Biography
Arturo Deza is the CEO & co-founder of Artificio, a Peruvian startup collecting and crunching driving data for self-driving vehicles’ training. Artificio’s innovation comes from collecting driving data from emerging markets, which is less “ordered” than Western driving data, helping self-driving vehicles handle more edge cases.
Prior to Artificio, Arturo earned a PhD in dynamical neuroscience from UC Santa Barbara. He was then a postdoctoral research associate at Harvard, and a postdoctoral research associate at MIT. He also taught at UTEC, a Peruvian university.
Artificio has raised a $250,000 pre-seed round. It has collected a total of 500 hours of driving data mainly between Lima, Peru’s capital and New York City, in the United States. They have also mapped half of the district of Barranco in Lima using photogrammetry techniques, with drones.
Can you explain the nature of your academic work, and how it informs how Artificio is built?
There are two currents when it comes to training AI models.
The first is feeding the model trillions of data points. At some point, the model will start recognizing patterns and make relevant inferences. The model’s performance depends on scale: the computing power you can access, and the volume of data you can feed your model. It’s a rather lazy method that requires very little heuristics or “inductive biases” (ie: human “rules of thumb” the model is fed) but it works.
The second is constructing your model to replicate aspects of how the human brain functions. This is the neuroscience approach, the one I specialized in. This exploits inductive biases and structure to take better advantage of the nature of the data. This requires exponentially less data points, but gets to an AI model that is aligned in representation to humans and thus more robust.
For example, it takes between 10 and 20 hours for a 16 year old to drive normally. They can be dropped in a brand new city, observe, and adapt to the different driving conditions in minutes. There’s clearly something exceptional going on inside the human brain.
This new sub-field has been called “NeuroAI”. I was lucky to be part of this revolution during my time at MIT’s Center for Brains, Minds and Machines [2020-2022], which has pioneered this research movement since 2013.
The current AI zeitgeist is too focused on the first current whereas the golden insights lie in the second: developing AI models that require less data.
In the self-driving space, we’ve recently witnessed the emergence of “visual language action models”. Those have beautifully combined both the data hungry and the inductive bias approach. These models learn the cause-and-effect relationships behind specific driving situations by mapping an action from the input video data (ex: when the video shows this, this is what the driver does); rather than simply labelling every object and predicting what the car should do.
Artificio’s technical ethos results from those elements: a focus on understanding how the human brain handles driving, and collecting dynamic, “cause-and-effect” driving data rather than just mass labeling objects and random video data.
How does this concretely translate into Artificio’s product?
We collect live driving data which includes two things: the actual visual driving data (through a camera on the driver’s dashboard) and the corresponding actions the driver took (determined by the GPS, an accelerometer, and an IMU or Inertial Measurement Unit.…).
Lima’s driving is notoriously chaotic and unpredictable, which is great for us. The more "bizarre" driving data we collect the better, as that data is more useful to autonomous driving companies. They want to ensure their cars know how to handle difficult “driving edge cases”.
In the field of machine learning, training systems with corner case scenarios is known as “adversarial training”. 10 years of research in the field has already demonstrated that systems trained with these inputs are more robust to edge cases when deployed in production.
We then hyperselect the out-of-distribution (OOD) data through our own AI model and license it by the hour. The price per hour depends on the complexity of the data a company wants to buy: whether it is the video data from cars, drones, or both. In many cases, this data comes with different options, such as extra meta-tags and GPS data. It depends on what the client needs for their tech stack.
We can thus deliver exponential value to clients by selling “golden ticket” packets of hundreds or thousands of driving hours, rather than millions (as our driving hours are far richer). That boosts training speed and robustness at a fraction of the cost.
We realized that the driving data we collected was also useful for mapping purposes, as Google Maps isn’t always up to date here. We’ve purchased drones, to add more preciseness to the maps we’re building. Currently, more than half of Lima’s Barranco neighborhood has been mapped through our drones.
RO insights: accurate maps for emerging markets
As Arturo mentions, “classic” digital maps providers such as Google Maps aren’t as precise in emerging markets than they are in the West. This is a combination of higher technical complexity and weak economic incentives.
Here’s how Tayef Sarker, co-founder of Bangladeshi startup Barikoi (which builds better maps for emerging markets) explains:
“Google adequately invests in markets they know will be lucrative. Bangladesh neighbors India ( (a larger, more affluent market), where Google obviously preferred to spend more resources.
Building reliable maps for Bangladesh required significant investments because a company like Google can’t apply a Western lens here. Addresses aren’t organized through names, they are situated in relation to a specific “landmark” (ex: the house is “just behind the big tree”). In the US, the government publishes useful data for map makers. Not so much in Bangladesh. The granular, data collection work has to be done from scratch, by hand.”
Excerpt from Barikoi: accurate maps for emerging markets, originally published in The Realistic Optimist
How does this concept of “adversarial training” play out in Artificio’s product?
Many self-driving car models are trained on clean data. Driving data from quiet, leafy, ordered suburbs. This omits the infinite number of “edge cases” a driver can be exposed to (a car going the wrong direction, a child jumping across the street to retrieve their ball, a drunk driver speeding up from behind you). Models need to be trained to handle this.
Our driving data comes from one of the most chaotic driving environments on the planet. This is great for models.
As mentioned in one of our blog posts (that cites some of my previous academic research and former collaborators from MIT):“by training systems in adverse and heavily out-of-distribution stimuli, neural network representations have a higher chance of performing better when faced with both the expected & the unexpected, in addition to being closer to learn brain-aligned representations” - Artificio (Berrios & Deza. SVRHM 2022), MIT (Harrington & Deza. ICLR 2022), Feather et al. (Nature Neuroscience 2023)”
This adversarial training comes naturally through the chaotic driving data we collect, but we also use the same concept when training our AI to “classify” objects. We’ll intentionally try to trick it, for it to learn better. Models have a way of getting tricked in the most random, dumb ways, and we have to make sure we trick them during training for them to adapt better.
Has Artificio manufactured its own hardware?