Three Key Factors Making AI Adoption Hard For Startups | April 2021

 

GETTY

The last decade has seen the advent of some remarkable technologies. We have witnessed mobile app ecosystems mature after iOS, and Android app stores were launched in the late 2000s. Anyone with a cool idea can build an app and launch it. Many tech giants such as Uber, Snapchat and Instagram were born out of this. 

We have also seen cloud computing become mainstream and enable anyone to get access to compute resources without the hassle of buying expensive servers. Artificial intelligence also got the spotlight in the same decade. Mobile apps and cloud computing are disruptive technologies that favor the underdogs (i.e., they leveled the playing field). But AI as a technology is biased in favor of large corporations over startups. 

There is no doubt that AI will stay in the digital world for a very long time. Businesses are already investing their resources to adopt AI, but it also comes with many challenges, especially for startups. A startup means fewer resources, a small team and more expectations to beat the competitive market. There are vast differences between how large corporations can use AI and how startups can use AI. 

Data Challenges

AI is data-driven, which means that the more data you have, the better results you have. Although businesses are collecting a significant amount of data today due to the world's digitalization and easy access to the internet, the amount of data startups collect is nowhere near the amount of data large corporations already have. If AI does not receive the right quantity of data, its results will be subpar. 

 Also read this What is Semantic AI

MORE FOR YOU

3 Key Cybersecurity Trends To Know For 2021 (and On ...)

What Is GPT-3 And Why Is It Revolutionizing Artificial Intelligence?

Recently, Facebook trained a computer vision model using a billion images from public Instagram accounts. Priya Goyal, a software engineer at Facebook AI Research, noted, "We inform Instagram account holders in our data policy that we use the information we have to support research and innovation, including in technological advancement like this." Regardless of how you feel about the privacy concerns, startups don't have the customer base or traffic needed to generate a significant amount of data to feed AI in this manner.  

Most startups rely on public datasets such as ImageNet for training data. ImageNet has only 14 million images even after nine years of data collection efforts by professors from the world's top universities.

Other than collecting important data, the other issue startups face is data quality. AI needs data that is categorized, labeled and correct. But as startups are at the initial phase of business, ensuring data quality and quantity is a big challenge. On the other hand, large corporations have the resources and customers to collect a significant amount of data and then get it labeled, which will significantly improve the data's quality as well.

Lack Of AI Talent 

Although it's the most overused tech buzzword, AI is still a growing field with a shortage of talent. An AI expert needs a good grasp of statistics and linear algebra, an understanding of how to build models and the ability to define a problem, its parameters and its architecture. Hence, the individuals experienced with AI are also limited, causing a shortage of technical talent. When it comes to startups, the situation becomes even worse. Due to more demand and less supply of AI experts, large corporations are in a better stage to hire skilled AI researchers and pay them high salary figures.

In fact, big tech companies are poaching AI professors, which has led to a further reduction in AI talent at the graduate level. For startups, paying a high salary is not reasonably possible, and it's more than likely that an AI expert would prefer to work with a large corporation with the resources where they can.

Computational Costs

The cost associated with AI is another spiking concern for startups. The training models of AI, such as deep learning, take significant time and computation for training. Developing a "good enough" model requires at least a few iterations of training for hyperparameter tuning and optimizations. These incur huge costs in computation power and development time as the model needs to be trained multiple times. 

Just retraining an existing model (e.g., BERT) can cost an engineer's monthly salary. Other than computational cost, startups also have to deal with MLOps infrastructure. For large corporations, this is not much of a challenge because they have significant capital, dedicated IT staff and resources to manage the computational and training costs associated with AI.

It can be challenging for startups to adopt AI compared to large corporations. With limited resources and a small customer base, deploying AI models and orienting business decisions around it is an excellent hassle for startups. AI is the future, but it is also true that it favors large corporations with the resources, expertise and time to research and deploy the latest AI models.

As we continue to see more skilled AI professionals enter into the market with MOOC initiatives and more open-source collaborative communities continue to spring up, researchers can collaborate. But still, there are no guarantees that the challenges of AI startups will be solved any time in the near future.