Every successful business is determined to achieve practical and scalable artificial intelligence and machine learning. However, everything is much easier said than done, as can attest to the AI leaders in some of the most information-hungry companies. For more insight into the challenges of building an AI-driven organization, we caught up with Jing huang, Senior Director of Engineering and Machine Learning at Momentary (formerly SurveyMonkey). who shared the lessons learned during the deployment of AI and ML.
Q: AI and machine learning initiatives have been going on for several years now. What lessons have companies learned in terms of the most productive adoption and deployment?
Huang: “Machine learning projects are much more complicated and large than ML model algorithms, so be prepared to build a strong team to support machine learning operations. It is extremely difficult to staff a team. world-class machine learning ML talent with experience are in high demand One option is to provide training and create a culture that fosters internal transfers; sometimes internal team growth can be the key to building an effective ML team. ”
“Before you create anything substantial, be sure to examine where the bottlenecks in the machine learning production pipeline are. While deciding to build or buy, when looking for a solution to accelerate your AI / ML capabilities, make sure that the solution you choose can be tailored, scaled, and easily integrated into your business applications. products. “
Q: What technologies or technological approaches make the most difference?
Huang: “From a broader industry perspective, machine translation and information retrieval, in general, have improved significantly after the adoption of deep learning. For example, at Momentive, we’re seeing a big difference in ML solutions that help customers find relevant and actionable information through massive amounts of effortless response data. “
Q: Are most AI initiatives executed in-house or supported by external services / parties (such as cloud providers or MSPs)?
Huang: “Depending on the use case and the organization, the requirements for AI initiatives are quite different. Some of them make more sense to take advantage of external services, others need to be performed in-house. -Party services for use cases that are independent and do not need to be tightly integrated with production systems. As we see better performing in-house solutions for use cases that need to be tightly integrated with production systems. “
Q: Where are the efforts of companies to achieve fairness and eliminate bias in AI results?
Huang: “The field as a whole is still learning 00 no one has all the answers. However, awareness of the impact of bias in AI has increased in recent years and progress is being made. Increasing efforts are being made to find solutions to mitigate the risk of bias in AI and discussions of bias and fairness in ML have become a new norm in both research and industry. . “
Q: Are companies doing enough to regularly review their AI performance? What’s the best way to do it?
Huang: “There will always be human prejudice – there is no escaping it – but one thing we have done is make sure that the people who work on this topic come from various walks of life to offer a wide representation and feeling included as well. It means inclusion, not just diversity, in order to highlight all the different kinds of concerns that might be at play. “
Q: Should IT managers and staff receive more training and awareness to mitigate AI bias?
Huang: “The research on biases in AI and its mitigations is quite recent compared to the history of computing, not to say compared to human history. Universities like Stanford and MIT have started incorporating ethical AI topics into their AI courses. The assumption is that the more senior IT managers, the more they can benefit from training that covers the latest developments in this area. employed on a quarterly basis. ”
Q: What areas of the organization are most successful with AI?
Huang: “It depends. Normally these are the areas where historical data is stored and can be easily accessed. Things started to change after deep learning technology became more widely adopted, with synthetic data and adversarial training playing an increasingly important role. ”
“There are many different parts of an organization that can successfully implement AI. For example, the IT organization within the enterprise can use ML / AI technology to improve the efficiency of business processes, the financial organization can leverage ML / AI to provide more accurate information. forecast, the company could integrate ML / AI solutions into its product offering to improve customer experience, etc.