Impact AI

In this episode, I talk with Gershom Kutliroff, CTO of Taranis, about precision agriculture. Taranis uses computer vision to monitor fields, providing critical insights to growers. Gershom and I talked about how they gather and annotate data, the challenges they encounter in working with aerial imagery, and how they validate their models and accommodate data drift with continuous learning.Quotes:“Taranis is using drone technology to capture imagery and then use AI to process that imagery to understand what's happening in grower's field.”“It becomes increasingly difficult to maintain consistent quality levels if you're working with tens or even hundreds of annotators. But when you have AI models, then you have the ability to control the quality of the insights that you're generating.”“There's a lot of discussion in the last few years in the AI space about data-centric versus model-centric. Model centric would be the case where in your development you invest a lot in  choosing the right architecture that optimizes your performance, gives you the best results for your models, or spending a lot of time with hyper parameters and that type of work. And data-centric is you spend a lot more time making sure that your data set is clean, that you've got that it's balanced, you've got the right amount of classes for the problem that you're trying to solve.”“We struggle with the problem of long tail distributions. If I take diseases as an example, there are some diseases that can cause a lot of damage to the crops. But they're very rare in terms of how often they actually occur in grower's fields.”“Because we're running our own operations and so we're flying our own drones, we've also  invested in the software that's running on the drones when we're flying. So the images the drone pilot captures in the field are validated in the field. We have algorithms running on the edge to be able to check the quality of those images. And then if the images are not the quality that we expect them to get, the pilot knows while he's still there at the field and he can fly again.”“For a lot of the models that we use you really need domain experts. You really need trained agronomists who can look at these images.”“A certain percentage of all of the missions that we've flown are sent for review by our in-house agronomists before we release them to customers. So that's a really critical piece of how we do validation, and that also gives us a high level of confidence internally that the product that we're releasing to our customers stands by the quality that we expect it to.”“We do suffer from this type of data drift where the data that we're seeing in production is not exactly in the same distribution as the data that we used to train. So the most effective technique that we've seen is to implement some kind of a continuous learning type of framework whereby we are able to take data that we're capturing in production, so when we're actually live   and servicing our customers' fields. And then the data that doesn't have a good correspondence with the distribution of the training data that was used for the models, we can then filter that data out. We can extract that data and use it to quickly retrain the models, to adapt the models, and then deploy those models back into production.”“The company started by offering a product based on manual tagging, which didn't have any AI technology at the beginning, which allows it to offer products and service customers and start building this very rich database that we leverage now.”Links:TaranisGerhsom KutliroffResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

In this episode, I talk with Vinod Subramanian, Chief Data and Product Development Officer at Syapse, about machine learning for healthcare and advancements in cancer treatment. Syapse is a real world evidence company dedicated to improving outcomes for cancer and other serious diseases. Vinod and I talked about the types of healthcare data they work with, the data challenges they encounter, how they validate their models, and how they mitigate bias.QUOTES:"Technology is not the answer exemplified of the intent. And the fundamental question, I think, that all of us are confronted by: what is the intent and what in the world that we want to try to help shape?""There are infinite possibilities in the terms of patient care with aggregated and harmonized data in healthcare. We all know about the point that data in general is fragmented and decentralized in the industry. Real world data comes from knowledge and knowledge comes from collecting information and of course, information stems from aggregating disparate data.""Machine learning today, especially in a life science setting, is leveraged as new ways right to garner new biological insights.""One of the things that we are also doing is not just about adopting and using (ML and NLP), we strongly believe that we want to share our work. And that would not only raise and mainstream the work of everybody doing it, but also it'll help us in adopting and applying in precision medicine through standards.""Now not all data is needed equal. When we can improve the way data is collected, connected, analyzed, and consumed, we can not only improve the lives of our community, but it also gives us a way to look at the care continuum very differently.""There's no guarantee when you get into an initiative which uses machine learning and AI, because it cannot be successful. It has to be a learning experience, but it, there's no guarantee that it will be  successful. And there needs to be willingness and appetite to experiment, learn, and iterate, and taking a Socratic approach, and accelerate the journey towards success, anchor down the culture."LINKS:SyapseVinod SubramanianResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

