Guiding the use of AI in research at Curtin

8 Mar 2024

The Curtin AI in Research Group, consisting of members from across the University, formed in March 2024 and is now calling for researchers to nominate their projects as potential candidates for AI acceleration.

The group will assess the expressions of interest based on the potential benefit of AI to the research described, and contact the relevant researcher(s) soon thereafter.

We aim to have selected a list of suitable research projects by the end of April 2024, and can begin AI support from May.

Follow the link on this page to express your interest.

13 Dec 2023

What does the rapid rise of AI mean for the human side of creativity, art and culture? David Karsten is joined by Associate Professors Rachel Robertson and Susanna Castleden to discuss.

Using around 200 million pages from local newspapers in the United States, collected over a time period of 170 years, these researchers used AI to measure economic sentiment.

This novel approach to extracting sentiment from millions of pages outperformed all previous work, presenting a deeper sentiment analysis as well as being able to process many more pages than ever completed.

Read their research paper

How did they do it?

The researchers used a combination of two techniques: Word2Vec and Sentprop.

Word2Vec is a method that captures the meaning of words and phrases within the context it is used. This allows for the system to understand the nuances of the language used in the newspaper at a deeper level compared to just positive or negative words.

The next technique, Sentprop, is a way to measure the overall sentiment of the newspaper or other text documents. This method uses a continuous sentiment measurement based on the Word2Vec results rather than the baseline word-count techniques.

Why was it not possible previously?

Previous work in this field used very basic techniques of counting preset positive or negative words, therefore not capturing an accurate sentiment analysis.

With the very recent improvements in language models in accuracy and speed it is now possible to process millions of documents with contextually aware sentiment.

How is it applicable to you?

With the improvements in the AI field, sentiment analysis has become a simple task that can process millions of documents and present accurate insights to the information provided.

This technique can be used on any set of text documents, not just news articles, and therefore can be very applicable to you as a researcher.

Similar research areas

• Survey responses
• Patient feedback in the healthcare system
• Social media discussions in many fields
• Gauging public opinion

In December 2023, researchers created Coscientist, an autonomous AI agent capable of researching, designing, executing, and analysing chemical research experiments. Researchers can simply give the Coscientist a prompt, such as “perform multiple Suzuki reactions”, and it will begin planning experiments. Coscientist is capable of browsing the web, interacting directly with lab machinery, and can even execute Python code to analyse results.

Autonomous agents such as Coscientist represent enormous potential benefits for researchers in all fields, who can now develop their own agents to have additional researchers who work 24/7 and are able to handle experiments from the initial planning to helping present final reports.

Read their research paper

How did they do it?

Coscientist is not a singular AI, but rather a collection of AI agents who interface with each other to research and execute experiments.

The core of Coscientist is the PLANNER module, a language model (GPT-4) which serves as the brain of Co-scientist, deciding which other agents to call.

There are 4 primary commands it can execute:

• GOOGLE – Uses a web searcher language model.
• DOCUMENTATION – How to interact with lab equipment.
• PYTHON – Executes python code for analysis.
• EXPERIMENT – Execute experiments on physical lab equipment over the internet.

Why was it not possible previously?

Recent breakthroughs with large language models (LLMs) such as ChatGPT allow an AI to understand plain English text and follow complicated instructions.

Researchers no longer need to be technical to work alongside AI and can simply instruct it in plain English.

How is this applicable to you?

The LLMs used in this project are not specifically trained for chemical research and a similar approach could be used for nearly any research field.

Autonomous agents are capable of planning, researching, and analysing results that can be beneficial to nearly every research field.

Similar research areas

• Data Analysis
• Research Paper Discovery
• Idea Generation
• Quality Control and Compliance

Using cutting edge language models such as ChatGPT and a technique called OCR (Optical Character Recognition) these researchers were able to automatically extract specific content from hard copies of histopathology reports and be put into a specified digital form.

Read their research paper

How did they do it?

The reports given were photos of handwritten or printed documents and therefore the first step was to convert the images into text using OCR.

Using a large language model, such as ChatGPT, it is possible to provide the model with an empty template and all extracted text and instruct it to fill in the template.

Putting these techniques together in sequence can automatically create structured digital templates from a series of handwritten documents.

Why was it not possible previously?

Large language models are fairly new advances in the scientific community and therefore handling unstructured data to fill in a set template was never possible.

Text extraction from images has been around for a while and has continued to improve, however language models now provide us with a way to turn this unstructured text into a structured digital format that we can use.

How is this applicable to you?

Being able to convert large amounts of handwritten text to a digital format can speed up the process of having to manually copy down information.

Using language models to automatically go through a large number of documents and find all relevant information is very beneficial and can save a researcher a lot of time in collecting data.

Similar research areas

• Digitising historical documents
• Extracting useful information from forms, surveys, questionnaires
• Extracting diagnoses, names and treatments from health records to aid in clinical research
• Extracting clauses and legal concepts from legal cases
• Extracting key concepts to get a quick understanding of a large number of research papers