Selective Noise Cancellation Headphones

Background:

Misophonia is a condition in which specific sounds cause acute emotional distress, anger, and /or panic. Some sound triggers are shared by many people with Misophonia, but many are unique to individuals.

Our Device:

During the 36 hour hackathon, my team and I built a prototype for a device which selectively cancels triggering noises through headphones to protect the user from induced panic. To do this, we transformed environmental sounds recorded on a microphone via the Fourier Transform to compare the result to identifiers of trigger sounds stored locally on an Arduino Mega. Further, we built a circuit which added the inverted sinusoid of the trigger sound to the original to provide noise cancellation through headphones in case of a match. Moving forward, we plan to implement a feature that allows the user to record triggering sounds which will be sent to a cloud platform where machine learning techniques will be used to compile identifiers of the most common triggering sounds to be stored locally.

Close Project

Febrile Seizure Detection

Background:

Febrile seizures are a medical phenomenon characterized by rapid, involuntary muscle contractions in patients experiencing a fever. Most common in infants between the age of six months to three years of age, data from the Center for Disease Control and Prevention predicts that there are 80,000 to 200,000 incidents in the United States alone of febrile seizures in infants anually. Currently, there are no portable devices on the market with the capability to detect febrile seizures and protect against their consequences. My sophomore design team and I thus sought to create a device which would meet this need.

Device Design:

Our device is designed as a beanie made from breathable fabric, onto which active electrodes are inkjet printed onto an interior elastic headband componenent. These electrodes continuously measure EEG data from the wearer and feed it into a processing unit. If pre-seizure brain state features are detected, this processing unit will trigger inflation of the protective cushions via release of C02 from a canister trigger valve system.

Detection Algorithm:

Our algorithm consists of three main components: data inputs, data preprocessing, feature extraction and analysis, and data outputs. Our inputs are recorded EEG signal data from pediatric seizure patients fed into our algorithm as if in real time. We perform preprocessesing to remove both line noise and biological noise using a Butterworth IIR filter and a moving average window, respectively. Thereafter for ten minute data segments, we used bandpass filters to extract gamma1 and gamma3 wavelengths from the signal, and compared the change in spectral power ratio of gamma1 to gamma3 frequency bands to a known value to map the time-period to either a pre-seizure or normal brain state label. In the case of a pre-seizure value indication, we output a signal to trigger cushion inflation.

Protective Cushion

Through a literature review, we determined a value for the maximum acceptable force an infant head can withstand. We thereafter determined experimentally that the optimal design specifications for our cushion when inflated are an approximately sphereical shape with a seven inch radius and construction from permeable fabric. These specifications allow optimal impact absorption and decceleration.

Close Project

Spatiotemporal Pattern Detection

Background:

The primary focus of the Keilholz Lab is to characterize the dynamics of neural functional connectivity in both pathogenic and non-pathogenic brain states. Over the course of this semester, I contributed to this mission by writing an algorithm which uses machine learning techniques to extract spatiotemporal patterns of functional network activity from fMRI scans of non-pathogenic human brains in the resting state.

Algorithm Details:

The inputs to my algorithm are preprocessed, volumetric fMRI data over time. I prepare this data for use by first applying a mask to the scan for each region of interest (ROI) and averaging the signal value at each time point to extract a sinusoidal time-course for each ROI. These represent the neural activity at those ROIs over time. Then, I transformed the data from the time domain to the spectral domain using a continuous wavelet transform. After, I isolated relevant data based on the frequencies known to remain in the signal after preprocessing. At this point, I was able to concatenate feature vectors of multiple time points to account for different windows of time, and compared the similarity of these using Kullback Leibler (KL) divergence. For different window lengths, I used this matrix of KL divergences for scans and sets of scans to run a t-distributed stochastic neighbor embedding algorithm. To build this algorithm, I adapted the procedure outlined in the journal article “Mapping the stereotyped behavior of freely moving fruit flies” by Gordon J. Berman, Daniel M. Choi, William Bialek, and Joshua W. Shaevitz. Thus far, this algorithm has been used to create a training data set which presents potential patterns in the brain scans used to make it. These patterns can be confirmed and used by the lab in the future as markers for brain pathology.

Close Project

Document Ingestion Service

Background:

The Gallery department ingests 9 to 15 thousand documents daily, each of which must be manually classified by type and read to extract metadata such as name, date of birth, account number, social security number, and other details. Currently, this process is outsourced and errors in transcription of metadata or routing of documents based on type can cause delays that span days in the process of case creation.

