Text: Laura Weißert
While Dr Aishwarya Venkataramanan was writing her master’s thesis at the University of Lorraine in France, she discovered her passion for machine learning—and for research. She had previously studied electrical engineering in her home country of India. »I love combining logical thinking with creativity and problem-solving,« she recounts. In her doctoral thesis, completed between 2020 and 2023, she investigated how microalgae in water can be automatically identified using deep learning models.
Even during that time, she had been closely following the research conducted at the University of Jena, in the Computer Vision Group led by Prof. Joachim Denzler. »I had a feeling that, with my expertise, I could make a genuine contribution here,« recalls the 29-year-old. Since June 2024, she has been conducting her own research here as a postdoctoral researcher, focusing on methods for quantifying uncertainties in deep learning models.
ChatGPT now processes more than 2.5 billion user queries on a daily basis—and the number is rising. However, it is not uncommon for large language models such as ChatGPT to be somewhat liberal with the facts and produce incorrect information. This phenomenon is also known as hallucinating. It is often not at all easy for users to assess whether an answer is correct. This is precisely where Dr Aishwarya Venkataramanan’s research comes in. Her aim is to enable these models to better assess their own reliability and to communicate this effectively.
»Models like these are trained on a specific type of dataset and learn to make a prediction for a specific input,« explains Venkataramanan. »These predictions tend to sound very convincing, even when they are incorrect.« In the worst-case scenario, this excessive »self-confidence« on the part of AI can have serious consequences—for example, when AI is used in medicine or in autonomous driving. Venkataramanan is researching methods by which to quantify uncertainty in AI predictions: A model should not only provide a solution, but also indicate how reliable it is.
Uncertainty in so-called vision-language models is a key focus of her work. These AI systems combine images and text. However, language can be ambiguous. Venkataramanan cites the word »bat« as an example, which can refer either to the small flying mammal or to a piece of sporting equipment. »When a model is presented with data that is ambiguous or that it has not encountered before, it should warn users not to trust the prediction,« explains Venkataramanan.
In her current research, she is developing methods for quantifying uncertainties in image-language models used for detailed classification tasks. These tasks require that a distinction be made between categories that appear almost identical—such as similar species of birds, plant species or visually similar objects. Such subtle distinctions are often difficult, even for experts, and frequently arise in applications such as automated bio-monitoring, where large numbers of images of animals and plants must be analysed. By developing uncertainty-conscious learning methods for these systems, Venkataramanan enables them to recognize when their predictions may be unreliable.
Artificial Intelligence is a rapidly developing field. We require new perspectives from a variety of people, including women.
Aishwarya Venkataramanan
Despite missing her family and friends in India—as well as the food and culture—Venkataramanan feels very much at home in Jena: »The city offers a fantastic mix of vibrant academic community and a relaxed environment that’s easy to explore on foot.« Thanks to its manageable size, you feel well-connected and at home here, yet it still has plenty of energy and much to offer in terms of cultural activities.
When not conducting her research, Venkataramanan likes to stay active and spend time outdoors. »Hiking and walking are my favourite hobbies and, living in Jena, I have some beautiful landscapes right on my doorstep.« When she needs a break from problem-solving, the researcher turns to reading, which often inspires new ideas for her work when she returns to her desk.
In many settings, computer science and AI are still largely male-dominated; something that Aishwarya Venkataramanan views as a challenge—both in India and in Europe. »There is a lot of potential for women in artificial intelligence—it involves programming, creativity, and logical thinking,« she remarks. However, attracting talent requires targeted encouragement and visibility. International conferences are increasingly turning to workshops designed to help researchers network and make it easier for them to get started.
However, it is crucial that this process starts at school. With events such as the MINT Festival at the University of Jena or the Long Night of Sciences, Venkataramanan sees promising signs in this regard. »Artificial Intelligence is a rapidly developing field. We require new perspectives from a variety of people, including women.«
