Understanding Visual Routines through Maze Solving
In this collaboration with Benjamin Balas and Ruth Rosenholtz, I seek to
determine what makes mazes harder or easier to solve. By constructing
mazes with varying properties (ie. wall thickness, wall shape, ratio of
white to black space, etc.) and asking subjects to quickly determine
whether a given maze is solvable or unsolvable, one can obtain data
regarding how long it takes a subject to determine the solvability of a
with respect to these properties. Another possible area to explore is
the role of peripheral vision in maze solving.
Project type
Research project with implementation
Steps
Construct several sample types of test stimuli (mazes with varying
properties)
Try to determine empirically which mazes are harder and easier to solve
Construct experiment(s) in Matlab
Run experiments on subjects
Analyze data
Contributions
Design and perform an experiment to determine whether modifying
various properties of mazes have an effect on the time it takes to solve
those mazes
Determine how the results relate to what we know about human vision
Aaron Blankstein
The Application of K-Lines to Metaphor Approval
If we are going to understand the role that language plays in human
intelligence, then we must understand how language is used creatively.
Metaphors are abundant in language, and particularly creative language
and literature.
Project type
Research Project with Implementation
Steps
Generate an example set of interesting metaphors.
Develop a conceptual model using k-theory to represent a object.
Implement this model using high level representations.
Apply a matching algorithm to this model to compare knowledge about
the world to possible metaphors.
Test this matching system against a set of examples.
Contributions
I will:
Present interesting examples of acceptable and unacceptable metaphors.
Develop a model based on k-lines for approving or rejecting metaphors.
Implement and test this model on the examples presented.
Martin Couturier
Hashing Over Verbs in a Self-Organizing Map of Threads
If we are to overcome memory limitations associated with the Genesis memory system, then we need to come up with a way of speeding up self-organizing maps in a way that directly connects to the Genesis requirements.
Project type
Research and Implementation
Steps
Divide the problem of word/thread storage into smaller and more manageable units by hashing over related verb types.
Each element type (transition, appearance, etc) recognized by Genesis will possess its own self-organizing map that has been hashed over related verbs.
I will test the results of this new memory storage system and determine if there is an improvement in the effectiveness and capacity of memory.
Contributions
I will:
Propose a new method of dividing the thread storage problem, which takes into account large data sets. This will hopefully improve performance.
Implement a memory storage program, adapting the existing Self-Organizing map model to accommodate smaller sub maps through the use of hashing.
Cory Ip and Mathew Cherian
Efficacy of linguistic aids in recall
If we are to understand human intelligence, then we must understand how humans can consciously recall past experiences. Extending the ideas of multimodal dynamics for perceptual grounding and interpretation presented by Coen, we postulate that actively engaging different components of the human language processing system in an experience will enhance recall of that experience. To investigate multimodal dynamics in recall, we plan to conduct a pilot study.
Project type
Pilot experiment.
Steps
We will have a control group and one or two experimental groups. Each group will consist of approximately five subjects. There will be a test session and a follow up session. During the test session, each subject will be asked to read a passage from a standardized test (SAT or LSAT). For the subjects in the control group, reading will be the sole activity during the test session. For those in the experimental group, we plan to select two supplemental activities from the following list ? reading aloud, free exposition of the content to a listener, reading a few questions related to the content of the passage before reading the passage, reading a few questions related to the content of the passage after reading the passage, free-writing after reading the passage.
A week after the experiment, all subjects will participate in a follow up session in which they will be asked to answer questions pertaining to the content of the passage to determine how much they recall.
We will try to determine whether the two supplemental activities aid the experimental subjects in remembering content more than the control group subjects. We will also examine whether one activity is superior to the other as a recall aid.
Contributions
We will:
Collect and analyze data relevant to oral and written aids for recall.
Matthew Fay and Sarah Beth Shiplett
In other News the Sky is Blue: Teaching Common Sense for Story Understanding
If we to replicate the human ability to reconsile differences in perspective through discourse, then we need to understand how two agents in a teacher-student relationship can come to a shared understanding a story.
Project type
Research project with implementation
Steps
Setup the Gauntlet system with two slightly diferrent versions of stories to be learned and two agents in the Gauntlet system with different common sense knowledge. Assigning the roles of teacher and student to the agents.
