Lucy Low

Deep Neural Network Autoencoder for Data Compression in High Energy Physics at CERN ATLAs

Large Hadron Collider is the world's largest and highest-energy particle collider with the ATLAS detector generating 1 petabyte of raw data and 40 million packets of protons colliding every second. High collision rate 20 MHz means not all events can be stored. A particle physics trigger system selects specific events. The goal is to reduce the size of the data by engineering a compression algorithm for the trigger system.

Autoencoder neural networks were used for data compression and anomaly detection. The two part encoder and decoder system compresses hadron jet event data from 4 to 3 variables. It was trained over a dataset to encode the inputs into a smaller memory space with PyTorch, FastAI Library, ROOT Data Analysis Framework, and CERN ATLAS Docker images. Analysis includes plots and graphs to explain the concepts of invariant mass, purity selection, trigger efficiency, and hadron event reconstruction.

Find out more

Yeezy Taught Me Text Generation - Train Next Character Prediction using Long Short Term Memory Model (LSTM)

LSTM is one of the most commercial AI achievement used in time series prediction, speech recognition, music learning, handwriting recognition, and sign language translations. Yeezy Taught Me is a web application for AI model training and text generation. Trained LSTM model to generate text based on patterns in a given text corpus. Yeezy's Model makes probability predictions of the character that follows the input sequence. Process is repeated in order to generate a character sequence of a given length hence the "text generation" part of the project.

Input file takes input text data. Model saved in IndexedDB database. User inputs Yeezy's model parameters to train model: [model layer size, number of epochs, examples per epoch, batch size, validation split, or learning rate].

Find out more

Stochastic SoundCloud - Lucy’s New Mozart Mixtape. Machine Learning Generative Music using RNN LSTMs

Stochastic SoundCloud uses machine learning to generate melodies as music is an art with a temporal and hierarchical structure. The output musical state is partially determined by the preceding musical state where the concrete musical state n+2 follows after the state n+1. Piano roll representation is a music storage data type where a music piece is represented by a score-like binary valued (0 XOR 1) matrix representing music notes over different time step. Piano roll of each bar and track for generated data is represented as a fixed-size matrix for M tracks.

Music generation with three novel recurrent neural networks (basic, lookback, and attention RNNs) using Magenta from Google's Tensorflow AI. Lookback and Attention RNNs are proposed to tackle the problem of creating the melody’s long-term structure. It was fed with a chord sequence and outputs a Prediction Matrix, which was transformed into a piano roll matrix and into a melody MIDI file. 10 stochastically generated "output.mid" music files are composed and opened up on Mac's Garageband.

Find out more

B00M-CANCER - Quantum Support Vector Machine (QSVM) for Cancer Medical Imaging

Lung cancer is the largest cancer making up 25% of cancer cases but receives least amount of funding (6%). The lung cancer five-year survival rate (19%) is lower than many other leading cancers, like colorectal cancer (65%), breast cancer (90%) and prostate cancer (98%). B00M-CANCER is a quantum machine learning lung cancer screening soluton with medical imaging for cancer diagnosis to help lung cancer patients. For radiologists, the limits of human vision causes diagnostic errors (overlooking tiny malignant lesions) with 35% of lung nodules being missed at the first initial radiology screening report.

Developed a Quantum Support Vector Machine (QSVM) for Anomaly Detection which translates the classical medical data into quantum states. The kernel-based SVM is expressed as a quadratic programming (QP) problem where energy function was minimized using Quadratic Unconstrained Binary Optimization (QUBO) and the discrete binary solution solved on D-Wave 2000Q Quantum Annealer. The quantum optimization algorithm solves the problem in logarithmic order compared to polynomic in classical making it more efficient with the increase in the number of qubits with lower computational cost.

Find out more

Covid Control - Long Short Term Memory (LSTM) Predictor and Reinforcement Learning (RL) Prescription Prediction Model

The pandemic caused regional governments, communities, and organizations to transform their operations, technology, and business models. Covid Control is a two-part machine learning optimization model that predicts the future 7-day moving average number of Covid19 daily cases for sequential decision making with weight parameters. This helps with intervention planning measures like lockdowns, social distancing, or the mandatory use of face masks. The first predictor model is a LSTM model trained on Oxford dataset. The eight parameters used to train the model included schools closing, workplace closing, cancel public events, restrictions on gatherings, close public transport, stay at home requirements, restrictions on internal movements travel between regions/cities, international travel controls. Each parameter was given an integer value from initial condition (no measures) to final condition (full measures). The previous 21 days of values for the eight NPIs were then fed into the action input for the prescription. The second prescription model uses a reinforcement learning agent to minimize the number of cases and outputs a set of actions. It takes data from the predictor as input, and outputs a set of actions to perform in each time step. The weights for each country were drawn from a uniform distribution within [0,1] and normalized to sum up to one. The predictions were averaged over a time period of 180 day to obtain the final objective number of covid cases for each country. This allowed reinforcement learning agents to make future prediction states of their environment by estimating the system state. In reward function search, the output of the predictor was the reward vector for Q value steps.

Find out more