Yifan Xu

Neural Model Library

Simple creation of feed forward and recurrent neural models in Tensorflow.

This library enables incredibly simple creation of feed forward and recurrent neural models in TensorFlow. Simply specify the input shape, output shape, and a list of components (or layers) to make up the network. The library then figures out the dimensions of each component, initializes each component, and builds the model automatically. Most importantly, the model is able to infer the sizes of intermediate layers (including recurrent inputs) automatically, requiring you to only specify the essentials. This enables you to make quick changes to your entire model in just a few lines.

Several of my friends have commented that, in working on this project, I've essentially just re-created Keras in Tensorflow. In retrospect, it is quite similar to Keras's Sequential model framework, with the addition of being capable of recurrent connections throughout the sequential model. Nonetheless, it was a good learning experience.

More detailed project description coming soon!
In the meantime, check out some of my other projects.

Close Project

NN Trainer Library

A library for automatic supervised training of neural models.

This library provides methods for training supervised neural network models without the need to implement all the necessary steps (eg. input handling, reformatting, batching, evaluation, feedback, etc.). It also allows for a large amount of customization in the training and evaluation process, while also making assumptions where possible, in order to minimize required user input.

More detailed project description coming soon!
In the meantime, check out some of my other projects.

Close Project

NN Experiments

A collection of various neural network models in TensorFlow.

This project consists of a collection of neural network models that I have trained on various datasets (eg. MNIST for digit classification with a CNN). Above, you'll see a simple example of a model that has learned to classify points of each of the four labels (i.e. colors) it's been given.

I started work on this collection of models as a means of learning applied machine learning.

More detailed project description coming soon!
In the meantime, check out some of my other projects.

Close Project

Collage Generator

Intelligently draws a message with images.

Collage Generator uses machine learning techniques, such as K-Means Clustering, to determine which images should be used to draw the message and which to use as the background.

Given a single set of images (.zip file) and a message to display, my Collage Generator returns a high resolution collage with the message drawn out using the images. It works by determining a subset of images that are the most distinct and drawing the message with those and the background with the others.

Close Project

OpenGL Game Engine

A custom game engine including dynamic lighting, terrain, and static and dynamic objects.

A custom game engine created from scratch in C++ using OpenGL.

This includes GLSL shader support for dynamic lighting (ambient, diffuse, and specular) and multiple light sources, algorithms to generate multiple different types of terrain (the images above show fractal-generated mountains and hills, respectively), and an intuitive object rendering system that handles both static and dynamic objects.

A previous version of this game engine created in Java using LWJGL and the JBullet Physics Library also includes a customizable object UI and real-time physics that can be used to simulate complicated collisions.

Close Project

Ray Casting Simulation

A simulation using AI techniques such as ray casting, spatial mapping, Bayesian filtering, pathfinding, etc.

An AI which simulates a robot that uses AI techniques to perform tasks such as estimating its location on a map and navigating towards a goal through an unknown maze in real time.

Given an image, the simulation creates a map and uses AI techniques such as ray casting, spatial mapping, Bayesian filtering, and pathfinding to perform tasks such as navigating from a start to a goal given no knowledge of the layout the map (as shown in the image above), or determine its location within the map based on the features of the map it has encountered so far.

The above video shows the AI navigate through the maze from start to goal without any prior knowledge of the map. This is done solely by collecting data from its sensors (the rays are shown in red) which report distances to obstacles, and then using techniques such as spatial mapping and pathfinding (modified A* algorithm).

Close Project

Seam Carver

An application for content-aware image resizing.

Using a technique called seam-carving, this application allows you to take an image and resize it without noticeably effecting the content.

Stretching, shrinking or cropping makes it obvious that an image has been modified, but seam-carving works by removing or adding pixels in the least noticeable locations in order to reduce or enlarge the size of the image. This application provides a UI for my implementation of the seam-carving algorithm. Above is an example of reducing the size of an image using content-aware resizing (the removed pixels are in red).

Close Project

Object Tracker

An object recognition and tracking tool.

An object recognition tool using OpenCV which uses color-based detection techniques, such as thresholding and the Hough Circle Transform to detect, learn, and then track different objects.

Using video from your webcam, Object Tracker will detect circular objects which you can then choose to track. After choosing an object, it will then determine the optimal HSV range for thresholding and then track the object based on that range.

Close Project

Audio Analyzer

An audio analyzer and controller for an Arduino-driven LED strip.

This application takes audio input from the microphone or speaker output, processes it, and then calculates certain audio metrics such as intensity gain, bass gain, average intensity, peak frequency, and estimated pitch. Coupled with the audio analyzer is an Arduino LED-controller, which I send serialized data about the changes in audio metrics in order to control a WS2812B LED-strip.

