Impact AI: Recent Episodes
Heather D. Couture
Learn how to build a mission-driven machine learning company from the innovators and entrepreneurs who are leading the way. A weekly show about ML challenges like data annotation, generalizability, explainability, bias, and collaboration across disciplines – and best practices for tackling them in a startup environment.
The accelerated development of medical AI could be life-changing for patients. Unfortunately, accessing large amounts of diverse, standardized data has been a major stumbling block to progress. That’s where Segmed comes in, a platform that allows researchers to access diverse, high-quality, and de-identified medical imaging data. Crucially, Segmed’s platform also provides data for medical AI training and validation.
I am joined today, by Segmed’s co-founder, Jie Wu, to discuss how they are solving key data issues to rapidly accelerate medical AI development. You’ll hear Jie break down some of the biggest challenges in curating medical image datasets — including the extra computational power needed to handle high-res medical images, like CT scans — and how they are addressing these obstacles. Jie also takes the time to emphasize the need for diversity when curating medical image datasets and the importance of mitigating bias during the data curation phase. To learn more about Segmed and how they are contributing to the development of medical AI, be sure to tune in today!
Key Points:
- A warm welcome to Jie Wu, co-founder of Segmed.
- Insight into how Segmed is solving data issues to accelerate medical AI development.
- Why solutions to these data issues are crucial for medical research.
- Segmed’s focus on medical imaging data.
- Their approach to different imaging modalities.
- An overview of the key challenges in curating medical image datasets.
- How Segmed determines the amount of data they will need.
- Best practices for curating a training set of medical images.
- Why collecting a diverse range of images is essential.
- An overview of how the quality of labels is assessed by experts.
- How imaging modality influences Segmed’s approach to creating datasets.
- The variations in datasets across different imaging pathologies.
- Special considerations that inform the validation set versus the training set.
- How bias manifests in models trained on medical images.
- Steps that can be taken to mitigate bias during the data curation phase.
- How the need for diverse datasets has increased along with greater awareness of bias.
- Jie’s thoughts on the future of foundation models in the medical AI space.
Quotes:
“A high-resolution of CT can take up to several gigabytes of storage itself.” — Jie Wu
“I think the most important piece is actually to collect as diversely as possible. So I ask that given the budget limit or maybe time limit, the size of the data set will be limited but it should be at least representative of the target population and targeted practice.” — Jie Wu
“The best quality labels are curated by experts and it is curated by multiple experts.” — Jie Wu
“A 3D image stores much more information than the 2D images, so you need less data for that.” — Jie Wu
“The external validation datasets require much more carefully curated datasets and much higher quality labels, and also it needs to be representative of the population, of the institutions, and also geographical locations.” — Jie Wu
“We hope that we can enter into the development of AI and make these algorithms go to market faster and benefit more people.” — Jie Wu
Links:
Jie Wu on LinkedIn
Segmed
These days, it seems that there are a lot of big problems in the world, especially in healthcare. Our guest today believes that there is massive value in tackling smaller problems, and, sometimes, the smaller problems are the most important to solve.
I welcome to the show today Joe Brew, Co-Founder and CEO of Hyfe, and he is here to talk about detecting and tracking coughing. We hear about what led to the founding of the company Hyfe and why they’ve narrowed their respiratory health innovations down to focus on cough. Joe talks about the role of machine learning, the process of gathering cough examples, and how they train their models. He touches on challenges they’ve faced, navigating model performance in varying environments, and the benefits of publishing their work. To hear more about why Joe believes now is the time to build this type of technology don’t miss out on this episode.
Key Points:
- How the movement of pathogens through our bodies and communities eventually led to the founding of Hyfe.
- What Hyfe does with respiratory health and why it’s important in overall healthcare.
- Why they’ve narrowed their focus down to the cough.
- The role machine learning plays in their cough-count technology.
- He explains more about acoustic epidemiology.
- The process of gathering cough examples and annotating them to train their models.
- We explore the challenges faced working with and training models on audio data.
- Navigating model performance in varying environments: working well in the real world.
- Joe shares thoughts on the benefits of publishing their work.
- Why now was the time to build this type of technology.
- How Joe and his team are measuring the impact of their technology
- Joe offers advice to other leaders of AI startups.
- We talk about the potential impact of Hyfe in three to five years.
Quotes:
“I realized that there are so many global health problems that are addressable, at least partially by tech. I hesitate to say, solvable, but addressable.” — Joe Brew
“The really big problem that Hyfe is tackling is around respiratory health.” — Joe Brew
“It felt to us that cough is perhaps, the lowest-hanging fruit, the area where the additionality of tech is greatest, because it's so prevalent and because it's currently just the status quo is so poor.” — Joe Brew
“If you really want reliable medical grade annotations, you need reliable medical grade input. Garbage in, garbage out. That's why the only way to really do that is through partnerships with medical professionals.” — Joe Brew
“A method, that if I were to start another company or to do another project, I would absolutely repeat, is to go quickly to the market, start collecting data, real-world data really quickly, and build a feedback loop where you're constantly training, testing, validating on real-world data.” — Joe Brew
“Our aim is not just to get nice comments on the App Store or nice emails. It's to impact the lives of millions. Everybody who breathes has lungs and everybody with lungs coughs. We think cough tracking is for everybody.” — Joe Brew
“Don't be turned off by problems that appear simple. Sometimes the simplest problems are the ones that are the most important to solve.” — Joe Brew
Links:
Joe Brew on LinkedIn
Joe Brew on Twitter
Hyfe AI
The impact of AI knows no bounds. Today, I am joined by Subit Chakrabarti, Vice President of Technology at Floodbase, a mission-driven, machine-learning-powered company specializing in flood monitoring and insurance. Having grown up in Eastern India, he knows the importance of adapting to global flood risk first-hand.
