Deep Learning: Modular in Theory, Inflexible in Practice

Today we are holding the Deep Learning in Healthcare Summit in London and celebrating World Health Day 2016, an annual global event created to draw attention to topics of major importance for world health.

The high-level view of deep learning is elegant: composing differentiable components together trained in an end-to-end fashion - it seems like a perfect tool for enabling novel medical imaging applications by tackling some of its unique challenges such as large and high-dimensional datasets with unorthodox structure, extremely fine signals and massive diversity with limited data.The reality isn't that simple, and the commonly used tools actually greatly limit what scientists are capable of doing. 

At the Summit today, Diogo Moitinho de Almeida, Senior Data Scientist at Enlitic, will discuss some of the unique challenges to medical deep learning, what we can do about it, and how those things can result in much better models in practice. Enlitic, named one of MIT Technology Review's 50 Smartest Companies of 2015, applies state-of-the-art deep learning technology to medicine with the goal of radically improving diagnostic outcomes for patients. I caught up with Diogo ahead of his presentation 'Deep Learning: Modular in Theory, Inflexible in Practice'.

What areas of healthcare have the biggest potential for disruption by AI?
Medicine is inherently a data problem, so all of it is up for disruption. A lot of experts have commented that deep learning is missing a killer app, where all we've gotten so far is a better image search, and we think medicine will be AI's killer app. We are particularly excited about diagnostics, treatment planning, and population health because this maximizes the potential worldwide impact.

What are the main risks associated with applying deep learning methods into healthcare and medicine?
A huge problem that medicine will face is people trying to cash in on the recent AI hype, building up unrealistic expectations and burning the trust of both decision makers in the healthcare industry as well as patients. We are on the brink of an AI revolution and we as an industry need to act responsibly, because the actions today could lead to an environment with regulations that prevent people from getting the best care possible.

What developments can we expect to see in deep learning in the next 5 years?
I believe we can expect to see spatial attention that works well enough that it will be considered a prerequisite to any state of the art visual model. This will greatly reduce computational requirements and enable analysis of higher dimensional problems that are fairly important in medicine. Today, deep learning models are used for only a subset of machine learning problems. I believe that we will see that subset grow significantly as we improve our abilities to optimize deep models and integrate the best parts of the "shallow" algorithms as differentiable components into deep ones. Finally, the data requirements for successful deep learning applications will be significantly lower: they will need less labeled data, they will effectively utilize noisy and incomplete data, and they will pull together data from multiple modalities to find patterns that we know should exist, but as of today have no ability to find.

How will hospital systems change with the application of AI?
The experience of the patient will be completely rethought in an optimized manner. Patients will no longer die waiting for care, because the ER won't have any lines. Assessment and triage will be done automatically and patients will be immediately directed to the most relevant expert. They will know the results of their MRI or CT scan before they even leave the machine. AI can filter through all the relevant information: diagnostic and family history, genetics, and test and imaging results to recommend a optimally personalized treatment plan.

What impact will AI have on diagnostics?
For one, we will see the end of subjective diagnoses and second opinions, and healthcare will regain the trust of patients. Diagnostics today is severely limited in its ability to determine the long term impact of choices, actions, and interventions. The tools that we're working today will allow us to better leverage the massive amount of data we have been generating, but not utilizing, to deduce these long term dependencies. This can allow us to make decisions not just to make one healthy a week, a month, or a year from now, but for one's entire lifetime. Developing nations currently greatly lack access to medical expertise, but thanks to its scalability, AI will democratize healthcare for everyone.

What area of deep learning advancements excites you most?
All of deep learning is really exciting to me, but the software side of things excites me the most because it just makes everything better. The tools that exist today are far from perfect, but even as is they've enabled the massive amount of results that has given the field its reputation. I believe this is just the beginning, and that we are, in the very near future, going to make software that will allow us to dwarf what has been done so far and make the tools we use today look like they're from the deep learning stone age.

Diogo Moitinho de Almeida is Senior Data Scientist at Enlitic, where he develops deep learning algorithms for medical diagnosis. Read more about Enlitic here.

To learn more about the applications of deep learning, see our upcoming 2016 events:
  • Deep Learning Summit, Boston, 12-13 May 2016
  • Machine Intelligence Summit, Berlin, 29-30 June 2016
  • Deep Learning Summit, London, 23-24 Sept 2016
  • Women in Machine Intelligence & Healthcare Dinner, London, 12 Oct 2016
  • Deep Learning Summit, Singapore, 20-21 Oct 2016
  • View all future RE•WORK events here.

    Deep Learning Algorithms Deep Learning Computer Vision Pattern Recognition AI Healthcare MedTech Deep Learning in Healthcare Summit World Health Day


      As Featured In


      Partners & Attendees

      Rbc research.001
      Mit tech review.001
      Kd nuggets.001
      Maluuba 2017.001
      This website uses cookies to ensure you get the best experience. Learn more