How The World’s Best Scientists Use AI | Faheem Ullah | AFP 52

The following is based on a conversation with Faheem Ullah, an Assistant Professor in Computer Science at the University of Adelaide.

What is the core architectural difference between Machine Learning and Deep Learning regarding feature extraction?

In traditional Machine Learning algorithms (such as Naive Bayes, K-Nearest Neighbors, Decision Trees, or Random Forests), feature selection must be performed manually. Researchers must use statistical methods, such as Principal Component Analysis (PCA), to isolate relevant variables and discard noise before model training. Conversely, Deep Learning automates feature extraction. The deep neural network independently evaluates input data, automatically identifying and prioritizing significant features while discarding irrelevant ones without requiring manual statistical filtering.

How do hyperparameters and model parameters differ during neural network training?

Hyperparameters are configuration settings defined by the researcher prior to the execution of the training process. Examples include the learning rate, choice of activation functions, and total training epochs. In contrast, model parameters, specifically weights and biases, are internal variables that are dynamically updated and adjusted by the model during the training process. These weights reflect the predictive importance assigned to specific features based on the input data.

What is the function of a pilot study in automated literature reviews?

A pilot study serves as a validation mechanism to test the accuracy and comprehensiveness of an automated or AI-generated search string before running a full systematic review. Researchers compile a control group of known, highly relevant benchmark papers on a topic. The designed search string is then executed across databases (e.g., Google Scholar, Scopus). If the algorithm returns all control papers, the search query is validated; if benchmark papers are missing, the search string must be refined to eliminate search bias and prevent missing critical literature.

Cleaned Transcript

Jeroen Schreel: Welcome to Apple Finch Pudding, your gateway into the world of science. Today’s scientist is Faheem Ullah, an Assistant Professor in Computer Science at the University of Adelaide. His work focuses on artificial intelligence. Welcome, Faheem.

Faheem Ullah: Thank you. Thank you for having me.

Jeroen Schreel: Faheem, before we start, do you have a fun science fact for our listeners?

Faheem Ullah: One thing that is really fascinating to me is that the human brain consumes around 20% of the body’s energy while weighing less than 2% of total human body mass. It is interesting to see that this part of the human body consumes so much energy while being extremely flexible. You see some people who have done wonders in their life, and others who have barely used it. My point is that it is always good to utilize our brains to their maximum potential because they draw so much energy from our bodies.

Jeroen Schreel: That is completely true, and you have to use your brain a lot. You have spent years studying computer science and machine learning. Was there a specific trigger that led you into that type of research?

Faheem Ullah: My bachelor’s degree was in computer science, which I started around 2009 when computers were a rapidly emerging field. Many people suggested I enter medical studies, but I felt that computer science was the future. I completed my bachelor’s degree at a great institution, and my PhD naturally followed in the same trajectory, specializing in AI and software engineering. Looking back, I feel it was the right decision because AI and computing are now pervasive across almost every domain of life.

Jeroen Schreel: AI really is booming. For the listeners, we often hear the terms AI, Machine Learning, and Deep Learning. What is the structural difference between those three?

Faheem Ullah: Artificial Intelligence is the broad umbrella term. It can be compared to human intelligence in the sense that an intelligent system processes prior data to make optimal decisions. For example, studies show that AI models can predict if someone is at risk for a heart attack up to five years in advance by analyzing patterns in medical data. AI is the overarching domain that mimics human cognitive functions.

Machine Learning is a subset of AI. In machine learning, we feed structured data into a model so it can learn patterns and project future outcomes. For instance, by feeding five years of flight history into algorithms like Naive Bayes, K-Nearest Neighbors (KNN), Decision Trees, or Random Forests, the model can predict which flights are likely to be delayed tomorrow.

Deep Learning is a specialized subset of Machine Learning. In traditional machine learning, we must perform manual feature selection—identifying which variables matter and discarding irrelevant data. In deep learning, feature selection is automated. The model independently determines which features are relevant and discards the rest.

Jeroen Schreel: Could you give an example of how feature selection works?

