MACHINE LEARNING
PROJECTS
Throughout my Computer Science courses, I expanded my programming proficiency by exploring diverse fields and mastering a range of libraries, enabling me to effectively navigate intricate problem-solving challenges. This portfolio features machine-learning projects that highlight my capacity to craft distinctive solutions through innovative approaches.
Computational Perception
Neural Network-Based Image Prediction from Bethge Model Neurons
Summary
During the Fall 2023 school semester, I completed Computational Perception (15786), a graduate-level Computer Science course at Carnegie Mellon University. During this time, I developed a project that applies neuronal stimuli preferences to recreate an image based on the relative activations of each neuron. I utilized the Bethge neuron-in-silicon model to get model neuron responses from a specific pre-processed image, then created FLPs and CNNs trained on 25000 pre-processed images to regenerate the target. I then used statistical analysis across 9000 pre-processed test images to determine the efficacy of each model.
Methods
The research methodology involves a dataset of 34,000 images that are pre-processed to better focus the model on the center of the image and make the algorithm more efficient. A publicly available convolutional neural network (CNN) is then employed to convert the images into neuronal responses.
​
Two information flows lead to four different models: Architecture 1, driven by pixel loss, and Architecture 2, driven by perceptual loss. Two neural network architectures are explored per architecture: a multi-layer perceptron for Models 1 and 2, and a CNN for Models 3 and 4. Data analysis includes pixel-level and neuron-level evaluations, comparing predicted and preprocessed images or neuronal responses. Various quantitative methods such as box plots, histograms, correlations, and mean squared error are employed to assess the models' effectiveness using both individual test images and aggregated results.
Libraries Used
Pytorch, Matplotlib, Scipy, Numpy, PIL
