AI for Everything

This Year’s Ai4 Conference — Day 1

Back in November, I had the opportunity to attend and write about the Ai4 Healthcare Conference. This year’s conference is a bit different. It was going to be September 1, 2 in Las Vegas and it was going to combine the previously separate conferences — Ai4 Healthcare, Ai4 Finance, Ai4 Retail, etc.

As with all things 2020, it’s been moved to a virtual format this year via the platform Dealroom Events which is something new to me and it’s pretty neat. A key idea of Ai4 is to be able to meet, network and learn about the goings on in the AI industry. Dealroom seems to be built on allowing networking and streaming talks to large groups.

Unfortunately, the conference started off with a glitch in the system leaving a number of confused attendees (myself included) wondering on whose side the issue lies. Personally, I couldn’t help but laugh imagining our future selves using a phrase like “classic 2020” to describe these moments. A number of the comments in the chat did poke fun at the technology issues at a technology conference — my favorites suggested auto generated subtitles, a few more epochs to get the model ready, and a lip reading model. On the upside, it did spur a really great side conversation in that about different use cases of AI.

Keynotes

For the keynotes, I’m presenting the abstracts below.

Algorithms of Oppression: How Search Engines Reinforce Racism — Safiya Noble

Safiya Noble discusses bias in algorithms. She is the author of Algorithms of Oppression. The landscape of information is rapidly shifting as new demands are increasing investment in digital technologies. Yet, critical scholars continue to demonstrate how many technologies are shaped by and infused with values that are not impartial, disembodied, or lacking positionality. Technologies hold racial, gender, and class politics.

Operationalize AI Throughout your Business with Intelligent Workflows — Madhu Kochar

AI helps you unlock the value of your data and gives you the ability to design intelligent workflows that are predictive in nature, allowing you to make better decisions. With prebuilt Watson applications, that speed time to value across key domains, such as customer service, financial planning, risk management and IT operations. Join us for this keynote to learn how to infuse Watson into your workflows and hear about clients who are already putting AI to work.

A New Era of Experiential Medicine: The Future of Technology as Cognitive Treatments — Adam Gazzaley

A fundamental challenge of our global healthcare system is the development and distribution of effective treatments to enhance cognition in those suffering from diverse psychiatric and neurological conditions. Dr. Gazzaley will describe the use of custom-designed, closed-loop video games to achieve cognitive benefits in both healthy individuals (Nature 2013; Nature Human Behavior 2019) and patients (Lancet 2020). This approach has now advanced to yield the first FDA-cleared digital treatment for ADHD, and the first video game cleared by the FDA as a medical device for any clinical condition. He will share with you the next stage of his research program, which integrates digitally-delivered interactive experiences with the innovations in machine learning, virtual reality, physiological recordings and non-invasive electrical brain stimulation to enhance cognition.

Track Talks

The conference features multiple tracks, each focusing on a particular subarea of AI. I spent my time on Healthcare and NLP since that’s where my interests lie. The talks run the gamut of very product oriented to very technical. They’re presented chronologically here.

How Ambient Clinical Intelligence is Transforming the Provider-Patient Experience — Robert Budman

Source: Budman, 2020 Ai4 Conference

Using voice to document patient-doctor interactions can help reduce the burden of paperwork and time spent entering notes for physicians. Ambient natural language understanding would be able to increase the accuracy, operating efficiency and throughput of a medical office. Nuance’s DAX aims to provide such an experience via a secure mobile app and can be used with a number of video chat and telehealth services.

The system also allows the physician to send orders in real time. An example in a video presented a physician was able to order a cast and medication for a mock patient in real time and in a manner that wasn’t disruptive to the patient-physician interaction. This helps provide a better experience for both provider and patient.

Training Conversational Agents on Noisy Data — Phoebe Liu

Typical process of building a conversational agent. (Source: Liu, 2020 Ai4 Conference)

This talk was focused on data collection and annotation for conversational agents, designing dialogues, and learning by imitation for social robots. Social robots can be thought of a hybrid between voice assistants and autonomous devices. They need to approach and engage people, and follow fuzzy social rules — while maintaining a personality that a human would want to interact with.

