A data scientist transitioned from an electronic communication engineer, Sahana Prabhu‘s research interests include diabetic retinopathy image analysis, meibomian image segmentation, emotion recognition via deep learning approaches, and retail analytics via RFID and stereo cameras. Analytics India Magazine caught up with Sahana, who is currently serving as a research scientist and technical architect at Bosch Engineering and Business Solutions, to understand her insights on deep learning, machine learning, etc.
AIM: As a research scientist, what factors do you believe contribute to your research success? Please throw some light on it for us.
Sahana Prabhu: Proper dataset collection that is representative of the problem to be solved is vital. Once the problem statement is formulated, one approach to innovating is to reduce the restrictions in previous research papers. An example of this is to add data augmentations to tackle variations in data from open settings instead of having a controlled setting for capturing data. Another valuable research method is combining two related problems and developing a comprehensive framework that optimises both solutions. For instance, one of my papers involves a framework to solve image matting and super-resolution simultaneously. Maintaining a daily log of useful web links and attempted experiments helps in structuring research work. Displaying the results of the intermediate steps of the algorithm and checking if it is along expected lines reduces erroneous assumptions.
AIM: Is the problem of unsupervised learning more difficult to solve? What is your view?
Sahana Prabhu: Unsupervised learning does not have pre-existing patterns to learn from in labelled training data. Nevertheless, it is necessary to harness the potential of huge unlabeled data and recognise unknown insights without the limitations of human bias. Since purely unsupervised learning is difficult, an approach to dealing with unlabeled data is self-supervised learning, an active area of research.
AIM: How can larger firms that invest in deep learning ensure that their efforts benefit others in the field?
Sahana Prabhu: Larger firms have more capacity to do experimental research for problems that are not obvious.
They can achieve greater operational efficiency from their proprietary data for real-world problems. Their existing customer network can look into various areas and formulate ideas to improve automation processes. Many large firms provide start-up incubation support, such as the Bosch Accelerator program, which goes a long way to further progress in this field.
AIM: What applications of deep learning excite you the most right now or soon?
Sahana Prabhu: Material recognition using texture analysis is a topic I am working on now, and it has applications in track-and-trace across automotive manufacturing, food produce, and pharmaceutical domains. Explainability is another research area emerging for many domains such as medical imaging, assisted driving, and manufacturing defect detection.
AIM: Which industry do you believe will be most disrupted in the future by deep learning?
Sahana Prabhu: Autonomous retail and autonomous driving are two emerging industries that will be made possible in the future. Smart automation of retail stores for self-checkout, tracking the flow of customers, and monitoring inventory can be facilitated by deep learning, and this is already under the experimental phase. Several automotive companies, including Bosch, have made significant advancements in autonomous driving. In addition, deep learning for computer vision, including semantic segmentation and image retrieval, is used extensively for assisted driving applications.
AIM: What are your opinions on the recent rise in interest in deep learning in the media?
Sahana Prabhu: There are two sides to it – while it has gained proper attention for automation in some domains such as manufacturing, it is also necessary to recognize its limitations. For example, deep learning can be an additional aid for routine screening in many medical image-based diagnosis tasks, but medical image-based diagnosis cannot fully eliminate doctor supervision. Moreover, there are many applications wherein conventional computer vision methods already provide the required accuracy for large-scale deployment, and deep learning may not be useful. There is also a perception that deep learning can give better predictions for cases where humans cannot, but deep learning gives only close to human accuracy in many applications.
AIM: What are some of the difficulties that someone new to machine learning might encounter? How should they approach them?
Sahana Prabhu: Information overload is a common problem that both newbies and researchers in machine learning face. It is tempting to apply machine learning for problems where conventional computer vision methods would suffice. We have a lot of online resources available, but it is important to focus on the problem and limit oneself to appropriate methods. New state-of-the-art algorithms keep getting added, and these can be found in paper listing sites such as “Papers with code” and “Awesome Deep Learning Resources”, rather than doing random searching on the internet.
AIM: Who is your role model in machine learning research?
Sahana Prabhu: Prof Andrew Ng has facilitated many research students (including myself) from related core fields such as computer vision to transition smoothly to machine learning. The way he straddles academic theory and industrial applications is inspiring.
AIM: Are there any research papers you think every data scientist should read, irrespective of whether they are just starting or have years of experience?
Sahana Prabhu: I will mention three breakthrough deep learning papers in computer vision for the three broad topics – classification, detection, and segmentation, respectively:
[1] Simonyan K, Zisserman A. “Very deep convolutional networks for large-scale image recognition.” 2014.
[2] Girshick R., et al. “Rich feature hierarchies for accurate object detection and semantic segmentation.” CVPR 2014.
[3] Ronneberger O., et al. “U-net: Convolutional networks for biomedical image segmentation.” MICCAI 2015.