In this episode, Subit shares insights on how Floodbase utilizes advanced AI and diverse satellite imagery to support the design of parametric flood insurance solutions. We discover how machine learning plays a crucial role in analyzing vast datasets and bridging the insurance gap for regions vulnerable to flooding. Join us as we explore the transformative potential of Floodbase's technology and its vision for a more secure and equitable future, in the context of global warming and the associated global flood risk.
Key Points:
- Introducing Subit Chakrabarti, Vice President of Technology at Floodbase.
- Subit's background: what led him to Floodbase.
- Insight into Floodbase and its focus on parametric flood insurance and disaster response.
- Subit explains parametric insurance.
- How Floodbase uses machine learning to design the index for parametric insurance.
- Their use of satellite imagery and other geospatial datasets in setting up their ML models.
- The challenges of working with such diverse data sets.
- What made it possible to build Floodbase's technology (spoiler alert: advanced AI).
- How Floodbase measures its impact.
- Subit’s advice for AI-powered startup leaders: address bias and build a skilled AI team.
- Floodbase’s three to five-year plan: make insurance more accessible.
Quotes:
“Adapting to global flood risk is something that is near and dear to my heart, having grown up in India and having seen a lot of damage from floods in Eastern India where I used to live.” — Subit Chakrabarti
“Parametric insurance pays out when a pre-agreed weather condition is made separate from the physical damage.” — Subit Chakrabarti
“What we use machine learning for is to design [the] index that the parametric insurance can be based on, and that is our proprietary AI technology.” — Subit Chakrabarti
“One of the most important challenges with satellite imagery is that satellite imagery represents the condition of a place at a certain point in time and it’s not the continuous movement of what that flood looks like at that place.” — Subit Chakrabarti
“Our policy at Floodbase is that we add more data to remove bias from the process.” — Subit Chakrabarti
“The biggest thing that we can measure is the flood protection gap. So like I said, 83% of losses are uninsured and we can measure that.” — Subit Chakrabarti
Links:
Subit Chakrabarti on LinkedIn
Subit Chakrabarti on Twitter
Floodbase
The role of AI in cancer detection grows more significant with each passing week. During this conversation, I welcome Marcel Gehrung, CEO and Co-Founder of Cyted, to discuss detecting gastrointestinal cancer. You’ll learn how Cyted leverages machine learning to diagnose Barrett’s Esophagus in upper GI samples. Marcel reveals some of the challenges he has faced at Cyted related to the limited autonomy an algorithm can realistically provide, and annotating data for training and validation. Hear how the company is responding to changes in AI, and why hiring for technical roles at Cyted has not been difficult, due to their location. You’ll hear Marcel’s perspective on hiring specialist generalists and some of his advice for leaders at AI-powered startups.
Key Points:
- Introducing founder and CEO at Cyted, Marcel Gehrung.
- His path to focusing on gastrointestinal cancer detection.
- How the technology at Cyted works to diagnose Barrett’s Esophagus.
- What Barrett’s is and who is most susceptible to it.
- The role of machine learning in detecting cancer in upper GI samples.
- Navigating the challenge of how much autonomy an algorithm can provide.
- Annotating data for training and validation.
- How Cyted is responding to changes in AI.
- Hiring for technical roles at Cyted.
- Onboarding challenges due to the verticality of technology Cyted works with.
- Why Marcel advocates for hiring specialized generalists.
- Marcel’s advice for leaders of AI-powered startups.
- Where he sees the impact of Cyted in three to five years.
Quotes:
“We’re essentially leveraging the best of both worlds. We’re working with cytoscreeners, which we also have on our staff to generate the initial annotations, and then we have someone who looks at it and then reclassifies if necessary.” — Marcel Gehrung
“The more ability the candidates have to horizontally integrate different types of knowledge from across the company or across the technology of the sector, the better.” — Marcel Gehrung
“Getting carried away just happens so easily, particularly when we follow the various news outlets in the world that overwhelm us with new exciting ideas and functions of that technology.” — Marcel Gehrung
Links:
Marcel Gehrung on LinkedIn
Marcel Gehrung on Twitter
Cyted
Climate change is one of the most pressing issues of our time, and today’s guest, Ankur Garg, and his team at BlocPower are using machine learning technology to mitigate it. BlocPower is a climate technology company that is focused on making buildings in low and middle-income areas more environmentally friendly. Their area of expertise lies in developing products and services to lower or eliminate the barriers that prevent access to energy efficiency and electrification retrofits. And this all starts with gathering, checking, annotating, and understanding enormous amounts of data (BlocPower currently has over 40 terabytes of data in its data lake!)
In this episode, Ankur talks about the innovative ways in which BlocPower deals with its data, the challenges that they face when it comes to the size and scope of its datasets, why machine learning technology is central to the work they do, and how they measure the impact of their technology.