Have both the teacher and student simultaneously process the story. The teach will identify mistakes in the students understanding as the occur.
Use the wavefront analogy for learning and felicity conditions to determine the best way for a teacher to respond to each of the student's mistakes by suggesting changes to common sense rules.
Evaluate the effectiveness of this method in enabling shared story understanding.
Contributions
We will:
Implenent the teacher-student learning methods into the Guantlet Systems.
Demonstrate that the wavefront analogy allows a teacher to respond appropriately to student mistakes.
Identify the felicity conditions needed for effective collaborative story understanding.
Alborz Geramifard
Task Adaption Through Online Feature Discovery
Humans are smart creatures. They take advantage of their past experiences to perform new tasks ahead of them. Often we hear teachers try to relate the new problems to the old ones in order to provide their pupils with interesting starting points. If we were to understand such synergistic trend in the human mind, we should comprehend how representation is evolved through the course of human life. In particular people tend to associate key correlation among perceptual features together in order to solve tasks. When faced with a new task, they reuse the existing representation and massage them to fit it to the existing task. This project explores task adaptation through migrating representations for autonomous agents.
Project type
Research project with implementation.
Steps
On the first step an incremental feature discovery code will be expanded to include the notion of forgetting of features. There are empirical evidence that the rate that memory forgets things is related with a power-law function [Chater and Oaksford, 1999] which will be used for forgetting.
Using the incremental feature discovery code, the agent is trained to solve a problem in the classical Blocksworld domain based on initial logical features (e.g. the red block is on top of the green block).
After a fixed number of interactions the old agent and a new agent will start a different problem which some similarities to the previous problem. The project inspects if dynamically massaging the representation through forgetting and expansion could help the first agent to reuse some of the learnt knowledge to excel in the new task.
I will conclude whether incremental feature discovery could be a plausible model for humans to adapt their representation through the course of their life span. If time permits I am interested to replicate the same experiment with humans as well and see how simulation results match human data.
Contributions
I will:
Add the notation of forgetting features to the existing incremental feature discovery framework.
Demonstrate that massaging the representation through forgetting and expansion is or is not a plausible approach for transferring knowledge among similar tasks.
Compare the simulated results with actual human data if time permits to rerun the process on human subjects.
Thananat Jitapunkul
Tracking Objects by Mapping Image to Log-polar Coordinates
Tracking objects can be a computationally intensive process. To determine locations of tracked objects, areas in the search space are compared to the reference template to find the image area that best matches the template. However, moving objects involves a large amount of scaling and rotation. Under Cartesian coordinate system, the operations make the comparison between object template and potential images more complicated.
Rotation and scaling can be represented in an easier format by mapping the Cartesian coordinates in normal images into log-polar coordinates. With a certain point as center of image, the log-polar coordinate system has one axis representing the logarithmic distance of the pixel from the center, and the other representing the polar angle with respect to the center to the pixel. Under this new coordinates, both rotation and scaling can be simply treated as translation, making the process much less computationally intensive.
In addition to the computational advantage explained above, implementing the tracking system based on log-polar coordinates can give insights into human cognition. Cognitive psychologists have earlier found that human visual system views the world based on the log-polar system rather Cartesian coordinates. Therefore, better performance under the proposed scheme would reiterate the significance of the finding.
Project type
Reimplementation
Steps
The first step of the project is to implement the system that maps the normal image into log-polar coordinates.
In parallel, a simple user interface should be implemented to demonstrate the performance of the tracking algorithm. The interface should show video sequences with tracking markers drawn over them.
Several short video sequences will be recorded as a testing data. If the implemented algorithm is sufficiently efficient, real-time data might be included.
Lastly, I will reimplement work in “Object Tracking Using Log-Polar Transformation” which is a Master's Thesis by Saikiran S. Thunuguntla from Louisiana State University and Agricultural and Mechanical College. (http://etd.lsu.edu/docs/available/etd-07072005-113808/unrestricted/Thunuguntla_thesis.pdf).
Contributions
The project will demonstrate the effectiveness of log-polar transformation in object tracking. The performance should be robust and reliable to a certain degree.
Avril Kenney
Learning Semantic Representations in the Genesis System
For computers to be fully capable of using natural language, they must be able to use semantic information to construct models of what the words and sentences actually mean.