Our Goal:

We sought to partially automate the document ingestion pipeline within Gallery to increase the efficiency and accuracy of customer case creation and reduce Capital One’s dependence on third party vendors.

Our Project:

We created a service that used Optical Character Recognition (OCR) to extract text from scanned documents, a Random Forest Decision Tree model to classify the documents’ extracted text by type, and finally an algorithm that parsed this text to extract metadata. The metadata and document type were output as a .json, and can be used to route documents to their correct location.

Our Process:

Foremost, our team held meetings with vendors offering OCR and ICR (intelligent character recognition) services, conducted demos with their products and wrote reports for future use recommending the products best suited for our documents if the company were to pursue a contract. These included GoogleVision, AWS Textract, Vidado, ScribeFusion, Abto, and more. For our service, we selected Google’s open source software, Tesseract, because it could be run locally on our computers thus allowing us to safely use documents with customers’ personal contact information. To build a model for document classification, we used two main approaches. First, we tested the suitability of using the extracted text as an input to sentiment analysis models using Naïve Bayes, a Logistic Regression, and Random Forest Decision Trees. The second method we tested was using the image itself as the input to a Convolutional Neural Network for classification. We then moved forward with the most accurate model—the Random Forest at 92%. Thereafter, we wrote an algorithm to parse the OCR text output and extract relevant metadata, and compiled this information with document type into a .json to be sent to the next process in the case creation pipeline.

Our Impact:

The service my team built answered a long standing question within the Gallery department of the feasibility and value of automating this pipeline. Our process allows all form-based documents to be streamlined through this process enabling faster and more accurate case creation and better customer service.

Close Project

CapChat App

Background:

Carbon Hackathon is a 48 hour annual event organized by Capital One that takes place during onboarding week of their Technology Intern Program. At this event, I formed a team with four other interns I had met just days earlier.

Our Idea:

We sought to build an application that would facilitate networking within communities such as Capital One. Our design is based on the hypothesis that networking will increase if people in a given space such as a café are aware of others’ interest in conversation. For example, people who want to engage in conversation with someone new are more likely to do so if they can see which other people in a room are also willing to do so at that time. Similarly, people are more likely to share space with others if they know they are welcome to work alongside them.

Our Application:

Based on this idea, we built an app that displayed the table layout of the user’s location, with tables shown in green, yellow, red, or no color. These colors indicated, respectively: willingness to talk, willingness to quietly work as neighbors, desire to be left alone to focus or save space for an incoming group, and an unoccupied table. Through the app, users can change the display of tables and assign time allotments for given tables. To add a second layer of confidence, we used Google Pixels to store information about each table’s value and illuminate corresponding LEDs lights. We built a RESTful API to ensure all applications on the same internet network updated tables simultaneously. With this technology, we allow people to see upon entering a room where to sit based on their preference for social interaction, and thus we hope this fosters new connections between people.

Close Project

Data Structures and Algorithms

Data Structures Covered:

Arrays, ArrayLists, LinkedLists (Singular, Double, Circular), Stacks, Queues, Dequeues, Priority Queues, Binary Trees, Binary Search Trees, AVL Trees, Heaps, 2-4 Trees, Hash Maps, Hash Tables, Graphs.

Algorithms Covered:

Sorting (Bubble, Insertion, Selection Cocktail Shaker, Merge, Quick, Radix), Pattern Matching (Brute Force, Boyer-Moore, Knuth-Morris-Pratt, Rabin-Karp), Graph Algorithms (Dijkstra’s Shortest Path, BFS, DFS, Kruskal’s, Prim’s), Dynamic Programming Algorithms (Longest Common Subsequence).

TA Responsibilities:

In my role as a Teaching Assistant (TA) for this course, I taught these core concepts to 25+ students in weekly recitations with my co-TA. In these lessons we lectured about content, answered questions about concepts, homework, and tests, and solved problems with the students interactively. Additionally, I held 3+ hours of recitations weekly in which I tutored any of the 300+ students in the class on an individual level, whether that be to help them understand concepts or to troubleshoot problems with their homework. Further, I helped write J-Units for homework assignments as well as grade the assignments for my recitation based on J-Units specific for that assignment. These homework assignments were data structures or algorithms the students coded in Java. Finally, I graded student exams.

My Impact :

In this role, I was able to help students understand and find excitement in data structures & algorithms, as well as deepen my own knowledge of the course concepts.

Close Project