Key Points:
- Ankur’s career journey prior to joining BlocPower.
- Why Ankur decided to join BlocPower.
- The inspiring work that BlocPower is doing to contribute to solving the problem of climate change.
- The central role that machine learning plays in BlocPower’s approach.
- Ankur gives examples of some of the different types of machine learning models that BlocPower uses.
- The size of BlocPower’s data lake and the types of data stored within it.
- BlocPower’s innovative approach to annotating data.
- The importance of high-quality training data sets in the machine learning space.
- Challenges that Ankur and his team face when training machine learning models on their core dataset.
- Technological advancements that have allowed BlocPower to achieve what it has.
- How BlocPower measures the impact of its technology.
- What Ankur believes the future holds for BlocPower
Quotes:
“Climate change is one of the primary problems of our generation, and BlocPower is making a huge dent in solving that.” — Ankur Garg
“Machine learning really excels at ingesting huge volumes of data and to be able to infer key relationships between these data points to come up with an optimal output or a solution.” — Ankur Garg
“Labeling the data and annotating is extremely critical. If your training data set is not of a good quality, no matter what algorithm you use, it won't really perform well.” — Ankur Garg
“You need a lot of high-quality data for machine learning and artificial intelligence to be productive.” — Ankur Garg
Links:
Ankur Garg on LinkedIn
BlocPower
BlocPower Email Address
If you are working in the life science research space and battling with image recognition issues, firstly, you are far from alone, and secondly, there is a solution! That solution comes in the form of KML Vision, an AI-powered start-up co-founded by today’s guest, Philipp Kainz. In this episode, Philipp explains how he became aware of the image analysis problem and the process that he and his team have gone through to develop machine learning models that provide a range of benefits to a diverse cohort of end users. There is still a large gap between what is technologically possible in a research or lab setting and what is actually out there and what people can use. Through their flagship product, IKOSA, Phillip is on a mission to change that. Listen to this episode to gain an understanding of how machine learning is being used to shape the future of life science research!
Key Points:
- The motivation behind the founding of KML Vision.
- The value that KML Vision’s cloud platform, IKOSA, brings to the life science research space.
- The diversity of end users of IKOSA.
- Benefits that IKOSA provides to its users.
- Examples of some of the most common use cases for IKOSA.
- The role that machine learning plays at KML Vision.
- How KML Vision trains their models.
- The challenges that the KML Vision team have run into when training their models.
- Philipp explains KML Vision’s approach to developing the machine learning aspects of a new product or feature.
- How KML Vision helps to solve the problem of reproducibility in the life science research space.
- Valuable advice for leaders of AI-powered start-ups.
Quotes:
“We basically set out to help people overcome this barrier of using new technologies for image analysis.” — Philipp Kainz
“There is still a big gap between what is technologically possible in a research or lab setting and what is actually out there and what people can use. So, we are actually focusing on bridging that gap.” — Philipp Kainz
“Nobody [really] has time to go into the inner workings of deep learning. They want to use it like we use this smartphone today. This is where we want to be in three to five years.” — Philipp Kainz
Links:
Philipp Kainz on LinkedIn
KML Vision
Sustainability is finally getting the attention it deserves as the global drive to reduce our carbon emissions gets more frantic each day. Thankfully, the progression of AI has accelerated the way materials and chemical manufacturers can go about their business in an environmentally friendly and sustainable manner.
Today I am joined by Greg Mulholland, the Co-Founder and CEO of Citrine Informatics, a technology company that is focused on accelerating the development of the next generation of materials and chemicals. We discuss the role of machine learning in Citrine’s technology, the challenges they are forced to overcome regarding their data sets, the model accuracy and explainability balance, and how Greg and his team validate their models. There is no doubt that Citrine’s work is vital for the global sustainability effort, and our guest explains his company’s collaborative programs, how publishing research articles has boosted Citrine’s profile, what this AI-powered business hopes to achieve in the next five years, and so much more!
Key Points:
- Introducing Greg Mulholland, his professional background, and how he ended up at Citrine.
- Greg explains what Citrine does and why this work is important for sustainability.
- The role of machine learning in Citrine’s technology.
- Taking a closer look at Citrine’s data sets and the data challenges that they encounter.
- The techniques that Greg and his team use to successfully handle small data sets.
- Examining the balance between model accuracy and explainability.
- How he validates his models.
- An explanation of Citrine’s collaborative program with external researchers.
- The benefits of publishing research articles.
- Greg’s advice to other leaders of AI-powered startups.
- His vision for Citrine’s impact and influence over the next five years.
