a specific nucleus via implanted electrodes. This modulates local neural activity, but the precise mechanism is unclear.

Deep brain stimulation is very effective in treating Parkinson’s disease and other movement disorders (dystonia) by stimulating subthalamic nucleus. Works about 40% of the time in major depression. Can be used for OCD, anxiety, and anorexia. Data Science and Computational Modeling Data science and computational modeling gives us a way reconstructing how the brain experiences something. By using ECoG, you can record the neural oscillations associated with some stimulus. After collecting many examples, a computational model is built to reconstruct the original signal based on brain activity. This can be used to control brain-computer interfaces (BCI) for people with neurological impairments. --------------------------------------------------------------------------------------------------------------------Lecture 21 - Dr. Virginia de Sa - Learning to See Big Idea: Dr. de Sa noted the difficulty in understanding a process like vision. Because of the variability and plasticity of the function, multiples approaches are necessary to understand and model the system. Idea: The Visual System is Not a Fixed Feed-Forward System What we see is influenced by our surroundings and experiences. All perception occurs within a context, and we make inferences based on context. It is heavily influenced by feedback. There are as many feedback connections as feedforward connections in the brain. Factors that Influence How We See Objects Past Experience Previous experience helps organize perception.

Surrounding Scenes The surroundings of the visual scene impact perception.

Recent Prior Exposure What we saw previously influences what we see. Example: The McCollough Effect and Other After Effects

Learned Familiarity with Special Objects

Learned familiarity with special objects affects what we see. Example: The Thatcher Illusion

Concurrent Input in Other Senses Concurrent input in other sensory modalities affects what we see. Example: McGurk effect and Sound-Flash Illusion

Idea: If you gain vision as an adult, you don’t automatically perceive the objects around you. This is because you don’t have the right experience to make sense of the incoming signals. Question: How can we evaluate how the brain processes incoming signals? Answer: 1. Single Cell Recordings 2. Optical Imaging 3. Microstimulation 4. Computation Models Single Cell Recordings Single cell recordings can be used in animals to identify patterns of action potentials in a neuron. This will yield a response curve.

Orientation Turning Curve An orientation turning curve tells you the responses to different bars of light.

Optical Imaging Optical imaging shows the activity of groups of neurons. This is useful for determining how nearby patches of cortex respond to stimuli.

Pinwheels ‘Pinwheels’ illustrate how nearby cortical columns respond to similar orientations.

Parallel Pathways for Sensory Processing There are parallel pathways for sensory processing, and they go through different visual cortical areas of the brain. Visual neurons in each pathway have different response properties.

V1 V1 is tuned to simple stimuli like edges and bars of light and represents the primary visual cortex. V2 V2 responds to more complex geometric shapes. MT MT, or the Middle Temporal area, responds to the direction of movement. IT IT responds to very complex stimuli like faces. General Principle The general principle seems to be that neurons respond better to more complex stimuli as you go further in the pathway. It also become less important where the stimulus occurs. Microstimulation We can test the functional role of specific neurons through the use of microstimulation. It was previously determined that neurons in area MT (Middle Temporal) respond to direction of movement. If you use microstimulation to activate neurons that previously responded to a particular direction, then the monkey will tend to report that direction. This is interpreted as activation neurons responsible for conscious perception of movement.

Computation Models Computational models allow you to test very specific hypotheses and evaluate their success. Suppose you want to test your hypothesis of how the visual gets shape information (3D) from shadows. Lehky & Sejnowski (1988) programed a computer with a learning algorithm to teach a neural network to see shapes. After training, the neural network was found to have receptive fields that are similar to the edge detectors in V1. This is supporting evidence that the algorithm was correct and that ‘edge detectors’ have a more complicated role than previously thought. --------------------------------------------------------------------------------------------------------------------Lecture 22 - Dr. Gary Cottrell - How Does the Brain Make the Mind? Big Idea: Dr. Cottrell further descried how deep neural networks could model and predict complex human behavior. Professor Gary Cottrell described the importance of using working models to explore cognitive phenomena. Working Model A working model takes the current state of knowledge and tries to implement it in a physical or computational framework. The better our understanding, the more accurate and precise the predictions of the model. The model can also help generate hypotheses about what we should expect. The model can only be as good as it information, so a good model of cognition needs input from behavioral measures (individual and social), large scale brain activity (e.g. EEG, fMRI), and fine scale neural activity (e.g. single cell, epigenetics). Dr. Cottrell’s 3 Axioms of Cognitive Science 1. The mind is what the brain does. [But also with help from the external world.] The “no spooky stuff” was meant to emphasize that it can be studied with physical measures. 2. What the brain does (i.e. thinking), is a kind of computation. All of the processes can be described in computational terms – recall Turing. 3. The kind of computations we use is probabilistic. It is constantly updated based on recent events (i.e. learning).

Question: Why are people still smarter than machines? Answer: Although computers may be faster and have more elegant programs (algorithms), computers are not exactly the same as the human brain. Brains are massively parallel computing machines. Different areas of the brain can respond to different characteristics of an object or scene. Historically, computers have been limited to serial processing – doing one calculation at a time. Neural networks were a new type of architecture based on our neural structure. Human-Style Computation Human-style computation and cognition involves combining information from our surroundings, unlike computers. Humans, as opposed to computes, are fast at combination information to do the following. Understand Sentences “U CN REED THIIS CANDT YU?” can be understood by humans, but computers have a hard time figuring this sentence out. Disambiguate Words In the sentence, “Billy picked up the truck and threw it across the room,” we know that this has to be a toy truck. Read Ambiguous Letters We can quickly recognize the word, fish.

Neural Networks Neural networks are built from individual units that are based on the neuron. Neural networks are made of simple units connected by weighted links. Computations are performed by spreading activation and inhibition to other units. They can be trained to perform many different tasks. If the output does not match what is desired for a particular input, an algorithm can change the weights. It forms a testable theory of how the brain works. The Interactive Activation Model When we read a word, we start at the feature level, followed by the letter level, followed by the word level. Step 1: Feature Level We look at the individual “bars” and shapes of each letter. Step 2: Letter Level We use these shapes to excite the compatible letters and to inhibit the incompatible letters. Step 3: Word Level We use these letters to excite the compatible words and to inhibit the incompatible words.

The Word Superiority Effect The word units feed back on the letter units, making them more active than they would be otherwise. We are better at seeing letters when they are part of a word than when they are in a nonword string of letters. Example: RED vs. PFB In the picture below, RED is much more likely than PFB.

--------------------------------------------------------------------------------------------------------------------Lecture 23 - Dr. Philip Guo - HCI/UX/Design Jobs for New College Grads Big Idea: Dr. Guo discusses the topics covered in HCI/UX/Design courses at UCSD, the different HCI/UX/Design jobs that are out there, and t...