In this episode, I talk with Patrick Leung, Co-founder and CTO of Earthshot Labs, about using machine learning to help predict and restore forests and our ecosystem. Earthshot Labs is building the technology and expert guidance to develop and finance nature-based carbon projects globally. Patrick and I talked about how Earthshot Labs gathers and annotates data, the challenges in working with remote sensing and other forms of data, the importance of collaboration across disciplines, and how machine learning tools can help save our ecosystems.QUOTES:"We are able to actually bridge that financing gap and unlock a whole bunch of new projects that can then be in the carbon marketplace, and also bring a host of benefits to both the ecosystem, as well as, the communities that live around the ecosystem.""Machine learning is really essential because what we're trying to do here is predict the future. We're trying to predict the next 30 years of a forest regrowing in a tropical region.""We must look at the past. We must look at whatever data we can gather from the past state of the ecosystem and use various machine learning methods to predict the future in order to provide a view on what's gonna happen on this land in the future when we do this project.""These are actual mathematical simulations that take into account the current conditions of the ecosystem and actually forecast them by using a kind of simulation that incorporates photosynthesis and evapotranspiration and other forms of ecological processes.""They would look at historical flood maps and essentially combine them with flood forecasting models i order to generate what is a given area going to look like if it gets flooded in the future because of climate change or for other reasons. And I was very enamored with that. I thought that was a very, very clever use of a technology.""I think what we're doing definitely encompasses biodiverse native ecosystems and just restoring as many of them as we can throughout the most critical parts of the biosphere, that there are in this world. And also helping to switch our societal systems into more of a harmonious, and regenerative relationship with those ecosystems."LINKS:Earthshot LabsPatrick LeungResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

In this episode, I talk with Matt Pipke, Co-founder and Chief Digital Health Officer of physIQ, about personalized medical predictions from physiology data. Matt and I talked about the challenges in working with physiology data, how to validate models and minimize bias, and the importance of collaboration.Quotes:“What physIQ does is it harvests data from those continuous data streams from wearable sensors and produces analytical results that are useful for clinical care when taking care of patients who are outside the four walls of the hospital and in scientific endeavors such as clinical trials where it's interesting to know what the efficacy of the drug is on a target disease, whether the health of the patients who might take those drugs is being improved or at least is not degrading any further.”“What we have to do is build our algorithms and our analytics based on machine learning techniques and, of course, the more recent really successful subgroup of deep neural net algorithms that can sift through this data and can highlight accurately the vital signs of physiology we need to make the assessments available.”“So part of the issue there, is to figure out how to differentiate the background variation that's normal for people as they move around in their daily lives from the telltale signs that they may be suffering from a derangement of physiology.”“There's a lot of companies and offerings out there that are in the consumer fitness market. They might be appropriate for healthy populations that are looking to track their activity, the amount of sleep that a healthy person might get, but they're really not the right target populations of interest for the medical system or for clinical trials where you have a population that's suffering from a disease that a drug is targeting.”“Now I know that a lot of companies out there tend to avoid the regulatory pathway for medical or health or fitness applications, and I don't think that's a good move. . . The FDA experience for us has been at times frustrating of course, as it is for anybody who has to deal with regulations, but at the same time, there is a core of meaningful value add there. Regulatory agencies around the world, FDA included, they have a pretty thankless job. They never get credit for what they do. They only get complained about. But what they're doing is really, really critical to outputting valuable, usable product in the healthcare and medical space.”“So bias in models really comes back to the representativeness of your data, right? So if you've got data that's not representing the target users, the target populations that you're going to analyze, you can end up with bias. You can end up with bias in surprising ways.”“If you aren't aware of what might be lurking in your data, you could be overfitting the wrong thing and then find out that your algorithm does not generalize, does not work in other areas.”“My feeling about this is that it's all about the data. physIQ got started a lot earlier than we probably should have and we've benefited in a strange way in that we've been in the game a lot longer than other players in this space. So we've been collecting data for a long time and we built a robust platform to collect data.”“There's a lot of resistance to change and, in fact, the layperson might be horrified to learn how the healthcare system actually works. But, stepping back, something definitely has to change in healthcare. We all know that it's not sustainable the way things are now. But we don't have any illusions at physIQ about how a little company like ours can change things by ourselves. It's really about timing. Right. And sometimes you have to look for those windows of opportunity when in large industries with huge amounts of existing business relationships and the way that they work today are ready for change.”Links:physIQResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

In this episode, I talk with Emi Gal, co-founder and CEO of Ezra, about cancer screening with a full body MRI scan. Ezra is on a mission to detect cancer early for everyone by making the process more accurate, faster, and cheaper. Emi and I talked about the challenges in working with MR data, how regulatory processes affect model development, and the importance of validation.Quotes:“What we've been able to achieve is to essentially reduce the cost and the time in a scanner of an MRI from about two to three hours for full body to 60 minutes. And we're actually working on a new AI that will roll out next year that will reduce the scan time to 30 minutes.”“What we do is we acquire the scan fewer times, and then we've built machine learning models that recognize what noise looks like and then just remove that noise. And then we kind of expanded that from not just noise. If you acquire scans with lower resolution, the resulting images are a little bit blurry so we can sharpen them.”“Our focus on the scanning front is to reduce scan time, which yields these images with increased noise artifacts, and then use machine learning to enhance these images so that a radiologist can then use them for interpretation.”“I think what having to receive FDA clearance for AI does, is it really forces the company from day one to think about what are all of the things that might influence the performance of said AI, and what can we do to ensure that we maximize the chances of success?”“We have had an instance when we had to go back to the drawing board and build the model again because we failed internal validation prior to formal validation that we had to submit to the FDA.”“I think the way you ensure that the technology we develop fits the clinical workflow is actually not starting with the technology, but starting with the end goal in mind and then figuring out what you need to do in order to achieve that.”“To screen a hundred million people a year, we think, is a huge endeavor and probably going to take a decade or two to achieve. And I'm personally committed to Ezra for the rest of my career.  In the next three to five years, I would hope we are making good progress towards that mission, and maybe in five years we're screening at least a million people a year.”Links:EzraEmi GalResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