Quotes:
“I trained as an electrical engineer and got into material science because I believed that material science was really an important technology set of disciplines that we needed, to solve the world's most pressing environmental challenges.” — Greg Mulholland
“We started the company 10 years ago now; we've been able to show that machine learning and artificial intelligence, among other things, can be used to really accelerate the future of the materials and chemicals industry. It was the vision all along, but it really required a lot of technology development and we're really proud of how far we've come.” — Greg Mulholland
“The scientists in our community are brilliant people.” — Greg Mulholland
“Explainability is important. Accuracy is also important. Neither is dominant over the other. It turns out, a less accurate model that is more explainable can often help unlock new thinking in a scientist's mind, that then unlocks the next-generation product.” — Greg Mulholland
“Publishing what we do as a starter for more conversations; I think it helps us attract good talent. It helps people understand that we're doing cutting-edge research and continue to invest in driving forward the field. I take it as a little bit of a feather in our cap and a source of pride that we get to help the world move along into this new era of AI.” — Greg Mulholland
“We've seen companies remove toxic chemicals from important products much more quickly than they could have otherwise. We've seen companies reduce their energy consumption. We've seen companies reduce costs and reduce carbon input. Those are all really exciting to me.” — Greg Mulholland
Links:
Greg Mulholland on LinkedIn
Greg Mulholland on Twitter
Citrine Informatics
Spatial biology is an important part of the research being done to gain biological insights and joining me today on Impact AI to discuss how his company, Enable Medicine, uses AI to decode biology is Aaron Mayer. You’ll hear about Aaron's background, what led him to create his company, what Enable Medicine does and why, and how they use machine learning in their endeavors. Aaron shares the struggles they face, why they publish their research, the timing their company has nailed, and so much more! Finally, he shares some words of wisdom for other leaders of AI-powered startups.
Key Points:
- Aaron’s background and what led him to create Enable Medicine.
- What Enable Medicine does and why it’s important for healthcare.
- The role machine learning plays and how it’s used with spatial biology data.
- The challenges Aaron has faced working with this data.
- How they prepared for the use of AI by creating a data infrastructure from scratch.
- The importance of trust and transparency and the benefits of publishing articles.
- Why this is the perfect time to build this kind of company.
- Aaron shares some advice for other leaders of AI-powered startups.
- Where Aaron sees the impact of Enable Medicine in the near future.
Quotes:
“The goal of Enable Medicine is really to organize biological data and make it searchable to deliver insights to the questions that we really care about.” — Aaron Mayer
“Machine learning and AI is deeply integrated into the platform and technology stack that we've been building [at Enable Medicine].” — Aaron Mayer
“We want to take these various AI models and put them into an environment where they can operate with an expert in a loop.” — Aaron Mayer
Links:
Aaron Mayer on LinkedIn
Enable Medicine
Today’s AI-powered company of focus, Envision, has made it its mission to improve the lives of the visually impaired so that they can live more independently. I am joined by Envision’s Co-Founder and CTO, Karthik Kannan, to discuss how he and his team gather data for their unique models, how they develop new products and features, and how they are able to ensure that their technology performs well with multiple users and across various environments. Technological advances and the recent boom in AI mean that now is the perfect time for Envision to thrive, and Karthik explains exactly how he and his team are taking advantage of this unique moment. We end this informative discussion with Karthik’s advice for other leaders of AI-powered startups, and what he hopes Envision will achieve in the next five years.
Key Points:
- Introducing Karthik Kannan as he explains what led him to found Envision.
- What Envision does and how its technology improves the lives of visually impaired people.
- How Karthik and his team gather data for training their models.
- His process of planning and developing a new machine learning feature.
- Diving deeper into the development of a new product or feature.
- How Karthik ensures that his technology performs across multiple users and environments.
- Why now is the perfect time for Envision’s technology to thrive.
- How Karthik measures the impact of his company’s technology.
- His advice to other AI-powered startup leaders, and his hopes for the future of Envision.
Quotes:
“I was fascinated with games. I had asked my dad how people make games and he said, ‘They write code.’ Then he put me onto a programming class and that's how I got started with writing code.” — Karthik Kannan
“I didn't go on to my master's or anything. I just did my bachelor's, just started working directly, because I was more eager to get my hands dirty into making software and stuff.” — Karthik Kannan
“[Envision’s] overarching theme is to constantly look at how we can translate the advances in computer vision, and broadly AI, into tools that can help a visually impaired person live a more independent life.” — Karthik Kannan
“We mix both data from open datasets plus we throw in a healthy mix of data that's captured from a visually impaired person's perspective — that's what makes the whole data collection and cleaning process quite unique at Envision.” — Karthik Kannan
“That whole process of [user] validation is extremely, extremely important because we're not the direct users of the product ourselves.” — Karthik Kannan
“In the AI space right now, the most important thing is to try and understand where, or have a very clear idea as to what kind of impact AI is making on your customers, and to double down on that.” — Karthik Kannan
Links:
Karthik Kannan on LinkedIn
Karthik Kannan on Twitter
Envision
Envision on Twitter
Envision on YouTube
‘Envisioners Day- Hear from our users!’
AI seems to be taking the world by storm, and it is easy to use this new technology for either good or bad. Today I am joined by Sérgio Pereira, the VP of research at Lunit, a company using AI for good by conquering cancer with machine learning.
You’ll hear about Sérgio’s professional background, Lunit’s missions, how they use AI for cancer screening and treatment planning, and so much more. Sérgio delves into how they read imaging before discussing the differences between supervised, self-supervised, and contrastive learning. Lunit has created an incredible dataset called Ocelot, and he tells us all about its benefits, how they published it, and why publishing a paper while ensuring that quality products are being produced is a challenge. Finally, Sérgio tells us his hopes for the future of Lunit.
Key Points:
- A brief overview of Sérgio Pereira's background, and what led him to Lunit.
- Lunit’s mission and why it’s important in fighting cancer.
- How Lunit applies machine learning for cancer screening and treatment planning.
- What most of the Lunit products are based on and how genetics come into play.
