AI Experts Suggest 13 Roadblocks & Limitations to AI Progression in 2021

0

With the success of RE-WORK's previous expert-led AI roadblocks blog, they decided to reach out to a new group of experts working in the field, and ask them what they think are the biggest challenges faced in 2021. Obviously, COVID has played a huge part in blocking potential AI development, but pandemic aside, they wanted to know what the experts think we need to overcome for development in 2021.

TL;DR 2021 Roadblocks:

  • Poor standards of education in the areas needed to develop AI skills at early school age
  • The levels of compute available to companies with lower budgets harbouring the levels of ML development actually available in CPU and GPU
  • The level of ML and AI hardware in certain locations, specifically in India and China, is not high enough for the training of the next generation of developers
  • The lack of training data widely available
  • Data collection without harbouring data privacy in certain countries
  • The need for the highest skill sets available in combination with humongous amounts of money for infrastructure purposes
  • Data standardization, data bias, and data silos are all problems to be solved
  • Mis-communication between those in and those outside the technical teams of AI-led companies

See the full expert opinions below.


Lavi Nigam, Data Scientist, Gartner

"Constantly evolving landscape of tools, algorithms, and practices makes it hard for any enterprise to adopt and deliver on Machine Learning. There are also many pieces of the data science puzzle which are yet to be matured like - model interpretability, model/data bias tracking, auto-deployment metrics, AutoML frameworks, etc. This constantly evolving process also leads to skillset issues in the industry since large data scientist populations have to constantly evolve at a faster pace and new entrants have to catch up a lot with new advancements."

Ravi Dalal, Senior Computer Vision Engineer, Walmart

"The biggest roadblock, in my opinion, is Compute. I don't think everyone can afford a GPU to train generative models which require tons of data and compute to even get started. The real escalation in the trajectory of ML development will only happen when we are able to democratize it for everyone. And that can only occur when we are able to build affordable chips that can provide GPU flops on a CPU itself. Then everyone will be able to contribute to this AI/ML development journey."

Zhiyong (Sean) Xie, Director, AI, Pfizer

Zhiyong gave three main roadblocks that he thinks are holding up AI and Machine Learning development:

1. "Data. In the medical field, we need medical experts to do annotation. Data sharing is not straightforward due to privacy and other issues. Data standardization, data bias, and data silos are all problems to be solved."

2. "Methods. When large training data is not available, new methods need to be developed based on the problem and available data. We may also need another revolution to overcome the limits  of current methods."

3. "Communication. ML scientists need to truly understand the problem."

Manmeet Singh, Machine Learning Lead, Apple

"There are many depending on the application of ML. Developing truly personalized models without violating user privacy is a hard problem. Running extremely large models (like BERT) requires a lot of parameter tuning and expertise. Not only do they require highly skilled professionals but also a humongous amount of money needs to be spent on infrastructure to meet the computation power."

Indu Khatri, Machine Learning Lead, HSBC

"I believe the biggest roadblock for developing Machine Learning is the lack of labeled data applicable for real-world applications. Think of the glass manufacturing industry - there is usually no open-source training data available for identifying manufacturing defects. You'd have to generate your own training data to solve this problem. There are only a handful of firms who even recognize this as a problem - we are far from working towards systematically solving it."

Remo Storni, Senior Software Engineer, Apple

"For the adoption and development in the industry big challenges are the development of ethical AI systems, the collection of data, computational power and engineering systems. All this coupled with the application to the well-scoped business problems."

Piero Molino, Staff ML / NLP Research Scientist, Stanford University. Previously co-founder and Senior Research Scientist at Uber AI

"Right now only very few companies and institutions have the combination of data and hardware needed to train cutting edge models and they also attract most of the talent. I believe there's no silver bullet, but work on democratization of all aspects of the stack, from open-source software to collection of open real-world datasets, to the development of cheaper and more specialized hardware, and finally, the development of higher-level abstractions that make ML more accessible as a tool to a broader slice of the population, are all efforts that together may lead to a more evenly spread access to ML."

Yaman Kumar, PhD Computer Science, University of Buffalo

Yaman suggested that it's a problem which affects different geographical companies in different ways. "(It's a) Hardware lottery. As soon as you give adequate hardware to kids in India and China, you will see the next million ML developers coming from there."

Jason Gauci, Software Engineering Manager, Facebook AI

Jason's answer was short and sweet. Simply put, Jason suggested that the main roadblock to AI development globally is "Poor education around statistics at all levels of education, particularly high school and early college."

Kiana Alikhademi, Research Assistant, University of Florida

Kiana's answer was three-fold. "There are several roadblocks to the development of machine learning systems including social implications, ethical consequences, and a clear view of the limitations of deep learning in a particular context. Nevertheless, the main roadblock, in my opinion, is to collect the data without compromising user privacy."

Shuo Zhang, Senior Machine Learning Engineer, Bose Corporation

As with many other answers to this question, both above and in other articles, Shuo's response centred around a very common problem which is seen in organisations both large and small. Simply, "Being able to learn more efficiently without needing a lot of training data."