Faheem Ullah: Imagine predicting student grades. We might evaluate four features: hours spent studying, class attendance, hours of sleep, and hours spent at the gym. A student’s grade correlates strongly with study hours and attendance, whereas sleep and gym time have far less direct statistical impact. In traditional machine learning, we must manually apply algorithms like Principal Component Analysis (PCA) to isolate study hours and attendance while discarding the rest. A deep learning model automatically identifies and weighs the critical features without human intervention.

Jeroen Schreel: So AI is the broader domain, Machine Learning is a subset, and Deep Learning is a specialized subset focused on automated feature selection. Does deep learning also assign different weights to these features?

Faheem Ullah: Absolutely. Machine learning and deep learning models rely on hyperparameters and parameters.

Hyperparameters are established by the researcher before training begins—such as setting the learning rate, choosing activation functions, or defining the number of epochs. Parameters, such as weights and biases, are calculated and adjusted dynamically by the model during the training process. The model assigns higher mathematical weight to features that directly influence the outcome (like study hours) and lower weight to secondary factors.

Jeroen Schreel: How will AI tools help researchers and scientists in their daily workflows?

Faheem Ullah: There are two primary schools of thought. One advocates for adopting AI tools to increase productivity and work smarter. The other opposes AI in education and research, arguing that reliance on these tools destroys the learning curve and restricts human creativity.

My view is that ignoring AI is like putting your head in the sand. AI is here to stay and will continue to evolve across healthcare, entertainment, sports, agriculture, and academia—largely driven by economic efficiency.

In education and research, we should absolutely embrace AI tools, provided we remain within the ethical boundaries set by institutions and publishers. Using AI to ghostwrite a research paper you don’t understand is unethical. However, using AI tools to streamline literature discovery is completely valid. When I started my PhD in 2017, constructing search strings and manually screening titles, abstracts, and full texts for a systematic review took a full week. If an AI tool can execute that initial discovery phase in one hour, you save six days that can be redirected toward critical reading and data analysis.

Jeroen Schreel: Is there a danger that relying on AI for literature reviews will cause us to miss critical papers?

Faheem Ullah: That risk exists, but it is also present when performing manual literature searches. Reviewers often point out missing studies regardless of the methodology.

To mitigate this risk, researchers conduct pilot studies. Before running a broad automated search on engines like Google Scholar or Scopus, you compile a benchmark set of 8 to 10 known, highly relevant papers on your topic. You then run your search string to verify if it successfully retrieves those target papers. If it does, your search parameter is validated.

Jeroen Schreel: AI can summarize thousands of papers in seconds. Do we lose research depth by relying on summaries rather than reading full papers?

Faheem Ullah: It depends on how the tool is used. I classify this into four scenarios:

1. Uploading 100 papers and asking AI to write the review: Unethical and counterproductive. You learn nothing.

2. Reading a one-page AI summary of 100 papers: Acceptable for getting a high-level overview of an unfamiliar field when in a rush, but insufficient for published literature reviews.

3. Targeted extraction: Asking AI to extract specific parameters—such as the exact datasets used across 100 papers—allows you to rapidly isolate data for rigorous manual synthesis. This is an effective use of AI.

4. Manual deep reading: Reading all 100 papers word-for-word yields the deepest understanding, but requires a significant time investment.

Jeroen Schreel: What about tools that convert research papers into audio or short summary videos?

Faheem Ullah: Audio translation tools that read the complete text aloud preserve the depth of the paper. However, relying solely on two-minute summary videos provides only surface-level information. It is like watching a one-minute movie trailer instead of the full two-hour film; you miss critical nuance.

Jeroen Schreel: If a groundbreaking paper in your field was generated entirely by AI, but the findings were completely accurate, how would you view it?

Faheem Ullah: Current AI generation is nowhere near capable of producing a top-tier, original research paper independently. AI text outputs remain visibly inferior to human academic writing.

Furthermore, publishers are establishing integrity frameworks requiring authors to formally disclose AI usage. Responsible academics must disclose these tools transparently.

Jeroen Schreel: How do you personally integrate AI into your academic workflow?