- Why subjectivity becomes an issue when it comes to reading images in machine learning.
- Sérgio explains what supervised, self-supervised, and contrastive learning is.
- The lessons from machine learning work that can be applied to other types of imaging.
- He tells us about Lunit’s dataset, Ocelot, and the benefits of it.
- How Lunit publishes their datasets.
- The challenges of getting a paper out while getting a product on the market.
- Sérgio shares some important things AI-powered startups need to consider.
- Where Sérgio sees the impact of Lunit in the near future.
Quotes:
“Our mission at Lunit is to conquer cancer through AI.” — Sérgio Pereira
“We don’t have many products at Lunit, that’s a fact, but the ones we have, we believe they are [the] best-in-class.” — Sérgio Pereira
“Mistakes in healthcare can have a very big impact, so we need to be able to show and demonstrate that our products work as we promised.” — Sérgio Pereira
“AI can be used for good and for bad. Let’s make sure we work on the good part.” — Sérgio Pereira
Links:
Sérgio Pereira on LinkedIn
Sérgio Pereira on Google Scholar
Sérgio Pereira on Twitter
Lunit Inc.
Paper: OCELOT: Overlapped Cell on Tissue Dataset for Histopathology
Dataset: OCELOT
Paper: Benchmarking Self-Supervised Learning on Diverse Pathology Datasets
Sustainable waste disposal has been a global pain point for many decades. While the recent push toward comprehensive recycling has eased the pressure a little, there’s still much more to be done if we are to build a sustainable society. Luckily for us, the progression of AI brings new hope for feasible waste disposal, and today’s guest, the CTO and Co-Founder of CleanRobotics, Tanner Cook, is here to tell us how his company is playing its part in improving the disposal of waste, recycling, and compost.
In our conversation, we learn about CleanRobotics and why the company’s work is vital for sustainability, the ins and outs of their smart recycling product, TrashBot, and how it uses machine learning, how CleanRobotics ensures that its technology is always improving and up to date, and the impact of their AI-powered systems on sustainable waste management. Plus, Tanner offers up some noteworthy advice for other leaders of AI-powered start-ups before sharing his vision of the future of CleanRobotics.
Key Points:
- Introducing the CTO and Co-Founder of CleanRobotics, Tanner Cook.
- Tanner’s background and how he ended up as a co-founder.
- What CleanRobotics does and why this work is important for sustainability.
- Assessing the information that CleanRobotics is able to extract from its TrashBot product.
- The role of machine learning in TrashBot technology.
- How they gather and annotate data with TrashBot.
- The challenges of training machine learning models on imagery.
- How Tanner and his team improve their technology and ensure that it’s always up to date.
- The way CleanRobotics measures the impact of its technology.
- Tanner’s advice to other leaders of AI-powered startups.
- What he’d like CleanRobotics to achieve over the next five years.
Quotes:
“[I] found myself looking at trash cans very closely with my co-founder, Charles Yhap, and realizing, at the bin level and where people dispose of things, there were a lot of problems going on, and a lot of problems that artificial intelligence and robotics could solve.” — Tanner Cook
“The number of rule sets are very diverse throughout the United States and throughout the world. The rules can easily change for what is and isn't recyclable when you drive 20 minutes outside of your city.” — Tanner Cook
“One of our personal tellers internally for CleanRobotics is sustainability. Putting in those checks and balances to make sure that we're actually doing something good - instead of just greenwashing - is very important to us.” — Tanner Cook
Links:
Tanner Cook on LinkedIn
CleanRobotics
TrashBot
In the traditional paradigm, it can take up to ten years for a drug to come to market. For this episode, I am joined by guest Aaron Morris, Co-founder and CEO of PostEra, to talk about using AI to accelerate medicinal chemistry and bring cures to patients faster than ever before.
Aaron breaks down the medicinal chemistry process and explains how PostEra applies machine learning to drug discovery. The data landscape within drug discovery is particularly challenging and today, we learn about PostEra’s approach to gathering data, the data sets they build from, and how they find new uses for project-specific data. Hear about the importance of model interpretability and how to get a competitive advantage as an AI-powered startup.
Key Points:
- Aaron Morris’ background in mathematics and how it led to the creation of PostEra.
- The scientific disciplines involved in developing a drug.
- PostEra’s focus: building the world’s most advanced ML platform for medicinal chemistry.
- Aaron explains the process of medicinal chemistry.
- How PostEra applies machine learning to the drug discovery process.
- The challenging data landscape within drug discovery and the data sets PostEra builds from.
- PostEra’s approach to gathering data, and how they use it.
- The challenge of finding new uses for project-specific data.
- How PostEra validates its models.
- Why PostEra makes its models less black box and how they go about it.
- The importance of model interpretability and how PostEra develops interpretable ML.
- Aaron’s advice for other leaders of AI-powered startups.
- His vision for PostEra’s impact in the next three to five years.
