Deep Learning

Saturday, 16 April, 2:00pm-5:45pm


  • Debdoot Sheet (Indian Institute of Technology Kharagpur, Kharagpur, India)

Topic and background

Deep learning is a genre of machine learning algorithms that attempt to solve tasks by learning abstraction in data following a stratified description paradigm using non-linear transformation architectures. When put in simple terms, say you want to make the machine recognize Mr. X standing in front of Mt. E on an image; this task is a stratified or hierarchical recognition task. At the base of the recognition pyramid would be kernels which can discriminate flats, lines, curves, sharp angles, color; higher up will be kernels which use this information to discriminate body parts, trees, natural scenery, clouds, etc.; higher up will use this knowledge to recognize humans, animals, mountains, etc.; and higher up will learn to recognize Mr. X and Mt. E and finally the apex lexical synthesizer module would say that Mr. X is standing in front of Mt. E. Deep learning is all about how you make machines synthesize this hierarchical logic and also learn these representative kernels all by itself.

Deep learning has been extensively used to efficiently solve and provide state of art solutions to problems like handwritten character recognition, speech recognition, lexical ordered speech synthesis, object and product recognition, image retrieval, content filtering, product visibility tracking, computational medical imaging. The interesting fact about deep learning is its ability to learn prime representations from unlabelled big-data abundantly available in biomedical and biological imaging, and fine tune to adapt to a specific problem solving using very limited labeled data for the purpose of classification.


This tutorial will focus on understanding the inherent hierarchy in solving most of the medical imaging and image analysis problems, then move over to the buzz surrounding this topic of deep learning and how firm does the buzz hold on to the claims it boasts of? Also we would host a hands-on tutorial with implementing a deep network using auto encoders for retinal vessel detection problems on the DRIVE and STARE datasets, and subsequently move on to using deep networks for feature discovery, using deep-hybrid architectures by coupling them with some ensemble learners like random forests, and finally end with how to address the dilemma of interpreting what the deep network has learnt and how it has learnt to do so? Following this we would end on a note to the major challenges in the field of deep learning existing today and some thought provoking research problems.


  1. Introduction to the concept of hierarchical embedding in medical images. This part of the talk would include summary of works which describe classical approaches of using classical machine learning techniques employing shallow learners with knowledge transfer in hierarchy for unfurling hierarchical embeddings in medical images.
  2. Deep learning – What’s the buzz all about. This part of the talk would tell the story of how the deep learning theory existing since 1980 rose to its current crazy state since 2010, from traditional shallow reasoning to current complex deep reasoning and how they have addressed great challenges in machine vision, with clear differences between common terms, viz. convolution neural networks (CNN), stacked auto-encoders (SAE), deep belief networks (DBN), the great scope for soft IP reuse this technology offers, birth of new products and on what industries is it creating demand on to invent for the upcoming marketspace; where do the people in this community socialize.
  3. Unsupervised Representation Learning – Auto Encoders (AE). The architecture, learning mechanism, tunability of architecture, denoising mechanisms, interpretation of learnt feature representations, translation of learnt feature representations to traditional architectures, application┬áscenarios in medical imaging viz. denoising, restoration, segmentation, feature learning from data, and some hand on coding tutorials using the DRIVE and STARE retinal imaging dataset.
  4. Hybrid Pre-Trained Deep Networks – Stacked AEs (SAE). The resemblance of multi-layer perceptron with a stacked AE and the learning rule induced differences, learning stability of SAE, architecture, learning rules, moving over to basics of knowledge propagation in solving hierarchical pattern recognition problems, some tutorials with DRIVE and STARE using SAEs and random forest hybrid architecture.
  5. Practical Challenges with Deep Learning and how to overcome them. This part of the talk would focus on how to select number of layers, number of hidden nodes, learning rates, other free parameters, how to monitor performance, offline vs. online vs. incremental learning, catastrophic memory loss (if any) and its impact, and the grand challenge of handling medical data, some example case scenarios viz. organ localization, super-resolution imaging, etc.


The tutorial is particularly aimed at enthusiastic early adopters of deep learning for biomedical and biological imaging and image analysis, who are not trained in basics of machine learning. The target audience is expected to be from students in research training (master’s, PhDs) and Post-Docs who are not trained with a basic curriculum in machine learning (e.g. undergraduate training in electrical or biomedical sciences, physics, biological or medical sciences), but are interested to explore deep learning in their work. Audience is expected to have a basic experience of computer programming with Matlab and are expected to be conversant with commonly used terminologies in the biomedical imaging community. This tutorial would address the getting-started demand of using deep learning in their medical imaging and image analysis perspectives, where despite have done MOOCs aimed at machine vision, they face the big-biomedical data dilemma, and do not have access to massive computational facility, thus refrain from trying deep learning. They would be the ones who would benefit the most of this tutorial which would have hands-on demonstration of our recent papers on deep learning in medical image analysis published at leading conferences including the International Symposium on Biomedical Imaging.

Other tutorials

T1: Designing Benchmarks and Challenges for Measuring Algorithm Performance in Biomedical Image Analysis
T2: SimpleITK: An Interactive, Python-Based Introduction to SimpleITK with the Insight Segmentation and Registration Toolkit (ITK)
T3: Heart Mechanics by Magnetic Resonance Imaging: Techniques and Applications