Gaurav Gite

This project which aims at reducing the electricity consumption of residents in the buildings by giving them personalized feedback on their electricity usage. We hypothesize that giving sentimental feedback, including the information about the neighbours, a news article summary about environmental issues and concerns, and other langauge based feedback features could motivate the residents to reduce their usage. A sample feedback message is shown in the gure below. Currently, we are sending feedback twice a week, every Monday and Friday, to an experimental group of around 90 apartments in a Columbia-owned building. The experimental group is compared to a control group of around 100 apartments from the same building. This system uses MCMC Gibbs Sampling inference to learn the Multivariate Gaussian Mixture Model, to form groups of apartments based on the response to the feedback. In each group, we use logistic regression to identify the important aspects of the feedback message such as sentiment of the message, presence of graph, information about the other similar apartments, etc. to personalize the feedback.

Recurrent neural networks (also called RNN) are gaining popularity in various interesting applications like next-word prediction, speech recognition, machine translation, or image captioning. In this blog, I have looked at a simple example, GenPredict, Name-based Gender Prediction. Genpredict could also be considered as a 'Hello World' example of the sequence classification task. We know that RNNs work because of their ability to recognize patterns within the training data. This project tries to understand which patterns are recognized by the neural network. Demo Video, Blog post , Code.

Facial expressions convey non-verbal cues, which play an important role in interpersonal relations. Automatic recognition of facial expressions can be an important component of natural human-machine interfaces; it may also be used in behavioural science and in clinical practice. Although humans recognise facial expressions virtually without effort or delay, reliable expression recognition by machine is still a challenge. The project provides a thorough comparison of various methods, like Neural Network, Lasso, Group Lasso and Fused Lasso, to solve this problem. It also identifies the strengths and weakness of each of these techniques. As part of the project, we gave a real-time prediction on facial images captured from a webcam. The data consists of 48x48 pixel grayscale images of faces. The faces have been automatically registered so that the face is more or less centered and occupies about the same amount of space in each image. The task is to categorize each face based on the emotion shown in the facial expression into one of seven categories (Anger, Disgust, Fear, Happiness, Sadness, Surprise, Neutral).

Of all the techniques applied, convolutional neural network achieved the maximum accuracy (~60 %).

Link to code.

Implemented dropout and dropconnect regularization techniques in Theano. Tried to get results from the Regularization of Neural Networks using DropConnect paper. Link to paper.

Link to code.

Using the data set of approximately 13000 essays from kaggle.com, an automated essay scoring system was constructed.

Textmining and NLTK libraries were used for text preprocessing and feature extraction.
1. Vocabulary Measure – A measure of quality of vocabulary used in the essay.
2. Numerical Features – Features such as total word count per essay, average word length per essay, sentence count indicate language fluency and dexterity. The essays were tokenized and split using python utilities.
3. Parts of Speech count – Various parts-of-speech such as nouns, adjectives, verbs and adverbs are good indicator of quality of essay. Features extracted using python NLTK part-of-speech tagger.
4. Orthography – Correct work spelling indicates command over language and facility of use. Extracted the number of spelling errors per essay, using PyEnchant 1.6.5 library.
5. Structure and Organization – Punctuations’ are good indicator of a well structured and organized essays. Punctuation based features were extracted using regular expressions matching from individual essays after tokenizing.

Learning Models – Regression with least square error works the best with accuracy of approximately 75-80% on the testing data.

The model can be improved by using more complex features. More advanced NLP features and features that are grammar and usage specific can further enhance the prediction model.

A score is provided to the courses based on the sentimental analysis of the course reviews. The reviews are obtained from CULPA - http://culpa.info/.

The web application for the project has been developed. Link

Backpropagation algorithm is one of the important methods to train an artificial neural network. But the convergence rate for Backpropagation algorithm is poor. Therefore very high number of epochs are required for the results to be accurate. Hence, the bottle neck of the system implementing backpropagation algorithm is the time needed to process it. A solution for this problem is to have a faster processor which would surely decrease the time required. Another alternative is to use FPGA (Field Programmable Gate Array) which works as a dedicated processor but can be easily reconfigured.

