PhD Poster Presentation | Identifying Human Actions Using Deep Graph Neural Network

Abstract: Observing human actions is necessary for practical reasons like surveillance, physiology and psychology. The modern-day approach to achieve this task is using machine learning models, which are trained on a vast amount of video data. In reaching for the promise of a lightweight model, graph-based machine learning algorithms which use a human skeletal graph, are proposed as comparable alternatives to existing RGB-based models. Our recent work
introduces a position encoding scheme called NAPE, which improves the embedding power of a graph neural network model. This results in a 1% increase in our human action recognition model on the NTU-RGB D benchmark dataset.

[Ordering of the human skeletal graph using the NAPE algorithm & Learned Binary vectors for each node in the human skeletal graph]

You can read more details about the NAPE algorithm, and how the generated position encoding affects the action recognition performance of our model.

MSc Individual Project | Video Representation Learning with Graph Neural Network, Using Data-Dependent Edge Weights

• The model takes in video data that has been converted into a graph. Each node of the graph is some extracted feature from a frame in the video. The model aimed to learn the weights of the adjacency matrix for the graph created, which was hypothesized to improve the embedding of the video data. The RML datasetLink opens in a new window on emotion recognition was used to train the model. The facial landmark points on each video frame were extracted and used as node features.
• Supervised by Dr Hongkai WenLink opens in a new window and Dr Tanaya GuhaLink opens in a new window.

[Heatmap of the learned adjacency matrix for some sample videos]

MSc Research Group Project | Modelling substantia-nigra neurons to quantify the effects of alpha-synuclein in Parkinson’s disease

• The toxic protein, α-synuclein, is thought to negatively impact the substantia-nigra neurons, causing the degenerative effects of Parkinson’s disease. In a recent study, experimental data was collected from substantia-nigra neurons which had been exposed to α-synuclein. We created dynamic I-V curves from this experimental data, in order to quantify the current-voltage relationship, and it was discovered that the exponential integrate-and-fire model provided an excellent fit to the cell which had been exposed to α-synuclein. An EIF model was used to fit the experimental data. Also, an LSTM model was trained to capture the dynamics in the experimental data.
• Supervised by Professor Magnus RichardsonLink opens in a new window.
• Fellow group members: Charles HepburnLink opens in a new window and Jack O'ConnorLink opens in a new window.

[Investigation into the presence of an adaptation current.]

[Long short-term memory network (LSTM) training and simulation data.]

MSc Individual Project | Particle Trajectory in Solid-body Vortex and Source-Sink Pair Flow Fields Using the Maxey-Riley Equation

• The question of how a particle moves in a fluid is the primary focus of this dissertation. The flow fields of interest are solid-body vortex and source-sink pair. Experiments reveal that buoyant particles would act as fluid tracers by following the fluid at each point with equal velocity. But for a particle with density greater (or lesser) than that of the fluid, the answer is not glaring for both flow fields of interest. For this, the Maxey-Riley equation is introduced, which describes the motion of particles in a fluid and is used to answer the above question.
  A mathematical model for an aerosol extractor (from COVID-19 patients) is designed using the solution of the Maxey-Riley equation for a source-sink pair flow field. And the effect of gravity on the aerosol extractor is considered alongside the effect of a constant flow stream.
• Supervised by Dr Cathal CumminsLink opens in a new window.

[(a) The trajectory (r(t), θ(t)) of a bubble (light particle) initially at rest at the edge of the teacup spiralling towards the centre of the stirred tea.]

[(b) Long term behaviour of the radial and angular velocity of the bubble]

[(a) The trajectories of particles less dense than the fluid (light particles) with a background source-sink pair flow field opposed by a laminar flow, but captured into the sink.]

[(b) The trajectories of aerosols from a breathing patient being captured by an aerosol extractor. The flow is influenced by gravity and a laminar flow which is in the direction of the sink.]

[(a) The trajectories of particles less dense than the fluid (light particles) with a background source-sink pair flow field opposed by a laminar flow.]

[(b) The trajectories of particles denser than the fluid (heavy particles) with a background source-sink pair flow field influenced by gravity.]

[(c) The trajectories of particles less dense than the fluid (light particles) with a background source-sink pair flow field.]

BSc Project | An Implementation of the Rivest–Shamir–Adleman (RSA) Public Key Cryptosystem (PKC)

• Cryptography has been considered in past times as an art, rather than a science. Also, most cryptosystems developed in past times are symmetric. This project work introduces the RSA public key cryptosystem which is asymmetric and considers the mathematical theorems upon which it is built, also providing their proofs and showing the logical sequence of how they emanate from each other. This work also provides the algorithm for the RSA, showing why it works and it's security-why it is not susceptible to just any form of attack. This work is also a practical implementation of the RSA by using it to encrypt (and decrypt) a text (Psalm 91: 1-2). In addition to the encryption, this project showed how the RSA can be adopted to the concept of digital signature. Lastly, a computer program was tailored to carry out both the encryption and decryption processes.
• Supervised by Ugorji Chimezie.

[Text to be encrypted using the tailored RSA algorithm (written in C++).]

[The resulting encrypted text.]

[A snapshot of the console window where the algorithm was implemented.]

[The decrypted text with a slight error at the end of the text due to an overflow in the computed number and its corresponding ASCII character.]