The rules of music theory can be embedded into simple autonomous agents that interact in “music space” to perhaps produce pleasing compositions. Further rules can be applied to select for different types/genres/eras of music.
Project type
Research project with implementation.
Steps
Identify the different parameters that make up a musical composition (pitch, volume, duration, key, tempo, timbre, etc.). These can be considered the axes of music space.
Create a program that will allow us to spawn agents in music space and assign rules to each one.
Identify fundamental rules of music theory (e.g. the notes that are in certain keys, harmonizing intervals) that serve as the first level of constraint on how the agents can move through music space. At this point, we should be able to produce very simple and theoretically sound pieces.
Devise additional rule layers and tweak them to see what sorts of compositions or genres we get out of it and if they are pleasing at all. An additional layer would be an agent spawner that sets the initial conditions for how many agents there are and where they begin in music space.
Contributions
We will have developed the notion of “music space” that defines the possible configurations of musical elements in a composition.
We will have implemented a program that can be used to experiment with the application of different rules to agent-based music composition.
Hopefully we will have produced an album of music that is composed by our agents!
Archana Ram
A Theory of Visual Abstraction for Object Recognition
I believe that people have multiple abstracted representations of
a single type of object, e.g. a face, some of which overemphasize
features in order to allow for the underemphasis of other
features.
Project type
Pilot study.
Steps
I’m creating a dataset of drawn faces based on the CUFS dataset.
There are two categories — “face” and “not face.” The “face”
category will contain multiple abstracted versions of faces from
multiple angles—these abstracted versions will overemphasize and
underemphasize facial features. The “not face” category will
consist of all attempted abstractions that are not easily
recognizable as a face.
Drawings will be categorized based on human study, i.e. a google
form that shows the participant an image and asks “does this look
like anything to you?” It’s important not to have these questions
be multiple choice, otherwise that could influence the
participants’ perception of the image.
Images for which >75(“person” also works) will be placed in the “face” category and
will otherwise be placed in the “not face” category. The two
categories should be of roughly equal size.
Contributions
This data will yield insights into the way in which people can
abstract visual stimuli as well as help explain why people are
able to recognize partially occluded images.
If part of an image is occluded (represented in the dataset as
severely underemphasized), according to the above theory the image
should still be recognizable if the unoccluded features are
sufficiently emphasized, allowing for some sort of mental
templatematching between the representation and the visual
stimulus.
Benjamin Xie
The Genetic Epistemology of Computational Thinking Concepts
If we are to understand how programmers' computational thinking knowledge develops, then we must understand the order in which they learn computational thinking concepts in an open, informal programming experience.
Project type
Research project with implementation.
Steps
I will extract data on programming blocks used in each created project for MIT App Inventor users with at least 20 projects.
I will cluster learners based on similar patterns of learning and identify representative users for each cluster.
I will use the representative users to generate genetic graphs for each cluster and see how the order in which computational thinking concepts are used.
Contributions
Collected and analyzed data pertaining the order learners use computational thinking concepts.
Generated genetic graphs to represent the origin and development of computational thinking knowledge of learners that represent the population.
Identified similar (or contrasting) patterns in the representative genetic graphs.
Berj Chilingirian and Brian Copeland
Discovering Relations Between Previously Unrelated Ideas
We believe that the ability to combine known concepts to create novel ideas is fundamental to intelligence. To understand this process, we must first uncover a mechanism for relating two previously unrelated ideas.
Project type
Research Project with Implementation
Steps
Build a list of both tangible and abstract concepts (e.g. banana, monarchy).
Determine the correlation of each pair of concepts using Google Search result counts. We use these counts as a reasonable estimation of concepts' correlation strength.
For any two concepts, identify how they are related through other concepts, using correlation strengths. The intermediate concept serves as the relating concept between the two previously unrelated ideas. For example, suppose our list of concepts contains banana, dinosaur, and eating. If we ask the system, "How are banana and dinosaur related?" the system will respond, "Dinosaur and eating are highly correlated. Eating and banana are highly correlated. Therefore, the relation between dinosaur and banana is a dinosaur can eat a banana."
