Practical artificial intelligence through research

Research

Language Models

Reasoning

NLP

Deep Neural Networks

Convolutional Models

Sequence Models

Transformers

Deep Reinforcement Learning

Gradient Boosted Machines

PhD Thesis:

"Teaching Smaller Language Models to Generalise to Unseen Compositional Questions"

https://researchspace.auckland.ac.nz/handle/2292/70611

Code, Data and Models accessable via Github:

https://github.com/timhartill/unseen_questions

Main Paper Link:

https://openreview.net/forum?id=d4Vr6E0jjm

Other Papers:

"Do Smaller Language Models Answer Contextualised Questions Through Memorisation Or Generalisation?"

https://arxiv.org/abs/2311.12337

"Answering Unseen Questions With Smaller Language Models Using Rationale Generation and Dense Retrieval"

https://arxiv.org/abs/2308.04711

Other Projects:

Convolutional neural network for Pneumonia classification and localisation.

Human Protein Atlas - Convolutional Neural Network for Protein localisation on sub-cellular entities in human cells.

Sequence modelling of an acoustic signal for earthquake prediction.

Localising objects in three dimensions using 3D bounding boxes.

Birds eye view of lidar data from autonomous vehicle.

Open Source

SensorMap - A realtime updated geographical map of detected objects. Typically object detections would come from AI models operating on multiple real-time video camera streams, but IoT streams or any other sensor types that are relevant to displaying on a spatial map would work.

Using Homography, Hierarchical clustering and Linear Assignment algorithms to translate objects detected from multiple multiple camera feeds onto a realtime updated 2D Map.

Github link to SensorMap: github.com/timhartill/sensormap

ConceptRules

A dataset for testing deductive reasoning properties in Language Models: ConceptRules Version 2