Unit 1 – Introduction to Machine Learning in Research
Assignment: 1
Question 1: Compare different ML paradigms (supervised, unsupervised, reinforcement)
with examples of how each supports hypothesis-driven research.
Answer: Machine learning is a branch of Artificial Intelligence that focuses on developing
models and algorithms that let computers learn from data without being explicitly
programmed for every task. In simple words, ML teaches systems to think and understand
like humans by learning from the data i.e. Machine Learning (ML) is a way of teaching
computers to learn from data and make decisions without being explicitly programmed for
every single step.
In research, ML methods serve as different kinds of “analytical engines” that help scientists
test their ideas (hypotheses) in novel ways.
1. Supervised Learning: This is the most common form of ML. It is like a student who is
learning from a teacher who has the answer key. You feed the system “labelled” data
(inputs and correct outputs) until it learns the relationship between them.
Helps with hypothesis-based research: Scientists use this when they have a specific
hypothesis about how one variable affects another.
Example: A medical researcher hypothesizes that certain markers in the blood can
predict a certain disease. They feed a model the records of thousands of patients in whom
the disease has already been diagnosed. The model learns that pattern, and researchers
test their hypotheses by seeing whether the model can accurately "predict" disease in a
new, never-before-seen group of patients.
2. Unsupervised Learning: In Unsupervised Learning the computer is given raw data and
asked to find interesting patterns or groups on its own. It's about exploring rather than
guessing.
Supporting hypothesis-based research: This is often used for "exploratory"
hypotheses. A researcher may suspect that a large category of things (such as
"consumers") actually has many different sub-categories that he or she has not yet
identified.
Example: A biologist has data on thousands of single cells, but does not know how
many different types of cells are present in a tissue sample. Using unsupervised
clustering, the model groups similar cells together. This supports the hypothesis that the
tissue is more complex than previously thought, leading to the discovery of new types of
cells.
3. Reinforcement Learning (RL): This method is based on 'trial and error' (learning by
trying). In this, an 'agent' interacts with an environment, and gets rewards for good
actions, and punishment for bad actions. Over time, this agent learns the best strategy to
maximize its total earnings (reward).
Support hypothesis-based research: This approach helps understand hypotheses related
to behaviour, optimization, or 'what-if' situations in dynamic systems.
Example: An engineer hypothesizes that a particular cooling strategy will reduce power
consumption in a data centre by 20%. They set up an RL agent to control the fans and
AC. As the agent tries out different settings and receives 'rewards' for lower electricity
bills, researchers are able to confirm whether their proposed strategy is actually the most
effective one.
References:
Chandra, V. (2021). Machine learning: A practitioner's approach. PHI Learning.
Chopra, R. (2024). Machine learning. Khanna Publishing House.
Geron, A. (2022). Hands-on machine learning with Scikit-Learn, Keras, and
TensorFlow. O'Reilly Media.
James, G., Witten, D., Hastie, T., & Tibshirani, R. (2021). An introduction to
statistical learning: With applications in R. Springer.
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction.
MIT Press.
Sridhar, S., & Vijayalakshmi, M. (2021). Machine learning. Oxford University
Press.
Question 2: Discuss current research challenges (bias, interpretability, scalability) and future
directions in ML-based research.
Answer: Although machine learning (ML) is a very powerful tool for modern research, it is
not a magic wand. As we become more dependent on these tools for scientific discovery, we
are facing some major “teething issues” that can harm the credibility of our results if we are
not careful.
Challenges of Current Research
1. Problem of Bias (Unbiasedness): The biggest problem is that ML models are only as
good as the data we give them. If a dataset contains long-established human biases—
such as race, gender, or socioeconomic biases—AI will learn them and amplify them.
This is dangerous in research, because it leads to "scientific" results that are
fundamentally wrong. For example, a health study that is based mostly on data from
one racial group may not work for other groups, leading to unequal treatment.
2. “Black box” (explainability): Many modern models, especially deep learning ones,
are overly complex. They can give you a correct estimate, but they can't tell you why
it happened. To a researcher, "the computer said so" is not a valid scientific
explanation. To turn a hypothesis into a real success, we need to understand the logic
behind it. Without the ability to interpret, we have risk making decisions based on
correlations that do not actually exist.
3. Scalability (cost and energy): As models grow larger, they require large amounts of
computing power and electricity. This creates a “digital divide,” where only wealthy
institutions can afford cutting-edge ML research. Furthermore, the environmental
impact of training these huge models is becoming a serious ethical concern for the
global scientific community.
Future Directions: The next phase of ML research is moving toward “Small Data”
and “Human-in-the-loop” systems. Instead of needing billions of data points,
researchers are looking for ways to let models learn from smaller, but higher-quality
datasets. There is also a push for Explainable AI (XAI)—tools specifically designed
to peer inside the “black box” so scientists can understand the reasoning behind a
result. Finally, we are seeing a shift towards Physics-Informed Neural Networks,
where AI is forced to obey the laws of nature (like gravity or thermodynamics) while
learning, ensuring that its discoveries are physically possible.
References
Chandra, V. (2021). Machine learning: A practitioner's approach. PHI Learning.
Dhar, P. (2020). The carbon footprint of machine learning. Nature Machine Intelligence,
2(8), 423-425.
Lipton, Z. C. (2018). The mythos of model interpretability. Communications of the ACM,
61(10), 36-43.
Subramanian, R. (2021). Ethics and bias in artificial intelligence. Pearson India.
Varma, S. (2023). Future directions in AI: From black boxes to glass boxes. Oxford
University Press.