Quotes:
“Though being reasonably competent on the machine learning side, I had a very, very steep learning curve when it came to getting up to speed with drug discovery chemistry and the applications of AI in that domain.” — Aaron Morris
“Drug discovery is going from biology to chemistry to medicine and PostEra squarely focuses, at least for now, on the chemistry angle. Our main focus is to build the world’s most advanced machine learning platform for what is referred to as medicinal chemistry.” — Aaron Morris
“PostEra is really the first AI company to pioneer machine learning across all three stages of how to design a molecule, how to make the molecule, and how to select the optimal set of molecules to test.” — Aaron Morris
“There is a lot of project-specific data that gets generated, and often what that means for PostEra is we’re having to be very inventive about how we try to get the most out of data even if it is not relevant.” — Aaron Morris
“If you want to build defensibility as a company, you have to have more than just innovations on model architecture.” — Aaron Morris
“Your typical drug today is taking anywhere between eight to ten years to come to market and obviously, we want to really accelerate that.” — Aaron Morris
Links:
Aaron Morris on LinkedIn
Aaron Morris on Twitter
PostEra
PostEra on Twitter
In this episode, I talk with Coleman Stavish and Julianna Ianni from Proscia about data-driven pathology. Coleman is the co-founder and CTO of Proscia and Julianna is the VP of AI Research & Development. We discussed the importance of quality control systems in an ML pipeline, model generalizability, and how the regulatory process affects ML development.
Quotes:
“Better accuracy in diagnosis means less overdiagnosis and less under diagnosis, which typically leads to better patient outcomes and quality of life.”
“Pathology is crucial in the drug development pipeline. It's helping pharmaceutical companies develop new treatments while assessing their safety and efficacy.”
“You'll often find slides that have been annotated with pen ink. That's something that can be quite common to do in some settings and that, if you're trying to train a diagnostic model, can really bias the model.”
“One of the heaviest impacts to development for us, just to give you an example, has been areas where we find a great level of disagreement in the ground truth data. So that will come out when you test, and we have to account for that disagreement during development.”
“It also requires thinking through, not just how are we going to validate, but then how are we going to keep tabs on the different deployments and ensure that we're not seeing performance degrade as maybe the data or the conditions within the laboratory change.”
“No matter how accurate or how valuable that information is that's produced by the model, if it's not actually introduced in the right way into the overall workflow, it's not going to be put into routine use.”
“Prepare to iterate. A solution that you build is probably not going to be the final destination, the final solution. And I think the fast pace of this field kind of demands some constant innovation.”
“I'd also say to heavily invest in your team. There's really nothing that replaces having good people and very skilled people working for you and building these AI products.”
“Something that we've learned ourselves is how to balance the investor pitch about AI and its potential with the near and immediate term. Smaller successes that build you a road to that more ambitious future.”
“They could have the ability to diagnose cases remotely without having and maybe assisting patients who are in far flung areas of the world that may not have access to subspecialty pathologist expertise.”
“Maybe it means someone gets the right diagnosis a little bit faster in aggregate. I think that could have a really big impact.”
Links:
Proscia
In this episode, I talk with Joe Peterson, co-founder and CTO of SimBioSys, about biophysical modeling of cancer. SimBioSys is trying to revolutionize precision cancer care through individualized treatment planning, accelerated drug development, clinical trial optimization, and comprehensive biomarker development. Joe and I talked about the challenges of working with heterogeneous forms of data and the ways bias can manifest when training models on medical data.
Quotes:
“We use AI or ML at effectively every point in the process, both in our clinical medical devices, but also for our internal R&D.”
“Have you ever seen the way weather scientists simulate a hurricane? We do a very similar thing within the body, or if you've ever seen mechanical engineers simulate the combustion of a gas and a gas turbine, we do a similar type of thing within these patient models.”
“If you're able to distill the processes that go on biologically, chemically and physically to their essence, you can create building blocks that can be mixed and matched.”
“Our thought was, let's not ask the models to do too much. Let's ask them to do one thing that we need them to do very, very well. This allows us to have more collected data or more directed data collection, as well as more clearly defined goals in terms of business value and delivering business value to each of the models.”
“All these different types of data are much more heterogeneous. They come from many different scales. They come from many different sources. They're encoded in many different ways, and so there's a huge effort, on the research and development side, just to extract what's meaningful in those different types of data sets so that we can begin to define those biophysical building blocks that ultimately make it into the clinical application.”
“It's just really about capturing the variability and trying to drive out as much variability up front as you possibly can.”
“We also develop models that are generally capturing any sort of drift in the data over time.”
“You wanna understand outside of just a research setting, but out there in the wild how well your models are going to work, how often you're going to return a null result or an inconclusive result to a physician and being able to track that over time is really important from a quality control standpoint.”
“It's all the quality control machine learning models and deep learning models that make up the bulk of those internally.”
“Our responsibility as practitioners of AI is to not only identify and understand that bias, that historical bias, but also try to account for it as best we can.”
“What we need to assess when developing drugs or algorithms or devices is how they were trained, how they were tested, and really stratify those patient populations as best we can to sort of understand, at the very least, how they're behaving.”
“We've spent a lot of time trying to account for that variability as best we can. That said, we don't have a perfect data set and we're constantly thinking about ways to improve it.”
“I think what it comes down to is being open and transparent and really looking at the data that you have at the end of the day, If doctors are going to trust medical devices and if they're going to trust AI, they need to have information about.”
“By looking into and stratifying the patient populations in that way we can better understand where we need to targetedly spend resources to collect potentially more data to better understand the performance in those places or to improve our algorithms.”
“Adopt good machine learning practices early, just like good clinical practice or good manufacturing practices that are standards that are now being drafted and adopted.”