The audio analyzer is able to extract useful information from audio in real-time. When processing the data, it takes into account the non-linearities of the human ear and modifies the data accordingly, in order to extract the most noticeable audio features. Then, it determines what information is required by the LED-controller and uses bit-packing to serialize and send the data as quickly as possible to give the smallest possible response time. Currently, the estimated pitch is mapped to a point on the color spectrum and used to control the hue of the lights, and the bass gain is used to control the brightness of the lights.

Below is an example of the Audio Analyzer listening to the song Hymn For The Weekend (Seeb remix) - Coldplay.

Audio Analyzer listening to Hymn For The Weekend (Seeb remix) - Coldplay

Close Project

Request

A library used for sending HTTP/HTTPS requests with simple data management functions.

A simple library used for sending HTTP and HTTPS requests, with many data management functions designed to make sending requests and parsing the response as easy as possible.

Clients can easily add or modify query parameters, form data, URL-encoded form data, raw data, JSON data, and binary data via a Request object, and then easily read the response from a Response object, which includes automatic parsing into JSON and HTML data structures.

Close Project

Compiler/Transpiler

Parses and compiles a custom programming language into C++.

I've created my own C-based scripting language. While it strongly resembles other languages, it has features that others alone do not have. It includes Java's static type checking, C++'s primitive types, Python's binary operations and nested functions, and JavaScript's expression evaluation and objects.

It is designed to overcome some of the more annoying aspects of C++, such as its inability to be used as a scripting language and resolve functions defined later in the file or nested within other functions. Simultaneously, it retains many of the great features of C++ as well as features from other languages like Python and JavaScript, all while keeping the speed C++ is known for. This is possible because the compiler uses ANTLR and a context-free grammar to parse the code and then transpile it directly into C++.

I'm still working on this project and looking to add more features. Below is a brief example of a the language implementation which demonstrates some of the features of the language so far, followed by the abstract syntax tree generated by the parser.

print(test(0));                           // Use as scripting language

unsigned short test(int n) {              // unsigned short return type
    long double x = 2;
    if (n) {                              // if n != 0
        return (unsigned short) x ** n;   // x to the power of n
    }
    return tryAgain();

    unsigned short tryAgain() {           // Nested function
        return test(n + 1);
    }
}
                  

Below are two more examples: Binary search and the Ackermann function.

// Binary Search
int binarySearch(int[] arr, int value, int left, int right) {
    while (left <= right) {
        int middle = (left + right) / 2;
        if (arr[middle] == value)
            return middle;
        else if (arr[middle] > value)
            right = middle - 1;
        else
            left = middle + 1;
    }
    return -1;
}
                  

// Ackermann Function
int ackermann(int m, int n) {
    if (m == 0) return n + 1;
    if (n == 0) return ackermann(m - 1, 1);

    return ackermann(m - 1, ackermann(m, n - 1));
}
                  

Close Project

Monte Carlo Localization Library

A library for n-dimensional Monte Carlo localization.

Monte Carlo Localization (MCL) is useful in scenarios where uncertainty arises in a system over time (as a result of error; eg. physical instruments in the error-prone, real world) and when observations can be made to reduce said uncertainty. I've created a library for MCL in any number of dimensions, so that any number of error-prone factors can be accounted for.

The image above shows an example of the library in use in a simulated environment. Shown is a simulated instrument, initialized with no belief as to its initial position or orientation. Over time, and through observation, the instrument develops a correct belief as to its position and orientation using this MCL library.

This library is the result, in part, of a project funded by The MITRE Corporation and MIT's Department of Electrical Engineering and Computer Science through a MIT-MITRE Undergraduate Research and Innovation scholarship, and was developed primarily for use in Self-Driving Microscopy: Bayesian Inference for Localization and Autonomous Navigation.

Close Project

Swiper

Allows a smartwatch to control presentations and other remote software.

More detailed project description coming soon!
In the meantime, check out some of my other projects.

Close Project

Financial Graph

Graphs and analyzes your personal financial history.

More detailed project description coming soon!
In the meantime, check out some of my other projects.

Close Project

Web Scraper

Various purpose scripts for easy web scraping.

Scripts used to efficiently scrape the web using various filters and heuristics to speed up the process. One example is my Wikipedia shortest path finder, which, given a starting page and goal page, will find the shortest path between the two using only links on Wikipedia.

More detailed project description coming soon!
In the meantime, check out some of my other projects.

Close Project