Anti-Data-Science: applying Taleb’s barbell in the AI world [2/2]
In the previous article, we started to apply the concept of Taleb’s barbell to the AI projects, trying to differentiate between stable and conservative AI applications and risky yet much more rewarding ones. We have looked only at the first group through the business model (and realized that it’s linear, routine-work-automation oriented), technology (with the obvious conclusion that it’s very data-hungry), and people (having data scientists and integrators-communicators and project owners with business, pattern-seeking experience). Of course, the alternative we’re looking at should be the opposite of the conservative approach, with more exposed risks, but even better potential returns. Let’s discover how do we build such AI ventures.
Useful mental models
Data, knowledge, and models
If you work in the AI/ML field, you constantly hear about people “training models” that can predict something from the inputs. Funny fact, but for us, as a humanity, were building models without AI and big data for centuries and even could go to the moon without all this fancy stuff. The point is that we can obtain models working in two directions:
- with small data by doing careful hands-on analysis and building equations from that (classic “blackboard science” from the movies)
- with big data relying on computational power and AI to find patterns and equations faster than humans (see the previous article)
If you merge those two together, you get physics-aware machine learning, which we discuss a bit later.
Exponential thinking
Linear iterative improvements with direct effort-result relationships are embodied in us and are very simple to understand. Exponential thinking requires seeing opportunities to invest today more compared to linear thinking and even lose short-term but obtain big advantages and extra-returns later when the investments will pay off. People are impatient with results and are short-term thinkers and energy consumption minimizes by biological design, however, when we try to design long-term abnormal returns opportunities, we need to practice this kind of thinking. Useful exercise: start spotting the curves instead of flat lines in your life (mathematically speaking of course).
Extremely aggressive investments in AI
The business model of aggressive AI
Two good reference applications for non-linear business outcomes are:
- drug discovery: if we can use AI to generate more drug compounds artificially and test them in-silico instead of the real labs' experiments, the curve of process speed-up is not linear, it’s exponential
- predictive maintenance: the earlier we can spot the pattern that leads to the machinery breakdown, the cheaper will be the cost of fixing and replacement. And this cost grows not linearly, but exponentially over time.
As you can notice, the data is coming from the labs or from the engineers and their machines. It’s not digital by nature, it’s digitalized. Since this information has to be measured from the physical world, it’s usually harder, longer, and more expensive to collect. That’s why in such applications you usually have much smaller datasets. That’s why zero-data. Start seeing the pattern yet? Do you know other examples of such improvement curves?
The technology of aggressive AI
We discussed before that with small-to-zero data there is a way to merge two methods of knowledge model building — handcrafting equations and training ML models like deep neural networks. One of the scenarios is using physical equation-based models to simulate more data for us. This way we can use small real data and vast simulated data to train our ML models and operate in the space closer to “big data scenarios”.
Another interesting point is, that often modeling in the physical world is narrowed down not to finding patterns (due to the lack of the data) but finding extremums. In the case of predictive maintenance, it’s finding extremal abnormal behavior, in the case of drug discovery, it’s the generation of a set of molecules that satisfy chemical and target constraints. Looks a bit more difficult than face detection exercises with deep learning, right?
People of aggressive AI
It’s also worth to mention, that many prominent companies who operate in such domains are created by scientists for scientists and aim to empower creative and scientific work. For example, Citrine Informatics builds a toolkit for material science research and Insilico Medicine is driving drug discovery research. CEOs have a profound scientific background, the end-users in both cases are scientists as well who generate new materials and molecules.
The big questions about small data
To summarize thoughts from those articles, I’ve compiled the tables above. How to use all this information? Apply the same rule as Taleb applied to the investments barbell: do not invest in the middle, invest only in safe, well-calculated investments or in risky yet high-return-promising ones. In the case of our projects, it means knowing very well on which side you are. Do you digitalize processes and deal with the documents? Evaluate data size, pipelines, and business knowledge of stakeholders. Doing predictive maintenance? Add physical models, find extremities, attract scientists and domain experts to the process. Why it’s so important? Let’s check a couple of cases where projects stood on the middle of the barbell trying to mix things that shouldn’t be mixed.
Why you shouldn’t mix the sides of the barbell?
Healthcare is one of my favorite domains that desires to be non-linear, has expensive data from the physical world, and AI applications have to help doctors to make the right decisions on a scale in the first place. However, everyone is still trying the “big data” paradigm there. The very known fail example is, of course, oncology-oriented AI by IBM Watson. What mistakes were done there?
- Trying to solve “routine” doctor tasks instead of creative and hard ones;
- 53K data samples, commoditized algorithms (which require much more data)
- Organized by business people, done by integrators, and in the end, it’s a no-go story
Let's look at the alternative story of DeepMind’s AlphaFold:
- The hard scientific challenge related to generation and discoveries
- 350K samples, although still “classic” approaches are used as a constraint optimization
- Done by scientists with scientists for scientists
The dichotomy of our AI barbell seems like holding here. Do you know some other examples?
Conclusions
Of course, this model is nothing but my personal observation and an attempt to structure the projects at my company Neurons Lab. Our focus is on the “risky” and “aggressive” projects and deep tech startups that need fast go-to-market. The analogies of business models, technologies, and people helped us to recommend the right actions and project approaches for our clients and I hope that this blog will help you to do so as well. If you’ll have any questions or AI R&D support, don’t hesitate to ping me via email or personal messages anywhere.
P.S.
I am open to discussions and collaborations in the technology space, you can connect with me on Facebook or LinkedIn, where I regularly post some AI-related articles or news opinions that are too short for Medium.