“Find the right partners to sort of drive the questions that you're addressing and ultimately the clinical actions that you're trying to address.”
“Models that are built to do a single task excellently well is a better approach than trying to build a model that does four or five tasks really well.”
Links:
SimBioSys
In this episode, I talk with Eric Adamson, CEO of Tortuga AgTech, about smarter farming. Tortuga AgTech builds robots for harvesting fruit and vegetables to help farms be more resilient, sustainable, and successful.
Quotes:
“Figuring out that pipeline from someone else's knowledge to the robot knows it is really critical.”
“If you build technology because the technology is cool or because you can, you are much more likely to fail than if you start with the customer problem and then figure out what kind of technology might help to solve that problem.”
“That learning happens with our machine learning engineers being in the field, being the ones who are actually taking data with handheld rigs.”
“Many of our team members’ first two weeks have been immediately flying to a farm and spending time on the farm with the robots, learning a problem in very, very deep detail. And I would encourage anybody building a technology based on machine learning or certainly robots to do the same.”
“We have a very efficient and effective pipeline that took us years to build. But it's exceptionally powerful for us to be able to, for example, go to a new site, run a couple robots or a small fleet of robots for a day, and then within a week have a brand new model that's been completely retrained on freshly labeled data from this new place.”
“That’s very critical for us because farm environments are changing so often. You really need to be able to be reactive and continue to improve your models as you develop.”
“We measure our scores based on golden data sets that we've sort of hand labeled ourselves. But we also have to make some judgment calls about what we really want in our performance versus what the conditions are in the field and what we're seeing on the farm.”
“We try to convert whatever model results are spit out into language that the customer intuitively understands.”
“It's really important to start with the customer problem and to start with the customer problem as an economic proposition.”
“There are already very large discussions happening in the farming community around what type of farming should be used in order to, for example, deal with climate change, to deal with drought, to deal with chemical regulations, to deal with a lowering of fruit quality and an increasing of fruit waste, the challenging labor environments.”
Links:
Tortuga AgTech
YouTube
In this episode, I talk with John Bertrand, CEO of Digital Diagnostics, about autonomous diagnostics. Digital Diagnostics transforms the quality, accessibility, equity, and affordability of healthcare with AI-powered diagnostics. They developed the first FDA-cleared autonomous AI system.
Quotes:
“So we look for diagnostics where there's an established understanding of what the disease is and there's a gold standard as to how to measure that.”
“We'll naturally start with an area where positive and negative is a very binary decision that is almost mathematically derived.”
“It goes back to picking the right types of disease states to make sure that the gold standard already exists.”
“How do you take images that have different coverage of the retina but make sure that you piece them together in a way that the processing part of the system is getting a consistent image that they're looking at every single time so that the algorithm remains consistent and we don't have to have different algorithms per vendor that we're interacting with.”
“We’re pretty proud of the fact we’ve been able to do that first kind of assistive feedback for the provider.”
“We want every single patient, regardless of their background, to receive consistent quality of diagnostic output. What that means is that we actually have to build our training data sets as well as our clinical validation studies and trials to take into account a diverse population set.”
“Continuous learning versus locked algorithms is another key factor. . . Would you really want that algorithm to adjust to the most recent data it's seeing, thinking it's attempting to become more accurate, when in fact it's really more optimizing for the ethnicity of the folks in that particular region, the sun rises on the east coast to the United States, everybody further east goes to bed. Now the algorithm’s been indexed towards another group from a ethnicity perspective, that’s no longer representative of where the testing’s being done as the sun rises in New York.”
“How do we ensure that we create confidence with regulators, with providers, and with patients that we've actually thought through this?”
“We can literally break down for you what the computer saw, why graded it out what it did, and why it gave you the results it did.”
“Your algorithm should be explainable so people trust the technology, understand how it works.”
“Also explainability helps you drive better accuracy and that you understand why you're getting the result that you're getting with the black box approach.”
“You really want to work within the healthcare system when you’re building these types of businesses.”
“If you're going to chart that course and really carry through to fruition, your vision of building an algorithm that impacts patient lives, I think you really need to center the culture of the business around a commonly shared vision for the mission of what you're trying to do.”
Links:
Digital Diagnostics
In this episode, I talk with David Schurman, co-founder and CTO of Perennial, about their verification platform for climate-smart agriculture. Perennial uses geospatial data and machine learning to unlock agricultural soils as the world’s largest carbon sink.
Highlights:
- How Perennial gathers and annotates training data from satellites and ground-based observations.
- Handling variations across satellites and geographic locations.
- Stratifying training data across the kinds of variables that matter.
- Collaboration between machine learning engineers, remote sensing scientists, and crop scientists.
- The importance of gathering more training data than you think you’ll need.
- Respecting the data.
- The nuance of communicating performance metrics.
Links:
Perennial’s website
Perennial on LinkedIn
David Schurman on LinkedIn
In this episode, I talk with Matt Alderdice, Head of Data Science at Sonrai Analytics, about precision medicine. Sonrai Analytics automates laborious data processes and speeds up new drug and healthcare developments.
Highlights:
- Machine learning for automating time-consuming and tedious analysis of microscopy images.
- Training for machine learning practitioners new to pathology by integrating domain experts with your team.
- Involving stakeholders throughout a project.