Faheem Ullah: I never use AI to write paper drafts. I use AI tools to generate presentation diagrams, draft slide outlines—which I then refine—and format quantitative data into complex visual plots.

However, cross-checking AI outputs is vital. Recently, I posted a social media query asking which of four professor profiles students would select as a PhD supervisor. I used AI to generate four avatar images for the options. I failed to inspect the output closely, and the AI generated four white male professors. The audience immediately focused on the implicit demographic bias of the images rather than the core academic question. As researchers, we are fully accountable for the outputs we publish, regardless of the tools used to create them.

Jeroen Schreel: Is literature discovery biased because AI tools struggle to index paywalled papers compared to open-access literature?

Faheem Ullah: While some dedicated academic search engines index metadata better than standard tools, most AI discovery systems search titles and abstracts rather than providing full PDF access. Once the tool identifies relevant papers based on metadata, the researcher must secure full text access through institutional subscriptions.

Jeroen Schreel: Do you use standard large language models like ChatGPT, or specialized tools?

Faheem Ullah: I rarely use general LLMs like ChatGPT, Claude, or Grok for literature mapping. I use specialized academic platforms like Paperpal, SciSpace, Consensus, or ResearchRabbit.

Because new tools launch constantly, I recommend trying free tiers to test specific features before committing to paid subscriptions.

Jeroen Schreel: Industry analysts recently noted that infrastructure costs for AI data centers are skyrocketing—projected at $600 billion for infrastructure that depreciates in four years. Some estimate consumers may eventually need to pay up to $100 a month to offset these costs. Do you expect consumer AI prices to rise significantly?

Faheem Ullah: Operating data centers requires vast financial and energy resources, but a price jump from $20 or $30 to $100 per month would cause massive user churn. Consumers would simply migrate to alternative open-source or cheaper models. Price increases will likely be incremental rather than drastic.

Jeroen Schreel: What about OpenAI integrating advertisements into ChatGPT to monetize free users?

Faheem Ullah: Ad-supported models are a logical step, similar to search engines or social media platforms. As long as sponsored results are explicitly tagged to distinguish paid placements from organic outputs, users can evaluate information objectively without inflating subscription costs.

Jeroen Schreel: If we have to manually fact-check every AI output, does it still save time?

Faheem Ullah: Fact-checking is the necessary tax we pay for the productivity gains AI provides. The required level of verification depends on the stakes. If you use AI to draft course materials for hundreds of students, rigorous verification is mandatory. If you are asking for personal product recommendations, occasional errors carry minimal risk.

Jeroen Schreel: Is AI rendering traditional educational assessment methods obsolete?

Faheem Ullah: Absolutely. Generative AI has disrupted traditional teaching and grading paradigms. Standard essay assignments are obsolete; ChatGPT can draft a high-quality essay instantly. Educational institutions must overhaul how they teach and evaluate critical thinking, moving away from easily automated tasks.

Jeroen Schreel: As AI content floods the internet, AI models are increasingly trained on synthetic, AI-generated data. Does this feedback loop degrade model quality?

Faheem Ullah: While recursive data loops pose risks, digital data logging is growing exponentially. User interactions, clicks, and workflows are continuously logged, providing vast streams of human behavioral data to refine future iterations.

Jeroen Schreel: Do you use AI to write your social media posts?

Faheem Ullah: Never. Social media algorithms actively detect and penalize AI-generated text. If an account posts automated low-effort content, platforms suppress its reach to prevent feed spam. Original, human-written content performs much better organically.

Jeroen Schreel: Can researchers use AI for science communication without algorithm penalties?

Faheem Ullah: Yes, by using AI for initial formatting—such as converting a published paper into an infographic outline—and then manually editing the design and text. If your goal is broad algorithmic reach, relying entirely on raw AI output will backfire. There are no shortcuts to authentic engagement.

Jeroen Schreel: Do you have a final take-home message for our listeners?

Faheem Ullah: AI tools can dramatically increase academic research productivity. We should embrace these tools, provided we operate transparently and within clear ethical boundaries.

Jeroen Schreel: I want to thank Faheem Ullah for sharing his expertise. Join us again for the next episode of Apple Finch Pudding.