This work extrapolates from research into Adversarial Patches on images, a technique which fools a neural network into misclassifying a seemingly unadulterated image by changing a small area of pixel values. When our patches are applied to the spectrogram of a sound file and converted back to audio, the files sound almost identical to the original, but are misclassified by a well trained neural network. The adversarial sounds in the files with the patches applied range from undetectable to minor static behind the spoken digit. 

This work uses reinforcement learning to train a robot to encounter a box using only a touch sensor on the front of the arm. The robot collects data-points about where the object is in space and, in my ongoing work, it will be able to use a neural network to classify which object it encountered entirely by touch. 

In Progress

When given a pointcloud of a human body, I have designed a version of PointNet++ to detect certain features on the human body. 

By reviewing criminals' previous history, this work uses various machine learning algorithms to effectively determine whether a criminal is likely to recommit a crime. I used a variety of algorithms with the aim of determining whether a more simple algorithm could predict recidivism just as well as a more complex model.  

The model outlined in this work uses the outputs of electrodes in a person who is told to move their right or left arm and is able to predict which direction the patient is intending on moving based solely off brain waves. 

This work reconstructs images that have been corrupted with the goal of regaining as much information as possible based off the superposition of waves within the image.

By separating and analysing the phonemes in spoken Arabic digits, the algorithm is able to recognise spoken digits.

This algorithm detects and determines which digits have been handwritten in an image.