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.

[INTRODUCTION]

[0:00:03] HC: Welcome to Impact AI, brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven machine learning-powered company.

This episode is part of a mini-series about foundation models. Really, I should say domain-specific foundation models. Following the trends of language processing, domain-specific foundation models are enabling new possibilities for a variety of applications with different types of data, not just text or images. In this series, I hope to shed light on this paradigm shift, including why it’s important, what the challenges are, how it impacts your business, and where this trend is heading. Enjoy. s

[INTERVIEW]

[0:00:49] HC: Today, I’m joined by guest, Dimitris Sakellariou, Founder and CEO of Piramidal, to talk about a foundation model for the brain. Dimitris, welcome to the show.

[0:01:00] DS: Oh, thank you for having me, Heather. Pleasure to be here.

[0:01:04] HC: Dimitris, could you share a bit about your background and how that led you to create Piramidal?

[0:01:08] DS: Yeah, for sure. I’ve studied physics as an undergrad. During my physics undergrad, I was purely interested in complex concepts, like quantum mechanics and field theory. I think it was three years into my physics undergrad when I visited a very close friend up in Edinburgh that was studying neuroscience and molecular biology. He explained to me how complex the brain is, or the system. Arguably, more complex than most of the systems in the universe.

Thereafter, I got very, very interested in understanding more about the brain, and I started participating in psychedelic studies and sleep studies that involved recording the brain with electrodes, with wires, something we call otherwise, EEG. I was struck by the fact that it’s very, very easy to record brainwaves because you just stick a couple of wires in your head. But on the other hand, these recordings, these brainwaves are very complex, and there’s no quantifiable way for humans to interpret this activity. The interpretation of the brainwaves is highly empirical still to this day. I was very interested in developing computational methods in order to understand these types of recordings.

[0:02:42] HC: What does Piramidal do? Why is it important?

[0:02:45] DS: As I just mentioned, maybe I should say that 2024 marks a 100 years since the creation of the EEG recording machine by Hans Berger. Despite the fact that, essentially, the past, we still do not fully understand what brainwaves are really telling us. The interpretation, as I mentioned, of the brainwaves is highly empirical based on a lot of trial and error. On the other hand, EEG brainwaves are pivotal for the diagnosis of many, many neurological disorders and brain disorders. At the same time, there’s an extreme shortage of neurologists in the US, but around the world as well.

Piramidal is primarily focused at the moment in automating, or otherwise democratizing the interpretation of these tests, these brainwave recordings, so that patients and people that have issues with their brain can get access to the diagnosis much, much, much faster, and in order to be able to receive successful treatment. That’s just the first step. In the longer term, we are very interested in also, being able to deliver introspection to everyone in the world. The value and the insights we can potentially extract from those brainwaves, it’s highly valuable for us to understand, for example, in a quantifiable way, what is the cognitive workload that we’re experiencing in a particular moment? How do we respond to a particular treatment or a new meditation technique that we recently started? We want to create large-scale models that are able to spawn personalized aging and start monitoring our brain in real-time and let us know when things go awry.

[0:04:54] HC: What role does machine learning play in this technology?

[0:04:56] DS: This is a very important question. Maybe I should start by saying that EEG, like brainwave patterns are extremely variable. If you compare, for example, the heart patterns, the ECG that now you can see on your Apple Watch, this is a repetitive pattern, right? The same thing more or less takes place over time, and doctors can actually track whether there are any changes in this repetitive pattern.

Now EEG is a very similar system, but it’s recording the brain and the brain generates a huge variety of patterns, like the one that our heart actually generates. Now, this huge variety of patterns is different across brain states. Different things happen when we are asleep, or during a particular sleep stages. Different things happen when we’re relaxed, or when, for example, we are experiencing some sort of neurological condition, like an epileptic seizure.

That means, A, that in order for doctors to extract meaningful insights from brain recordings, they need to record for a much, much longer time. There’s an example of all those different patterns. The problems start, the difficulty starts, because when these recordings are very lengthy, so doctors need to go over very long periods of brain waves in order to extract value. Two, because of the fact that these patterns are not exactly the same across patients and subjects, those different genders, ages, for example, the thickness of the scalp can influence all these patterns, actually look. It is for these reasons that humans have a very difficult time drawing these correlations and associations across patients. What are the similarities and the patterns in this huge variety of the expressions of the brain?

