The AI forum session of InnoVEX 2019 featured great speakers from prominent companies. One of them is Dr. Min Sun, the Chief AI Scientist of Appier who discussed how AI can make real business impact.
3 Stages of AI History
Dr. Sun divided the development of AI into 3 stages; anything before 2012 is considered the history of AI, between 2012 and 2019 is the third wave of AI, and after 2019 is the future of AI. Historically, the development of AI has been a bumpy ride; with the community having experienced 2 AI winters; a period of reduced funding and research in the technology. AI winters generally start with an AI technology breakthrough such as rule based AI and expert systems that show a lot of potential for possible future uses. However, the breakthroughs were then considered as not having enough business impact which will prompt the government to shutdown funding while industrial investment also slows down.
Now in the third wave of AI, there is another major breakthrough: Deep Learning. The main difference is how it is a representation learning tool that is used to remove or replace feature engineering tasks for humans. Not only that, Deep Learning is very generalizable and has shown great performance on speech recognition, image domain, and natural language processing (NLP). Deep learning continues to breakthrough and has started to make innovations in visual and speech recognition.
In the past few years; deep learning has been combined with other AI technology such as reinforcement learning which resulted in AlphaGo. In addition, deep learning has also shown great potential to learn representation from raw text and can learn on its own with potential for comprehension of over 100 languages just by learning from Wikipedia alone.
More than just Technology
Dr. Sun stated that AI is not only about the technology; as the third wave of AI happens because AI had help from other technologies and industries. Currently it is possible to collect a massive amount of data and in the future with more and more sensors in circulation, more IoT data will exist in the ecosystem. Computing platforms also continue to increase in speed and capability; including CPUs, GPUs, APUs, etc.
In addition, the open source movement also greatly helps greatly in the development of AI. Because of the open source platforms, research in academia and industry labs can publish in open source platforms such as GitHub and be validated by people everywhere. Perhaps in just half a year, people can build downstream applications through the technologies and start creating businesses from them. The question then is how can we work together to maintain the growth of AI in the future? According to Dr. Sun, there are 2 missing piece needed to reach the next generation of AI: Human centered AI and Automated Machine Learning.
All of the developed AI will interact or be used by humans. As such, it is important to consider how AI can work with humans and increase their productivity; rather than thinking about how to make AI that surpasses human capabilities.
B2B users of AI models will most likely need certain minor tweaks or modifications for the model to serve their needs better. These adjustments can be done automatically through Automated Machine Learning so the model can be more easily generalized and scale with little to no extra costs.
The key to know the unknown
Dr. Sun proposed 7 steps that are needed before an AI model can be deployed: Problem definition, Data, Feature, Model, Optimization, Evaluation, and finally Deployment. During the first 2 steps; it is crucial for the developers to define the right problems that need to be solved. In the context of human centered AI, the developers need to think about how to define the right problems where AI will generate added value; special business impacts to work with the humans.
AI is very powerful tool, but not necessarily applied in everywhere. Developers also need to consider what problems can be solved with the added values generated from AI. Once the problems have been clearly defined, the data collection method is next. While most of the processes can be automated, developers need to realize that at the end of the day, their solutions’ purpose is to solve human problems. This means humans and the human factor need to be inserted strategically in the loop at the right place to make sure that collected data is high quality in terms of correctness and diversity.
From Dr. Sun’s experience, the attention should be expanded to cover more than just the technology in terms of human centered AI. Once the processes have been completed; good problems defined; and good potentials shown; the next question will be on the issue of scaling. How can the solutions be scaled up to reach and serve more customers? Developers might need to do different feature engineering for different customers and design slightly different model to be able to put the model into production in a robust way.
In the research of Automated Machine Learning, tools such as Deep Learning can save a lot of time from having to manually do feature engineering. Other tools such as Architecture Search can also help developers design slightly different neural network models for different customers. In addition, certain companies including Appier are also building their own engineering processes and toolkits to help push the models into production at a very robust and efficient way to be able to scale as effectively as possible.
By keeping the problems that users want to solve in mind, 2 more questions that need to be answered are: how can AI bring unique values in terms of addressing those problems? How can AI make a difference?
Traditionally, solutions come from rigorous data analysis and visualization that focuses too much on the past experiences of the users. Leveraging AI’s predictive powers mean developers can predict what might happen in the future and then suggest the right decision to make based on those possibilities. AI can understand a large amount of unstructured data and can extract insights or intelligence from the unstructured data which could not be analyzed before. In a way, this means AI can be a key to understanding what was previously unknown.
Conclusion
Dr. Sun stated that currently AI is experiencing major growth from 2012 to 2019. To maintain the momentum going, it is vital that the community and ecosystem focus on 2 topics for the next generation of AI: the human centered AI and automated machine learning.
Developers must remember that when they are developing problems and collecting data; they must put the user value first and think about how AI can generate added value to bring business impact. Automated Machine Learning is an existing tool that can be used to scale the AI model’s performance to multiple customers at low to no cost; l this tool will also help the developers customize their models to fit the users’ requirements at little to no cost.
To watch the full forum session, visit our YouTube channel here.