Project type
Implementation
Steps
Implement a system that learns to identify the appropriate Genesis representation (event/world structure) for a sentence given the semantic output from the START parser.
Train the system using labeled positive and negative examples of a few classes of representations.
Assess the modified system's performance by giving it novel sentences and determining whether its representational outputs are in accordance with human intuition.
Contributions
Characterizes a correspondence between semantic structure (linguistic) and event structure (non-linguistic).
Proposes a method for using examples to learn how to infer what type of representation to use for a given input.
Increases the power of the Genesis system to draw relevant information from semantic structure.
Yuri Lin
Trajectories as a Tool for Translation
In order to understand how people effectively learn a second language, we must first understand how they learn associations between concepts and expressions in the language they are learning. To help build this understanding, we can use a computer to model the process of using concepts to translate between two languages.
Project type
Research project with implementation.
Steps
This project focuses on trajectories as an example of a tool to be used for translation between Chinese and English; major steps will be to:
Establish an internal representation for trajectories for the system.
Determine rules for expressing simple trajectories in two languages, English and Chinese, and teach the system about these rules as well as some dictionary words for use in translation.
See whether the system is able to translate trajectory sentences between the two languages by using the rules and its understanding of trajectories.
Contributions
With this project, I will:
Build a system that is able to translate trajectory phrases between languages, given rules of how to express a trajectory in that language.
Demonstrate, with this code, the potential effectiveness for humans of learning new languages by learning how to express concepts such as trajectories and transitions.
Robert McQueen
The Effect of Society Influence on Personal Choice
If we are to understand the role of societal influence on choosing goals, we must understand how varying levels of trust between influential agents affect the internal, decision-making models of individuals
Project type
Research project with implementation
Steps
Define how an internal model affects the decisions of an individual. I will also need to identify how internal models of two individuals interact depending on their relationship.
Write a program that demonstrates an internal model and how it can be used to influence other internal models.
Program a data structure that models an influence network that consists of internal models
Contributions
I will:
Extend Minsky's Theory on Imprimers with a weighted influence network composed of nodes representing influential objects and weighted edges as trust relationships between influential objects.
Implement a program in Java that constructs internal value models of people based on societal influence.
Implement an influence network that enables internal models to interact with one another that is consistent with Minsky's Theory on Imprimers
Rizal Muslimin
Tangible Analogy in Visual Perception
Humans are capable of solving many visual problems by simply imagining the shape being transformed with different analogies. If we want to understand the role of visual analogy in reconstructing shapes, then we need to understand how the input geometries are segmented and associated in our cognitive map to reform them back into a meaningful shape.
Project type
Research Proposal
Steps
First, I will set up the agents of Tangible Analogy: [a] Geometry: point, line, plane, volume; [b] Tensor: a vector with coordinate and direction to be applied into the geometry; [c] Template: a set of basic rules for tensor to behave; [d] Icon: a set of performative geometry (a geometry with a rule-assigned tensor); and [e] Iconic Net: a network of Icons in n-dimension.
Second, I am going to propose a visual processing mechanism of how these agents transform the input geometry into different performative shapes. [1] Fragmentation: Segmenting and grouping the initial shape to certain geometry [2] Mapping: Linking the fragment with the tensor in a different level in Iconic Net [3] Substitution: associating different icons to perform different analogy in the Iconic Net.
Third, illustrate Tangible Analogy scenarios with the following condition:
[1] Use one shape with different templates and icons.
[2] Use two different shapes with the same template and icon.
[3] Use the two previous scenarios in different Iconic Net dimension.
This research proposal would probably be better considered as an ensemble of related readings from 6.XXX. For instance, I will exercise Jackendof's Place-Path function for Tensor behavior; Minsky's K-line for the Iconic Net; Winston's Frames/inheritance together with Borchards Causal Reconstruction for the Substitution model; and Winston's Arch Learning to classify the icon based on their initial properties.
Contributions
I will:
Map the iconic net to show Tangible Analogy mechanism in perceiving shapes in a different way.
Illustrate different scenarios for the shape-templates-icon relationships to evoke different analogy.
Priya Ramaswamy
New Method of Neuron Learning
If we are to understand how humans learn, we must first understand the most basic agent of learning: the neuron.