Now, coming to your questions to machine learning, we know that computational systems can actually do a very good job of drawing associations between different variations of the same patterns. This is where machine learning becomes important. There are certain narrow machine learning tools for each gene that perform very well across more tasks. What we’re trying to do here is to build like, now maybe a large-scale machine learning system that can holistically encapsulate these variety of patterns, these variety of expressions that take place in our brains.

[0:08:04] HC: Through machine learning, you’ll be able to understand the patterns better. What would be the output of a machine learning model, like a supervised model where you’re trying to predict something in the end in order to diagnose a condition, for example?

[0:08:17] DS: Yeah. I cannot share a lot of information about the architectural components of our model, because it’s at this moment still proprietary. I should say, that we want a high level to generate models that are able to exactly as you’ve put it, detect, for example, abnormalities as they take place in real-time. Maybe, also to be able to predict the onset of a disease in advance, so that we can provide early on treatment, or even prevention of the occurrence of a disease as well. Beyond medicine, as I mentioned earlier, we want to provide translate these brainwave signals into meaningful and actionable insights for our end user.

[0:09:16] HC: Part of your solution involves building a foundation model. Why build a foundation model in order to solve the tasks that you just talked about?

[0:09:25] DS: At a very high level, I should say that narrow ML systems haven’t historically worked, or maybe better put, generalized very well across different brain states and different tasks, different patients and different EEG hardware devices. Now, the goal here as in any other foundational model is to be able to reuse these model objects in order to address different downstream tasks, without having to retrain a model from scratch.

[0:10:08] HC: What are some of the challenges you’ve encountered in building a foundation model for EEG?

[0:10:13] DS: Yeah. This is a very good question. I think some of the challenges have to do with the computational data that is required for training a very big model than the one we are trying to build. Now, on the other hand, in order to build a foundation model for a domain that is different than where we commonly see the foundation model these days, like language modeling, it’s important to understand, well, the domain that you’re trying to do a foundational model at because that’s going to inform the architectural components of the model so that it is able to provide good inference for the users.

[0:11:05] HC: What did it take to build the foundation model you’re working on? Maybe you have some examples of the quantity of data, or the amount of compute, or the more details on the types of expertise and algorithms that went into this, whatever you can share about what it takes to build this type of solution.

[0:11:22] DS: Yeah. I’m not very comfortable to share a lot of information on that yet. I should say, that I think we definitely needed to raise a lot of money in order to be able to support the model-building components and get access to the computational resources we needed for the foundational model. Our raise was about 6 million and almost half of it is currently being spent towards compute.

On the other hand, we took a lot of time to prototype the model architecture at a very high level. I should say, that it’s important to identify the key ingredients of the recipe you’re trying to build, like what is, for example, the equivalent of an eigenvector transformation, that no matter what else is missing, you will still reach your destination or your goal. That’s very, very important because on one hand, it informs the design of the core computational graph of the model. That means that it’s about how you bake biases into your learning system. How do you position your model to experience the world, so that it’s able to perform the particular task, or tasks that it will be asked to train on?

This has been a relatively difficult aspect of model building. It’s very important to have discussions with neuroscientists and clinical experts in order to understand what is the end-to-end pipeline from receiving data all the way to inference.

[0:13:16] HC: Foundation models are notorious for needing a ton of data in order to train them, to put it lightly. How do you get access to this data? Are there any insights on how you’re able to gather a data in order to train a foundation model?

[0:13:31] DS: Yeah. For one, we have found that there’s a lot of open-source data repositories that in many cases are siloed, or not very well known. At least on the Internet, someone could get through open-source papers and literature and identify the big number of siloed data repositories that could be harmonized and aggregated in order to build a large enough data set to equal for the training of a model like ours.

On the other hand, it’s important to get access to proprietary data. As I said, EEG and brainwaves in our case is something that takes place in almost every big hospital around the world. It’s very important for diagnosis of neurological conditions. You know that there’s a lot of data being collected at any point in time. We’re working toward artery data agreements that I’m not in the position to share right now, but also, another very important aspect of getting access to data and high-quality data as well. [0:14:51] HC: You mentioned earlier in our conversation that generalizability is one of your motivations for building a foundation model, to create something that can work on a diverse set of patients with data collected from different places. In gathering data to train your foundation model, do you have to consider the diversity? Do you have to consider the sources of the data coming into it and think carefully about where you need to gather more data from?

[0:15:15] DS: Such an important question. There are two moving parts to this equation. I should say, one is, as I said, having, or creating a diverse enough universe of data, a diverse enough corpus of data, so that the model gets exposed to the different patterns and interesting things that take place in the brain.

