What Challenges Do AI Startups Face?
8 February, 2022
Complex Algorithms and Training of AI Models
In a typical AI or SaaS startup, you mainly deal with challenges that are not even there or have not been previously solved. This is not the case with AI-centric applications where even the best AI model will never perform with 100% accuracy.. That is primarily because we deal with uncertainties. After all, the foundation of AI and machine learning (ML) is probability and statistics. When deploying an algorithm to forecast the sales, the weather, or the stock price, all of that is based on statistics based on dynamic data. Therefore, in such a case, you are handling a non-deterministic case.
Similarly, in the case of building autonomous cars, AI engineers have to set up an algorithm that will most likely recognize the obstacles in front of them. However, someday there might be a car in front with a bicycle image or other image on it that the AI needs to recognize, and at that moment, if the AI fails to identify it, there might be an accident. There are many edge cases in a day like this, and finding the right solution for this is crucial.
For AI to Succeed, You Need Continual Data and Process with Feedback Loop
As more and more businesses are inclining towards AI technology, there is an influx of better solutions. The first step towards this would be to build an intelligent system that resembles the human brain. The focus of AI is to create something that imitates the human brain or the humanistic ways of thinking. However, today's most updated technology does not have the algorithm or the capability to build that yet.
We have been focusing on the algorithm or the ML models, but we must put bigger and clearer attention to the systems and processes. No one algorithm can fix the complicated non-deterministic cases in a dynamic environment. Again, if we draw the example of an autonomous car, it's a dynamic environment, and there are many possibilities in one street. There are other possibilities that something can take place as well. A combination of building processes based on autonomous and Human-In-The loop systems is the key for successful AI deployment in the real world. This is widely known as ModelOps or MLOPs.
We kept a stream of data that our system would autonomously analyze. Once it detects something strange or beyond the Operating Conditions, they would notify the human, and from then, they would take over the design and make their decision. This is called the Active Learning mechanism and applies to both human users AND the team of engineers in the background, where they have to take over the car, or reannotate, retrain and redeploy the model. The problem with this system is that it may sound straightforward, but in reality, the architects designing the systems will have to work through the processes to perfection.
In this case, we broke down the processes into making the entire system automated:
- Data operations: the system would acquire the data and then have real human work on it such as cleaning and annotating the data.
- Building the model: using the particular data sets, we built training models for that and have a bunch of models, which are then validated.
- Roll into production: we rolled our system into production, and after that, the system would autonomously monitor to see if there was uncertainty or drift on the production data.
Over time as our process and model perfected, we will find ourselves dealing with lesser and lesser non-deterministic cases or edge cases. Although consumers or users perceive AI and ML to be simple, in actuality, it is super complicated. No single tool or platform can handle all of these processes. We need to assemble the required tools and build the process in a way that we are designing the process from scratch.
One Size Does Not Fit All
- Note that there is no single tool yet anyone can rely on when dealing with AI to build the system. You have to be mindful and make a lot of errors to do the end-to-end process and automate it.
- As the head of AI, your job is not to create the algorithm or ML models; your main role is to develop the team, processes, and tools while still exploring the state-of-the-art AI methods to be infused into the production process. The environment and challenges will dynamically keep changing, and it's essential to be watchful that the 'one size fits all' principle may not work here. When your process and tools start performing better, you can think of fully automating them.
- Have a complete feedback loop to amplify changes and make things more stable.
Scale your coaching effort for your engineering and product teams
Develop yourself to become a stronger engineering / product leader
Jord Sips, Senior Product Manager at Mews, shares his expertise on a common challenge for product managers – finding root causes and solutions.
Senior Product Manager at Mews
Jonathan Belcher, Engineering Manager at Curative, explains how to balance team cohesion and individual focus time, tapping into his experiences of working remotely for seven years.
Engineering Manager - Patient Experience at Curative
Snehal Shaha, Lead Technical Program Manager at Momentive (fka SurveyMonkey), details her short-term technical strategy to unify processes among teams following an acquisition.
Senior EPM/TPM at Apple Inc.
Pavel Safarik, Head of Product at ROI Hunter, shares his insights on how to deal with disagreements about prioritization when building a product.
Head of Product at ROI Hunter
Kamal Qadri, Senior Manager at FICO, drives the importance of setting expectations when optimizing large-scale requirements.
Head of Software Quality Assurance at FICO
You're a great engineer.
Become a great engineering leader.
Plato (platohq.com) is the world's biggest mentorship platform for engineering managers & product managers. We've curated a community of mentors who are the tech industry's best engineering & product leaders from companies like Facebook, Lyft, Slack, Airbnb, Gusto, and more.