Neural networks, or deep learning, is a capability that is changing the way people live and work. From language translations to medical diagnosis, to speech recognition and self-driving cars, deep learning is in the fabric of a technology revolution.

But what is deep learning, and how much knowledge does a nontechnical or computer science stakeholder need to have to contribute to or run projects, or to spot opportunities for applications? How do healthcare executives know the potential data objectives can be addressed with deep learning?

To add more complexity, the marketplace is filled with content and claims that will confuse even the most ardent expert. Commercial interests stretching claims, a lack of clarity of what is involved in a healthy project, and science-fiction expectations all have created a snowball effect of fuzziness on even the most basic understandings.

Next month in his HIMSS22 educational session entitled "Demystifying Deep Learning and Neural Networks," Michael Meighu, a director at CGI, an IT and business consulting firm, will simplify the subject matter so that all attendees can understand what is at stake.

Meighu is a life science specialist with more than 20 years of experience in the field, holding a doctorate from Alliance Manchester Business School in the U.K.

Healthcare IT News sat down with him to glean some of his expertise and get an advance look at his HIMSS22 Global Conference & Exhibition session.

Q. Please help a layperson better understand deep learning/neural networks by reframing the concept in everyday language.

A. Deep learning/neural networks is not magic or something mystical. In fact, I will propose to you that you are already an expert in it, even if you don't know how to use a computer.

You see, the human brain in its excellence, and perhaps the thing that separates us from other species, is our ability to learn, to learn fast and to act upon that learning. The brilliant part is that the human brain does not even need to understand the nitty gritty of what we are learning. We learn by pattern recognition, trial and error, and experience.

Essentially, phenomena can remain a black box, but our brains have found a winning formula to circumvent these hurdles through pattern recognition.

Consider for example our first language. Truth be told, I don't know the rules or logic behind English. But I have been told I am a good writer. I have learned to write based on the patterns of what I have read and heard before.

My brain isn't working out language logic step by step, or the rules of language as I write. However, I know when something isn't sitting well. I may not even know why. I just have an "Ah, that doesn't sound right" feeling.

Consider also, putting five kids in a room and giving them 100 labeled photos of cats and dogs to learn from. After that exercise, you can give them new unlabeled dogs and cat photographs, and by some learning human mechanism, they can label them on their own correctly. They don't need a specification to understand what constitutes a dog or a cat.

This is all possible because of our natural neural networks. The ability of our brains to fire neuron sequences to recognize patterns, according to inputs from all of our five senses. Computer neural networks or deep learning is a pattern-recognition algorithm inspired by the biological one.

In computer practice, the main building blocks are an appropriately configured neural network and training labeled data. The use cases are virtually endless: from OCR software, to cameras on smartphones recognizing foreign language texts on signs, to X-ray diagnosis to translations.

Computer neural networks already are here and all around us. And like human learning itself, there is no end or saturation to applications and creativity.

Q. What are the key components of a deep learning/neural networks project?

A. There are many key components in a deep learning project, and ironically the actual neural network is probably the simplest component. It is the simplest step, because most algorithms are already available off the shelf, or part of a commercial product, and the neural network exercise is about fine-tuning hyper-parameters.

But back to the example of having children in a room being trained with labeled pictures of dogs and cats with the expectation of identifying new picture categories. What if there were two pictures of a dog and 98 pictures of cats? Would we really be surprised if the kids struggle to correctly identify new pics of dogs?

Or what if mingled up in the dogs picture sample there were also pics of wolves, and in the cat sample there were pictures of tigers? Would we be surprised if the kids identify wolves as dogs with new unlabeled pictures?

What if the homework assignment was actually to identify new pictures of cats, dogs, mice and squirrels, but somehow that message was lost in collecting the training photographs? This represents a metaphor regarding the key components in a deep learning project.

The key components of a deep learning project go way beyond the technology and must include actual alignment of the training data, the business question and the personnel performing the work.

The established key components from Geron's "Hands-On Machine Learning, with Scikit-Learn Keras & TensorFlow" suggest:

1. Look at the big picture.
2. Get the data.
3. Discover and visualize the data to gain insights,
4. Prepare the data for machine learning algorithms,
5. Select a model and train it.
6. Fine-tune the model.
7. Present the solution;
8. Launch, monitor and maintain the system.

This is consistent with my experiences.

Q. What are a couple of warning signs of what can go wrong in such projects?

A. The essential issue with deep learning projects and machine learning projects in general is the idea of the "No Free Lunch Theorem." There is no model a priori made to work for a data set. You have to do the legwork to ensure you are using the right model, configured in the correct way for the dataset in hand to solve a particular business issue/value to expectations.

Surely, there are accelerators in the market, models already predefined with assumed training data for a particular problem, but the process dynamics must be followed to assure the best ROI and meeting expectations.

Some of the key red flags (but there are more) to keep an eye out for include:

• Marketing victim/unrealistic expectations. Technology vendors tend to oversell capabilities and undersell project requirements in order to minimize the "no free lunch theorem." Look out for a poorly executed proof of concept with bias data, and one that did not map out scale-up issues and resolutions. If there was no POC to map business value to capabilities, then that is also a red flag.
• Data, data, data. Garbage in is greater than garbage out holds even more true in machine learning projects. The data analysis phase is probably the most key part of the project, and the most time-consuming. [An observation] that visualization on the data was not performed, or few people are intimate with the meaning of the datasets, is a red flag.
• "One-man show." In some cases, the AI project is a lone person championing an objective, even with the best of intentions. Successful AI projects comprise teams of individuals with a mixture of diverse skills sets with sponsorship at an appropriate level. If there is a sense that an AI project is a "one man show," then this is a red flag.
• Not aligned with the business, or objectives are not clear. AI is not magic. It is a collection of various capabilities aimed to answer a particular question or business need. It is not the inverse: Provide the business with capabilities and hope they find use cases for it. A red flag would be capabilities not tied to business objectives.

Michael Meighu's HIMSS22 session, "Demystifying Deep Learning and Neural Networks" is scheduled for Tuesday, March 15, from 1:30-2:30 p.m. in the Orange County Convention Center, room W311E.

Twitter: @SiwickiHealthIT
Email the writer: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.