In this episode, I talk with Amanda Marrs, senior director of product at AMP Robotics about modernizing the world’s recycling infrastructure. Amanda and I talked about how they ensure their models work for a diverse set of objects, measuring the success of their technology, and some tips for building a successful ML team.Quotes:“At AMP we have a broad mission of enabling a world without waste.”“We work backwards on everything that ends up in a landfill to develop the technology we need to keep that from happening.”“We really have two main areas that we work in. One is technology that we will put in place at a material recovery facility. . . The other half of what we do at AMP is use our own technology for what's called a secondary sortation facility.”“All of this technology really has three main components. You have to be able to see the objects on the belt, and that's where the machine learning comes in. You have to be able to sort the objects effectively, and there's some ML behind that as well. And then you have to be able to report, see what's happening, and draw conclusions and make decisions and optimize further in the facilities.”“A majority of the data fits nicely within these primary categories. But, in AI, typically there's this natural long tail, and we have that as well.”“Diversity is the name of the game in this industry where you have to be able to recognize everything. And so a huge sample set of data really helps us overcome that.”“The wonderful thing about AI, it doesn't get tired, it doesn't get dizzy. And it can keep its inference at the same rate.”“What we try to do when we translate this to customers, to non deeply technical folks – they're technical in other ways, but they're not dealing with AI all day – is we really try to translate it to the outcomes.”“Start your hiring process early so that you're expecting it might take a while before you really, really need that team member joined, onboarded, trained up and enabled to help deliver on projects.”“I think, for us, recruiting and thinking about what mix of talent we really need on a team, it's looking across all of those different areas and building out a team that really compliments each other's skillsets.”Links:AMP RoboticsResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

In this episode, I talk with Mahyar Salek, co-founder and CTO of Deepcell, about an AI powered technology for single cell analysis through the lense of high content cell morphology.  Deepcell's platform blends deep learning, microfluidics, and high resolution optics to deliver novel insights about cell biology and has the capability to sort, label-free for downstream multi-omic and functional analysis for use in research, translational studies, and therapeutic research.  We discussed some of the challenges and opportunities in working with single cell images and how they used self-supervised learning.Quotes:“We really use the power of computer vision and AI capabilities combined with the advances in microfluidics and imaging to create this high dimensional, high content interpretation of single cell images. And we use that in real time to purify and separate cells of interest.”“We have to see millions of cells even in just one go, one run. So you can't really do that without the scalability of an algorithm, right? And then we have to be consistent and robust.”“When I hear challenges, I equate them with opportunities and I'll tell you why. So, for instance, one of the challenges, not just with us, but any sort of AI solution that looks at biological samples is the susceptibility to artifacts.”“But as soon as you roll it out, there's a difference between your lab and the lab, you know, a block down the road because of the artifacts. So it's artifacts are definitely challenging, but for us, it's an opportunity as I mentioned, because we generate the data through our own platform and that means that we have a very controlled environment.”“Because, again, we have the full control over the imaging path and where the cells lie, where we image them, we could actually do these sort of things and come up with models that are very less reliant on labels.”“By being able to run a biological assay and validate whether the existing model, like basically errors in the existing models and existing labels, and that way you're able to iterate very quickly on your learning without even relying on arguably erroneous human labels, erroneous and obviously expensive human labels.”“Any modern life science companies that rely on data, you have to have a very tight collaboration between machine learning and data scientists and the domain experts.”“It is really important to, as you kind of come up with a development strategy and the product strategy, understand where you could rely on AI today versus where you hope that the AI could deliver, you know, two years down the road.”Links:DeepcellResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

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:ProsciaResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

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:SimBioSysResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.

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 AgTechTwitterYouTubeResources for Computer Vision Teams:LinkedIn – Connect with Heather.Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.Computer Vision Advisory Services – Monthly advisory services to help you strategically plan your CV/ML capabilities, reduce the trial-and-error of model development, and get to market faster.