A Natural Language Question Answering System on Mathematics

Amy Wei


Our system parses and answers mathematical questions that are written in an intuitive, natural language format. Questions include: “Is (0.0,0.0) (0.0,1.0) (1.0,1.0) an acute triangle,” and “Are (0.0,0.0) (0.0,1.0) and (1.0,0.0) (1.0,1.0) perpendicular.” In order to answer these questions, we defined a grammar structure fitted with a personalized math vocabulary, sorted into categories, such as integers, floating point numbers, coordinate points, etc. For example, our system parses the question “Is 100 even” as a verb, integer, and adjective. Recognizing the sentence structure, it calls a corresponding predicate to determine the final result. If the user inputs a sentence the system does not recognize, or misspells a word, then the system will return ‘Invalid input.’ In summary: our system parses the user’s question based on a defined grammar structure, and then calls a corresponding predicate to return a final result. Through this process, it can understand and answer mathematical questions written in natural language.


Electronics: Thermal Sensors and Operational Amplifiers  

Kanishk Tihaiya

This project was meant to serve as an introduction to the field of electrical engineering. The main learnings include how different components serve a role in the overall circuit. Components such as resistors, capacitors, power supplies, and operational amplifiers were used in the final circuit. Additionally, a special type of resistor, known as a thermistor, was also used in the circuit. When the temperature of the thermistor knob increases, the resistance decreases, allowing for a higher voltage to pass through. The final circuit was composed of two initially separate parts. One part was the thermistor circuit, which used the thermistor. The other part was the oscillating circuit, which was designed to switch (oscillate) between two states. The output of the total circuit included an LED, which is supposed to blink when the thermistor knob is being touched. In the future, this circuit can have many applications, including a temperature alarm system. For example, a speaker can be added so there can be an alarm sound when the temperature of the thermistor reaches a certain level. A laser can also be incorporated into the circuit as a way to transmit a signal. Overall, this was meant to be introductory and working on the project resulted in a gain of understanding about the various components, as well as core electrical engineering concepts.


Emotion Classifier

Daniel Zhou

The objective of this project was to use machine learning to train a model to detect emotions. This model was trained to be able to sort images into 7 classes, these being “angry”, “disgust”, “fear”, “happy”, “neutral”, “sad”, and “surprise”. This project used the FER2013 dataset (facial emotion recognition) to train the model. This dataset contained 28,709 images that were split using an 80-20 ratio for training and validation. The model itself was made using keras.sequential and the Adam optimizer. The model utilized dropout and data augmentation to prevent overfitting and was trained with 25 epochs. The best model was saved to use with a cascade face detector to be able to detect emotions from a video input. As of currently, the model is still a work in progress. The model is currently around 55% accurate and while it works with cascade, it cannot currently show the classification of the input on the screen. Plans for the future include improving the accuracy and rectifying the classification issue.


On the improvement of COVID-19 pneumonia classification on chest radiographs through transfer learning

Grace Liu

A base transfer learning model is developed to classify the image data set from the SIIM-FISABIO-RSNA COVID-19 Detection competition on Kaggle.  Due to the input data mismatch between the medical images (grey) and most pre-trained models (RGB), different methods are explored to complement the two missing channels. Based on the tests, simply broadcasting the data is not the best choice. In addition, different data augmentation functions are tested to improve accuracy and reduce overfitting. Surprisingly, none of these functions have a significant impact. This could be explained by the fact that these images are already augmented for the competition dataset. Next, we propose a channel-based data augmentation scheme. This method helps to improve the accuracy by 3 percent. After that, different transfer learning models are investigated. InceptionResNetV2 and EfficientB7 were found to have the best performance. Finally, we propose an ensembled transfer learning model that is based on multiple transfer learning models.  All the features are combined before the final classification. It is shown that this method further improves the model accuracy by 5 percent compared to InceptionResNetV2 and EfficientB7 separately.  An overall accuracy of 81% was achieved through the processes described above