This is a continuation of insights shared during a recent Assembly Talk featuring Sushant Tripathy; Research Scientist in Machine Learning at Google.
Most practical examples of hybrid-AI in businesses have to do with organizations effectively deploying hybrid-AI to improve the efficiency of computing tasks. The following are a few examples of promising applications:
- Social media platform image/video tagging and Not Safe For Work (NSFW) content identification: Users predominantly access these platforms through robust devices such as iPhones, laptops, and desktops. By employing hybrid-AI, the process of tagging and identifying NSFW content can be conducted on-device, followed by a gradual confirmation process on the cloud. This approach ensures an optimal user experience and prevents any undue augmentation of the company’s cloud infrastructure, as it adheres to their customary batch processing routines.
- Bolstering eCommerce search-ranking and recommendation: Several research performed by Etsy internally as well as other entities like job boards have shown that tweaking recommendations based on the most recent browsing activity of users improves users’ engagement such as clicks on adds and purchases of items. This is made possible with the integration of hybrid-AI onto eCommerce search-ranking and recommendation mechanisms that update in real-time based on recent browsing activity of users.
- Email/Chat phrase response-generation: Email/Chat phrase response generation tools are small enough to be encoded within devices and applications to help generate automatic responses like “Thank you”, and “That sounds great!”. By running these simple models on users’ devices, computer servers such as those used to process emails avoid having to work harder to process a significantly large number of messages at peak hours. In return, organizations are able to conserve up to 30% of their resources that they would have had to invest on scaling cloud computing capacity to avoid loss of service during busy periods.
- Contact affinity generation: In this case, hybrid-AI is used to create a system where the top contacts in a user’s list are ranked by the likelihood of receiving voice or video calls, and text messages from the user. This ranked list is then sent to a central server to help with voice searches on Google Home where those contacts are also saved. What’s important to note here is that any data not allowed to be saved on the server won’t be sent there in the first place. Similarly, the list of top contacts sent to the server can’t be used to figure out the user’s original contact data. As such, this application of hybrid-AI helps organizations comply with privacy policies and protect their users’ data. Speaking of data privacy, there is a push to start generating ads based on user browsing behavior and content consumption patterns. However, this application of hybrid-AI is heavily monitored by emerging privacy policies. The future of this technology lies in overcoming this privacy protection layer, and in being able to develop and deploy generative AI models that are small enough to run on user devices.