- Literature reviews to search for associated publications and potential solutions to avoid overly complicated solutions.
- Validating models with ethnically diverse datasets.
- Analytical validation for differing stains, scanners, and operators.
- Clinical validation on a held out dataset in the same environment as would be in the clinic.
- Identifying relevant metrics from conversations with pathologists, oncologists, nurses, and patients.
- Focus on the problem you’re trying to solve – AI is just a tool.
Links:
Sonrai Analytics’ website
Matt Alderdice on LinkedIn
In this episode, I talk with David Golan, co-founder and CTO of Viz.ai, about diagnosis of acute and emergent diseases. Viz.ai increases the speed of diagnosis and care for a variety of conditions to improve the lives of patients.
Highlights:
- Increasing access to lifesaving treatments.
- The importance of the full system, not just the machine learning component, in accelerating workflows.
- Their clinical AI team includes med students, MDs, biomedical engineers, and neuropsychologists.
- Bias can be created by a lower performance on a subset of the population in a way that is unknown to developers, users, and clinicians.
- Careful monitoring of algorithms to identify subsets of data with poor performance.
- Unbiased collection and stratification of data for FDA submission.
- The importance of good annotation and monitoring infrastructure.
- Relatively simple model architectures can take you a long way.
Links:
Viz.ai’s website
David Golan on LinkedIn
In this episode, I talk with Steve Brumby, co-founder, CEO and CTO of Impact Observatory, about sustainability and environmental risk analysis. Impact Observatory uses satellite imagery and machine learning to empower decision-makers with planetary insights.
Highlights:
- Using machine learning to generate thematic maps to represent land cover and land use.
- Geospatial data from the European Space Agency’s Copernicus program that is available on a variety of platforms.
- The importance of identifying the relevant output for end users and others in the value chain.
- How machine learning engineers sometimes discover things used by remote sensing scientists that are no longer necessary.
- Keeping models simple.
- Mitigating bias in models by using large and globally diverse datasets.
- Get to know your customer and their pain points, then craft a machine learning solution that works for them – if you’re lucky, it’ll also work for others.
- Finding the things you’re passionate about – both the technology and helping the customers in that space.
Links:
- Impact Observatory's website
- Steve Brumby on LinkedIn
- K. Karra, C. Kontgis, Z. Statman-Weil, J. C. Mazzariello, M. Mathis and S. P. Brumby, "Global land use / land cover with Sentinel 2 and deep learning," 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, 2021, pp. 4704-4707, doi: 10.1109/IGARSS47720.2021.9553499
- Brown, C.F., Brumby, S.P., Guzder-Williams, B. et al. Dynamic World, Near real-time global 10 m land use land cover mapping. Scientific Data 9, 251 (2022). https://doi.org/10.1038/s41597-022-01307-4
- Impact Observatory’s data
- https://www.arcgis.com/home/item.html?id=d6642f8a4f6d4685a24ae2dc0c73d4ac
- https://planetarycomputer.microsoft.com/dataset/io-lulc-9-class
- UN Biodiversity Lab: country dashboard for Australia
In this episode, I talk with Dean Freestone, co-founder and CEO of Seer, about epilepsy. Seer uses home monitoring to diagnose and manage neurological conditions, relieving bottlenecks in the healthcare system.
Highlights:
- Using machine learning to summarize data to reduce the labor intensive search for episodic events.
- Handling imbalanced datasets.
- Controlling the workflows to enable gathering and annotating huge datasets.
- Working with technicians to speed up review of EEG data.
- Using machine learning to capture features that doctors can’t describe.
- Dealing with low inter-reviewer agreement from clinicians.
- How bias can manifest is neurological data.
- Do not underestimate the cost and amount of work to build a healthcare AI startup.
Links:
Seer’s website
Seer on LinkedIn
Dean Freestone on LinkedIn
Welcome to Impact AI, the podcast for startups who want to create a better future through the use of machine learning.
I'm your host, Heather Couture.
In this podcast, you’ll learn how to build a mission-driven machine learning company.
I’ll be interviewing innovators and entrepreneurs from a variety of industries: healthcare, drug development, environmental, agriculture, and many more.
Each is striving to solve a problem that they are passionate about. They will talk about the role machine learning plays in their technology and the impact of their product.
They will also help me uncover machine learning challenges like data annotation, generalizability, explainability, bias, and collaboration across disciplines – and best practices for tackling them in a startup environment.
Now, who am I?
I’m a consultant with almost 2 decades of experience in computer vision and machine learning for a variety of applications. From manufacturing to planetary science to commercial media to cancer research.
I completed a Masters at Carnegie Mellon University and a PhD in Computer Science at the University of North Carolina. As a researcher, I published in top-tier computer vision and medical imaging venues. Now I write regularly on LinkedIn, for my newsletter Pathology ML Insights, and for a variety of trade publications.
I offer consulting services through my company Pixel Scientia Labs to help startups get to market faster by building more generalizable computer vision models. I make use of the latest machine learning research to amplify their results and support their in-house team for the long term. My mission is to fight cancer and climate change with AI – and I do that by strengthening the machine learning component of my clients’ most impactful projects.
My hope for this podcast is to share machine learning best practices more widely so that many others can benefit as they work towards solving important problems.
Thanks for listening.
Please hit subscribe to be notified about new episodes.