Contributions
Developed a mechanism to find the correlation between any two concepts.
Created a system to relate two previously unrelated ideas.
Daniel Sosa and Srinidhi Viswanathan
What Difference Does It Make?
Understand the innovations associated uniquely with human intelligence with a multidisciplinary approach.
Project type
Literature Review
Steps
Collect data on the genomic differences between humans and other animals using actively transcribed genes in human brain tissue.
Characterize differences between human behavior and other close evolutionary relatives such as chimpanzees.
Understand the role of the merge operation in various representations of intelligence, including Minsky’s k-lines, Gentner’s structural alignment and mapping, and Borchardt’s symbolic representation.
Contributions
Discerned candidate genes for intellectual innovation using existing cell-type specific and ancestral genomic data.
Identified unique aspects of human behavior based on observing primates.
Demonstrated the centrality of merge for several different representations of intelligence.
Proposed future directions for study in understanding intelligence that is uniquely human.
Dishaan Ahuja
Incorporating User Feedback in Emotional Analysis of Story Characters
In order to truly understand a story, a system must be able to determine and describe both the emotional state of the characters in the story, as well as what events in the story likely caused the emotional state. This description might take the form of sentences such as "John feels angry because Jack threw a tomato at him." A useful feature of such a system is the ability to learn from user feedback. The goal of this research is to develop a proposal for collection and incorporation of user feedback in the Genesis system sub-module that performs emotional analysis on characters in a story, and outputs statements similar to the example provided.
Project type
Research Proposal
Steps
I will determine which aspects of emotional analysis output would benefit from incorporation of user feedback. The two most obvious aspects are the emotion that the system determined, and the cause of the emotion that the system determined.
I will propose and design an interface for collection of user feedback. This interface will allow the user to select an error in the system output, and somehow provide the right answer according to the user. For example, a user could replace an emotion with another from a drop-down list, or change the causative event by selecting from a list of events in the story as told by Genesis.
I will develop a method for incorporation of user feedback into the emotional analysis system, so that on future runs on new stories, the system is able to use the insight from this feedback to provide a more accurate assesment of a character's emotions.
Contributions
Developed a method to enhance the accuracy of an emotional analysis system, based on user feedback on various aspects of system output.
Designed an interface, with descriptions and wireframes, to collect the necessary user feedback.
Dishita Turakhiac
Investigating role of social interaction in creative thought process
Examining the role of social interaction - verbal and visual in the creative conception if ideas and development of design. As designers, we often indulge in discussion or brainstorming as part of developing a design idea or thinking of a solution. A group discussion often results in better solutions. What is it about the creative thought processes that involves social interaction that makes it more efficient? Is our thinking process directly related and influenced by other thinking processes around us? Do human thinking processes perform efficient computations when interacting with peers?
Project type
Pilot experiment
Steps
Experiments involving creative brainstorming involving multiple combinations of people with varied levels of social interactions
Experiments exploring both verbal communication and visual/drawing communication
Literature review of available papers and studies on significance and role of social interaction in creative thinking
Literature review of social cognitive processes that assist in creative thought processes
Contributions
Paper presenting studies
Summary of available literature
Proposal for implementation of the findings – can a program be developed to assist humans to engage in effective and efficient thinking processes using modes of social interactions?
...
Ege Ozgirin
Seeing the story
In order to identify the role of perception in human story understanding we need to understand the role of visual representations in memory and reasoning.
Project type
Pilot Study
Steps
Small stories—varying in length and context—that composed of simple events and their combinations(e.g., "Steam rises. The steam contacts the metal plate", "the cat goes inside the circle. The cat sleeps. The cat goes out of the circle", "The pirate looks for treasure in the island, he considered alternative routes to go to 'X'.")will be presented to the participants in two formats. Control group will read the stories in written format and for the test group, the story will be presented in sequences of illustrations and they will be eye-tracked.
A week later, participants will be asked of which stories and what type of details related to stories they recall. Furthermore, they will be asked about the underlying concept in each story.
Data will be used to discern individual differences in reasoning strategies, conceptual thinking and memory.
Contributions
Gathered data about individual differences in recall and reasoning.