As a part of the project we have implemented the backpropagation algorithm in VHDL. We have constructed different modules for the hidden nodes and output nodes which gives us the flexibility of constructing any network. We have verified the accuracy and correctness of our implementation by simulating the XOR problem as a small dataset problem and Fisher Iris problem as a large dataset problem. The XOR problem is also implemented on the FPGA to check the working of the network.

The time taken to solve Xor problem and Fisher Iris problem was 3ms and 17ms on FPGA whereas the same implementation on python 2.7 took 0.405 sec and 0.723 sec.

Testing Accuracy for Fisher Iris – 93% (on FPGA) and 96.67% (on Python).

The device projects keyboard on a flat surface using laser technology which can then be used to type on the laptop and other electronic devices.
Working
It consists of 3 main modules; illumination, sensing and projection module.
Illumination - Illumination module consists of a 650 nm class II line laser which projects a beam of red light, at an angle of 120 degrees, above and parallel to the surface.
Sensing - Sensing module consists of a webcam with a built-in CMOS sensor. It detects the light reflected from the user’s finger when crossing through the red beam of light.
Projection - Projection module consists of a self-made projector, only used to project an image of the keyboard onto a flat surface.
Image processing
The image processing was done on Python 2.7 using the OpenCV library. Images are captured continuously and threshold operation is performed on them to retain only the red color pixels. By taking the average of the coordinates of the red color pixels, the key pressed is obtained. A sleep break of 0.5 sec is designated which limits the typing speed to 120 characters per min.

I had designed a simple CPU which can perform basic arithmetic, logical, data-transfer, and control-flow operations from the basic gates in VHDL.
CPU specifications: -
CPU External Busses: 16-bit external data bus and a 13- bit address bus.
General Purpose Registers: a 16-bit register: accumulator (acc).
Memory Organization: 2 sets of memory - Instruction memory and Data memory.
Data Memory: Capable of addressing 8192 words; each word has a 16 bit width.
Instruction Format: Each instruction is a 16-bit word and occupies a memory word. Operands: An Explicit operand (the memory location whose address is specified in the instruction) and an implicit operand (accumulator). The CPU has a total of 8 instructions, divided into three classes:
•Arithmetic-logical instructions (ADD, SUB, AND, and NOT)
• Data-transfer instructions (LDA, STA)
• Control-flow instructions (JMP, JZ)
Addressing Mode: Direct Addressing

In this project, we had assembled the following circuit which can be connected to any of the home appliances (lamp, fan, radio, etc) to turn it on/off from a TV or DVD remote control.
Major components of the circuit are IR receiver module TSOP1738 and decade counter-IC CD4017.
Working:
1. Initially the count of decade counter is zero, O0 at pin 3 is high. Hence, LED1 (red) connected to the pin 3 glows to indicate the off state of the appliance.
2. The 38 KHz infrared (IR) rays generated by the remote control are received by IR receiver module TSOP1738 of the circuit. The output of TSOP1738 is amplified by the transistor T1. The amplified signal is fed to clock pin 14 of decade counter IC CD4017. The count of the decade counter increases to one. Therefore, O0 at pin 3 goes low and O1 at pin 2 goes high. Hence, LED2 (green) connected to the pin 3 glows indicating the off state of the appliance.
3. Transistor T2 (BC548) connected to pin 2 of IC1 drives relay RL1. Diode 1N4007 (D1) acts as a freewheeling diode. The appliance to be controlled is connected between the pole of the relay and neutral terminal of mains. It gets connected to live terminal of AC mains via normally opened (N/O) contact when the relay energies. Hence, the output O1 turns the home appliance ON via relay.
4. When the next signal from the remote control is received, the count of the decade counter increase to two and output O2 at pin 4 goes high. As pin 4 is connected to the reset at pin 15, the decade counter is reset to the original state with the count equal to zero. Thus, returning to the off state.