Project type
Simulation and Data Analysis Experiment
Steps
Work will be done with Prof. Staelin. I will be using his novel neuron learning architecture to carry out simulations to understand the instant learning mechanism of the neuron. This will be done by varying and optimizing parameters of the model.
Contributions
I will:
Modify parameters of Staelin's neural learning model to understand how it effects the model. I will connect findings to neuroscience.
Daniel Rosenberg
From Scribbles to Drawings: Model for self-supervised scribbling-drawing learning
If we are to explain how humans are able to create novel things,* then we must explain how the process of scrabbling develops into drawing through the recursive cross-modal interactions between the senses, and the construction of new metal images through association and analogy.
*Novel things are understood here as the construction of things that did not exist in the world and therefore cannot be represented before they are created—humans bring them into being.
Project type
Research Proposal with Model (for future implementation)
Steps
Define the different stages of the scribbling-drawing development in children (Piaget’s and Arnheim’s studies)
Review Minsky’s, Rao’s and Coen’s key AI concepts and how they can be related to creative processes.
Use the psychological background on scribbling-drawing development in children in relation to the AI background on action, perception, and self-supervised learning to construct a model for self-supervised scribbling-drawing learning.
Contributions
We will:
Relate different reading from 6.XXX with the scribbling- drawing development in children. Particularly Minsky’s Key-Lines theory, Rao’s Visual Attention thesis, Coen’s Sensory-motor Learning model and implementation, and Gentner’s Analogy and Similarity theory.
Extend Coen’s work on Multimodal Dynamics and Self-supervised Learning to Artificial Systems able to Draw.
Propose a Model for self-supervised scribbling-drawing development.
Eli Stickgold
Learning From a Small Set of Examples Using Threads
If we are to understand the way humans learn, we need to be able to recreate the human ability to learn with only a small sample set. The Lattice Learning algorithm provides us with a way to generalize learning quickly on objects stored with heirarchical organization.
Project type
Re-implementation
Steps
I will reimplement the lattice learning algorithm described by Klein in java code.
I will test this implementation with a variety of sample sets to ensure it doesn't draw false conclusions.
I will test the number of data points from a given set the algorithm needs to reach a state where it can correctly predict the status of all other data points, once assuming optimally chosen examples, once assuming randomly chosen samples.
If I have enough time, I will also extend beyond Klein's implementation to allow the algorithm to recognize the concept of a negative exception.
Contributions
Demonstrate an implementation of the lattice learning algorithm.
Assess the effectiveness of this algorithm in learning on a classification problem.
If possible, develop the algorithm to allow for negative exceptions.
Andrew Sugaya
Battlecode: Modularity and Intermediate Functions in a Genetic Algorithms
To create an unbeatable AI in the framework of Battlecode, we need to develop a system that is capable of adapting to its opponents, environment, and goals.
Project type
Research Project with Implementation
Steps
Be comfortable working with the Battlecode engine and be capable of writing basic players.
Develop a genetic algorithm-based player that will explore as much of the map as possible.
Develop further modules that will allow the player to seek and attack its opponents.
Report on the experimental results. In particular, discuss the successes and failures of the various design choices and alternatives.
Contributions
I will:
Develop a Modular Genetic Algorithm-based Battlecode AI that uses intermediate functions and is capable of seeking and attacking an opponent.
Analyze the advantages and shortcomings of a genetic algorithm-based Battlecode AI.
Varvara Toulkeridou
Dynamic Descriptions
Many theorists suggest that creative behavior is embedded in perception. If we are to envision intelligent design machines, we have to understand that they should be able to perceive their environment and form intrinsic representations, “entities” that bear content independent of our interpretations.
Project type
Reimplementation
Steps
I will highlight what I consider as an essential feature that a computational system capable of creative activity should have.
To make my point of view explicit I will reimplement the SOM algorithm described in Larson’s Thesis. I will use input images of linear patterns and test whether the system can generate its own descriptions on which it can further apply transformation rules.
Contributions
I will:
Demonstrate that the ability of a system to perceive and constantly generate intrinsic descriptions from its environment, is fundamental for creative activity.
Propose a possible framework for a design machine.
Reimplement and test the SOM algorithm on an aspect of the framework proposed.