Observed that the time it takes for text-readers to recall stories were considerably longer.
Verified that illustrations aid individuals to remember and reason about the story better.
Eeway Hsu
Attack Planning for Control Systems
Cybersecurity is a growing issue in our society. Malicious attacks can range from stealing multi-billion dollars worth of credit card and personal information, to shutting off control systems leading to massive infrastructure failures.
The ability to hack control systems can become incredibly dangerous. To decrease exposure, we must consider attack planning. Attack planning builds trees of possible attacks, allowing preventative methods to be implemented.
Project type
Research Proposal/Project with Implementation.
Steps
I will meet with Dr. Howard Shrobe and read papers to create a good understanding of the existing work in attack planning and control system exploitation.
I will review existing vulnerabilities found in existing control systems (such as in real-time OS, remote connection) and decide what new rules should be implemented in Joshua.
I will decide on one small portion of control system hacking to focus and extend the attack planning method to work on this problem.
I want to look into heuristic selection of possible attacks based on factors such as ease of attack (monetary cost), openness to vulnerability (social cost).
Contributions
Develop new ideas necessary to extend existing attack planning methods specifically for control systems.
Implement new rules in the existing Joshua system for attack planning.
Harry Rein
The Emergence of Abstract Entities/Introducing Randomness into Conway’s Game of Life
Abstract entities are often composed of other abstract entities
- for example, the human body, while it always recycles its constituent parts, remains a human body. Human-like entities were the natural evolution of simpler organisms. We believe that complexity will naturally arise given a single rule: every entity has
a finite lifespan. Furthermore, less robust entities have shorter lifespans.
Single cellular organisms are less robust than multicellular
organisms. Given a sufficient amount of time, with some probability, a single cellular organism will be killed. Multicellular organisms are much more capable of sustaining themselves in their environments. At some stage of evolution, some single cellular
organisms had to evolve such that they acted in the mutual interest of other organisms, eventually reaching a point, at which, within their actions, was encoded a notion that they are a part of a multicellular organism. This is necessary for multicellular
organisms to exist, but it is at the expense of the single celled organisms that make it up. The multicellular organism maintains its identity even if some of its cells die. This is the natural result of our principle.
These multicellular organisms give rise to even further macroorganisms,
ad infinitum. We believe that this is one of the key principles in the development of intelligent life, and that we can better understand the process by building a simplified system.
Project type
Research Project with Implementation
Steps
Build a rudimentary game of life.
Hypothesize and experiment with different randomness based rule sets. The randomness is introduced so that our principle of finite
lifespan is guaranteed.
Search for the creation of organisms, that is entities with
a robust notion of self, defined by a continued existence, despite the specific cells (as defined in Game of Life) that make them up continuously changing.
Search for multiple layered organism, such that lower level organisms
can be identified as part of the whole.
(((Search for multiple layers of subsumption, that is, a group
of entities acting for the betterment of the whole.)))
Contributions
Resolved randomness role in the evolution of species.
Demonstrated the value of our rule in creating complex evolutionary processes.
Designed and implemented an extension of Conway’s Game of Life.
Jason Tong and Phoebe Tse
Understanding the Unspoken
If we are to understand the story in human dialogue, then we need to understand how humans recognize and respond to conversational implicatures. Take the following example:
Alice: Where does Eve live?
Bob: Somewhere in California.
Alice: Okay, I will ask David if he knows.
The implicature that Alice draws from Bob’s response is that Bob does not know where Eve specifically lives (e.g. region, city, and street name). If Alice does not recognize this implicature, she may ask “Where in California?” or a similar follow-up question. Bob would then most likely respond with an exasperated “I don’t know!”
Alice’s follow-up question would reflect a failure to understand that Eve’s address is likely absent from Bob’s knowledge space. Bob’s ignorance is not explicitly conveyed in the dialogue, but is understood by both participants. Recognizing this implicature enables more efficient (and less frustrating!) communication.
The vision is to examine the mechanisms involved in dialog understanding and determine what computational imperatives are necessary for processing implicatures.