On the other hand, it’s very easy to spend a lot of compute. The important part is A, to build a diverse enough corpus of data, so that the model learns and manages to internalize a wide variety of patterns of the brain and wide variety of systemical signal attributes that originate all the way back to the hardware, and store a lot to some corner of its model parameters.

On the other hand, it’s important to monitor how the model is learning, how its parameters get saturated over time by being exposed to one particular data set. If it’s not receiving updates and at some point, it’s not learning too much from that particular data set, maybe it’s wise to stop learning from that data source for a little bit and maybe move on to a different data source, so that the model also gets exposed to different things before going through too many steps that means computational resources and therefore, costs.

[0:16:47] HC: How are you using this foundation model? Are there things you’re fine-tuning it for today, models you’ve already built off of it? how do you plan to use it in the future?

[0:16:55] DS: Yeah. This is a super interesting question. A couple of things I wanted to say here. Going back to my initial point, currently doctors spend an absurd amount of time reviewing EEGs. These recordings, like in long-term monitoring units, or in ICUs, these recordings can last for days, to different weeks. For this reason, in many cases, diagnosis is massively delayed or missed. There’s many, many cases of misdiagnosis for neurologic patients around the world. In order to make diagnosis accessible and democratized for all these patients, the key component is to make the process of reading the EEGs and extracting the clinical value, much, much more easy.

We, on one hand, our high-level goal is to make the interpretation, the analysis of the EEG as frictionless as possible, or even fully automated, as a first step. Now, as a second step, we don’t necessarily want to create all the vertical use cases ourselves, but we would like to be able to provide our foundational model through an API so that neuro-tech developers and tech developers can ping our model in order to create their own vertical applications, that being in healthcare, or in use cases that we haven’t even thought about yet.

[0:18:40] HC: These other use cases, well, others have access to your model in order to be able to adapt it to new applications?

[0:18:48] DS: Yeah. Down the line, this is our goal, for our model to actually act as the middleware between these recordings. The middleware between the hardware and the developers that want to use the model to find unit towards a particular goal that they want to serve their users, or provide inference for patients.

[0:19:16] HC: How will others get access to your model? Is this an open-source, open-data component, or is this something that’s going to be licensed?

[0:19:24] DS: For the time being, we keep the model and the data proprietary, and we haven’t made any decisions with that yet.

[0:19:33] HC: Are there any lessons you’ve learned in developing foundation models that could be applied more broadly to other types of data?

[0:19:38] DS: this is an interesting question. I feel, this is very similar to an answer I gave earlier. But I think the first important lesson is have very good knowledge, very good grasp of the domain that you’re building a foundational model in, in order to be able to introduce biases to through the architectural design backbone of the model. Two, get high-quality evils around what the model should be doing so that learning is geared towards the correct direction. There is at point more important around how a team hub project can spend money in a wise way and carefully scaling up the model before the project runs out of money.

[0:20:36] HC: Thinking more broadly, beyond foundation models and to your role as a founder, is there any advice you could offer to other leaders of AI-powered startups, or future founders?

[0:20:46] DS: Yeah. Yeah. Find a co-founder, if you don’t have one. It is a very interesting journey that requires a lot of decision-making. Finding the right person. Someone that is very keen to build together with you and make important and difficult decisions can change massively a trajectory of your company. Two, try to get into Y Combinator. I think that also massively changed the way I was thinking about a startup and how we operate in general before going there. We managed 50 cents.

[0:21:31] HC: Finally, where do you see the impact of Piramidal in three to five years?

[0:21:35] DS: I don’t think we are very far from a future where everyday devices carry, or have neural sensors embedded onto them. With models like ours, I would definitely see a future where personalized agents monitor our brain health and a lot of things around our wellbeing. For example, our cognitive workload, or how our stress levels may have increased recently, because of overexposure to screen time, how our attention has been decreasing, because of certain activities that we took up recently, or even implement interesting strategies to boost memory enhancement, or memory consolidation during intense periods of learning. All that powered by these personalized agents launched by models like ours.

[0:22:38] HC: This has been great. Dimitris, I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:22:46] DS: We have a website at piramidal.ai. We have a LinkedIn page as well. Most of the information is actually there.

[0:22:56] HC: Perfect. Thanks for joining me today.

[0:22:58] DS: Thank you. Thank you so much, Heather. It’s been a pleasure. And thank you for having me.

[0:23:03] HC: All right, everyone. Thanks for listening. I’m Heather Couture and I hope you join me again next time for Impact AI.

[END OF INTERVIEW]

[0:23:12] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share with a friend. If you’d like to learn more about computer vision applications for people and planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]