Challenges 1 and 2 for designing dialogue. Liu, 2020 Ai4 Conference

Naturally this poses some challenges in creating the bots including no way to collect human intents, hard to model the flow of real-world conversation, and data collection is noisy and costly.

Challenge 3 in designing dialogue. (Source: Liu, 2020 Ai4 Conference)

Noisy data can be a large part of a data set. The example below in Slide 3 contains 47% noisy data. Often the data contains filler words such as, like, um, uh/ah, etc. Appen was able to use an AI supported system and two tiered payment system to improve the quality of labeled data.

It’s possible to circumvent some of these challenges by continuously creating offline content by teleoperation, however this is expensive and creates a poor user experience.

Can improvements be made by in-situ human-human interaction and train the robot to learn off of this genuine interactions without annotation? Yes, it can be done and it does a good job despite some challenges in the training data such as, only 53% of the case study was noise-free or very minor errors, and questions asked by humans have natural variation, e.g., “what is the resolution?”, “how many megapixels?”.

Comparison of performance for social bot using ASR vs Behavior Correctness. (Source: Liu, 2020 Ai4 Conference)

Natural Language Processing — Challenges and Opportunities in Healthcare — Enrico Santus

Overview of how data can be transformed from unstructured data to actionable knowledge. (Source: Enrico Santus, 2020 Ai4 Conference)

How can one make data-driven decisions out of a mountain of unstructured and disorganized data? A hot area of machine learning focuses on natural language processing — taking spoken or written language and producing some useful information.

Santus discussed how his team was able to extract information from studies and make the data more accessible for analysis in retrospective studies. He also described work on improving patient selection for CRT (cardiac resynchronization therapy). The CRT study was able to identify 26% of patients that had no additional benefit from the CRT procedure.

Methodology for CRT Study (Source: Enrico Santus, 2020 Ai4 Conference)

Deep Learning for Automatic Extraction of Cancer Data from Unstructured Pathology Reports — Ross Mitchell

Much of the information used in the day to day in cancer treatment is stored in unstructured pathology reports:

Example of annotated clinical data. (Source: Mitchell, 2020 Ai4 Conference)

BERT (Bidirectional Encoder Representations from Transformers)has been able to help improve the ability to parse this type of data. Mitchell’s goal was to use BERT to extract tumor site and histology as it is critical information for many downstream tasks. This required overcoming three major issues: adapting BERT to extract accurate site and histology descriptions, handle diverse terminology, and ICD-O3 Codes. [Note: I had some intermittent technical glitches during the remainder of the presentation, the details might be slightly off but the big picture is accurate.]

To overcome the first hurdle and teach BERT the language, Google’s BERT was trained on PubMed Abstracts to produce BioBERT. Additional tuning was done with Stanford’s Q&A Dataset, BioASQ Q&A, and Moffett CTR Data.

Dealing with the diverse terminology is challenging. For example, code 8070/3 prefers “squamous cell carcinoma, NOS” but there are 161 phrases that are used in the real world. A library of SEER Coding Materials and Moffett CTR Phrases were utilized to handle tautologies. ICD-O3 Codes + Preferred and Alternate Terms were used to train DistilBERT which can then be translated to an ICD-O3 Code in a fault tolerant manner.

Results on the test set of pathology reports:

Mitchell, 2020 Ai4 Conference
Mitchell, 2020 Ai4 Conference

This particular talk has a manuscript in prep.

Meetings

One big difference from the in person conference is how networking is done. This time it’s completed using DealRoom’s platform (see below) to schedule meetings.

In person, I was able to stop and ask “What brings you to the conference?” to pretty much anyone I walked by. Naturally, it’s a lot harder to do virtually and there’s incentive to spend that time doing something else. However, the DealRoom platform provides a private chat room for the networking sessions which make it easier to have a deeper conversation in 20 minutes since all of the distractions are gone.

Conclusion

All told, this year’s conference is a bit different than previous ones because of COVID-19’s impact. The networking is still there and is easy to take advantage of. The keynote talks were glitchy on Day 1 but the track talks were very high quality and were just the right length at 20 minutes.