Project type
Research Proposal
Steps
During the first week, we will explore the existing Genesis infrastructure by reading papers and experimenting with the system. The purpose would be to gauge Genesis’s current capacity to understand reported exchanges between characters in stories.
We will find examples of conversations that would induce gaps in Genesis’s understanding due to the presence of implicatures.
The next week, we will outline the computational imperatives used by humans to process implicatures. We will apply the computational imperatives to selected story examples to evaluate the results against what humans would reasonably produce.
In the final weeks, we will outline the requisite steps for integrating the computational imperatives into the Genesis System.
Contributions
We demonstrated that the failure to recognize and draw inferences from implicature significantly impairs dialogue understanding and the ability to build accurate models of others’ knowledge and belief states.
We detailed the computational imperatives and the associated representations necessary for understanding implicature.
We demonstrated that dialogue comprehension can be explained by our computational imperatives and representations through applying them to instances of implicature.
Julian Brown and Natalie Lao
Image Fooling in Simultaneous Neural Networks
If we want to build secure systems using neural networks, we must first attempt to prevent neural network fooling of the type demonstrated by Nguyen.
Project type
Reimplementation of the Nguyen paper.
Steps
Two different neural nets, Tensorflow and CMU-Shufelt, will be trained on the same subset of the CMU face images directory. This directory consists of labeled photos of 20 subjects of varying pose, expression, and attire.
After training is complete, the two neural nets will be applied to the rest of the image data. Confidence levels for facial-ID recognition will be recorded for both neural nets for each image.
We will use a hill-climbing algorithm and an evolutionary algorithm to create fooling images for each neural net. A fooling image should map to one individual/ID with high confidence for the neural net that it was created on.
Fooling images created from one neural net will be applied to the other neural net. The resulting confidence levels will be recorded.
We will use the hill-climbing and evolutionary algorithms to attempt to create fooling images for both neural nets at the same time.
We will draw conclusions from this data about whether implementing simultaneous neural nets helps prevent fooling.
Contributions
Implemented hill-climbing and evolutionary algorithms to create fooling images on two different neural nets simultaneously.
Nguyen presented research that cast light on the differences between human vision and current deep neural nets. His results suggest that current neural nets are only capable of learning low-level and middle-level features rather than a global structure, allowing them to be easily fooled by nonsense images. We determined that simultaneously applying multiple neural nets decreases (or does not decrease) the rate of fooling.
Neha Patki
Understanding the Strength of Emotional Empathic Alignment
If humans can recognize actions by aligning their body and imagining the physical sensations, then the same process should also allow them to imagine a mental state and identify an emotion. My goal is to quantitatively measure empathetic alignment in the space of human emotions, and determine how the absence of sensory data or prior experience affects the degree of alignment.
Project type
Pilot Experiment.
Steps
Subjects will be split into three groups. The control group will be shown a randomized series of short 30-second video clips representing scenarios where the main character experiences a range of different emotions. The second group will be shown the same videos with the audio turned off. The final group will only have access to the audio, and will not get to see the video.
After watching and/or listening to each of the scenarios, subjects will be presented with 6 emotions: anger, fear, sadness, enjoyment, disgust, and surprise. (These are based on Paul Ekman's work on basic emotions.) For each of these emotions, subjects will indicate their estimate of the main character feeling the emotion, on a scale of 1 to 5 (1 being unlikely; 3 being unsure; 5 being very likely). Subjects will also be asked to rate if they have ever been in a role similar to the main character, on a scale of 1 to 5 (1 meaning never been in the situation; 5 meaning experiencing the exact same situation).
I will draw conclusions about how the availability of data (audio and visual), and the experience of similar situations affected the subjects' emotional responses to the scenarios. In particular, I will consider the confidence of the responses based on how far away the response is from unknown, and the accuracy of responses based on a comparison with the control group.
Contributions
Conducted experiment and analyzed data relevant to determining how the amount of sensory data and previous experience affects the ability to align to a different person's emotional state.
Demonstrated that the degree of similarity of the experience causes the emotional alignment to be more definitive (or has no effect on how people identify emotions).
Demonstrated that the absence of sensory data (video or audio) causes emotional alignment to be less definitive (or has no effect on how people identify emotions), and less accurate (or more accurate) as compared to having full sensory data.
Oscar Rosello
Steps towards visual computational creativity
If we want to build a creative computer programs, we need to first define a model for creativity.
Project type
Research proposal .
Steps
Research literature on creativity.
Research literature on models for creativity.
Research literature on computational creativity.
Research on neural network applications.
Identify previous case study precedents.
Contributions
Surveyed the current state of computational creativity.
Proposed a conceptual model for visual computational creativity.
Sami Alsheikh
Physical Scene Understanding through Simulation
If we are to understand how humans interpret the physical world, then we need to fully consider approximate probabilistic physical simulations as a potential mechanism.
Project type
Pilot Study/Reimplementation of Tenenbaum's "Simulation as an engine of physical scene understanding."
(http://web.mit.edu/~pbatt/www/publications/BattHamrTene13PNAS.pdf)
Steps
Read and study Tenenbaum's "Simulation as an engine of physical scene understanding."
Implement a computer application that displays various color blocks on a table (as seen in Tenenbaum’s paper) and can simulate ‘ground-truth’ physical results.
Implement a rough approximation to the ground-truth physical simulation (called an `intuitive physics engine' (IPE) in the paper) that ostensibly produces more human-like predictions than a ground-truth physics engine (faster, but potentially less accurate).
Construct multiple block scenes and ask various questions to subjects as done in the paper. For example, when shown a group of blocks stacked in a specific configuration, one could ask `Will this tower fall?'
Compare the answers from subjects with the answers produced from the intuitive physics engine and the ground-truth physics engine.
Contributions
Designed and implemented a 3D geometric representation of a block world with a ground-truth physics engine.
Designed and implemented an IPE inspired by the Tenenbaum paper.
Collected data on human predictions to basic questions about block scenes and compared those predictions to the ground-truth and IPE
Scott Penman
Computing Design Iteration
In order to understand the capacity of computation to model creativity, we must investigate the creative processes used by designers. Iteration is a generative and exploratory technique employed in design thinking. By defining the characteristics and evaluating the methods of iteration, I intend to explore its potential as a computational process.
Project type
Research project with implementation
Steps
Conduct literature review on the topic of design iteration (i.e., iteration as a technique in the design process).
Define iteration in terms of computation.
Develop computational model of design iteration.
Evaluate computational model against human creative process.
Contributions
Definition of iteration as a technique in the design process.
Computational model of design iteration.
Shidan Xu
Investigating Difference in Depth of Ynderstanding in First/Second Language
There is a fundamental difference in depth of understanding for humans using first vs. second language. Often, using first language, one can understand the concepts deeper and have more meaningful conversations. Such thoughtful conversations and ideas may be deemed more intelligent, however the human subject is no different. What does this depth of understanding depend on? Is it through the merge operation, memory capacity, use of exact (non-suitcase) words, etc.?
Project type
Research Proposal / Pilot Study
Steps
Reread papers on symbolic merging and explore some papers on L1, L2.
Create an experiment that use human subjects to test aptitudes in merge operation, memory capacity, skimming abilities, etc., in their first / second languages.
Analyze the results to identify which aspect caused the difference in depth of understanding
Contributions
Surveyed a range of possible areas for causes for depth of understanding.
Identified the most important cause and the most different aptitude and analyzed which basic operation was affected.
Steven Fine
Near Miss Teaching
If we are to understand how people readily learn new concepts, then we must understand the specific advantage that is gained from near-miss learning.
Project type
Pilot Experiment
Steps
Subjects will be divided up into two groups. In the experimental group, subjects will be asked to read two similar stories about a specific concept. One containing the concept of interest, and one not. In the control group, subjects will be asked to read only the story that contains the concept of interest.
We will ask the subjects questions to determine what aspects of the concept they understand, and which aspects they do not.
We will draw conclusions from this data about what specific aspects of concept comprehension are gained from being exposed to two stories, one of which is a near miss.
Contributions
Collected and analyzed data relevant to determining the type of advantages near miss learning brings.
Demonstrated that being exposed to two stories, one of which is a near miss, drastically improves one's understanding of a specified concept.
