Sensory prides itself on being one of the few deep learning companies working with vision and voice that quite intentionally does not try to collect personal information.
We believe that private information should stay private, and we traditionally felt that the best way to accomplish this was to not let any data leave the device. For that reason, Sensory’s technology and models have always run on-device with no data whatsoever going to the cloud.
While our embedded speech and vision AI models are accurate and secure by design, there are many companies that either prefer or require cloud-based solutions. For example, some applications require the power of GPU processing while others need the flexibility to leverage real device data for continuous accuracy improvements.
Now Sensory is going to add cloud-based AI solutions to our technology stack, but the core tenets of our belief in privacy are not changing. In fact, the benefits of on device – speed, always available, and privacy will not go away with our hybrid solutions that will enable more flexibility and better accuracy, while offering some opportunity to lower device-based costs and power consumption.
But let me focus on privacy and data in this blog. Companies collect voice and vision data for two main purposes:
- To build better AI models. To build better models whether with voice or face, we need the real-world usage data: How people speak, the statistics of what they say, what front-end noise filters are doing to the sound, how they position their face in the camera, etc. But none of this requires knowing the user, seeing the user, or even hearing their actual voice. For example, a text translation of what the users say could provide the statistics and grammars around what gets spoken.
- To better target the user for sales/advertising. Most of today’s trillion-dollar companies are based around some form of advertising. That’s how Google and Facebook make most of their money. And to do this well, the companies need to know us. They want to know who we are (demographics like age, gender, race, etc), what mood we are in, and what we are interested in. Thus, they collect a lot of data, build profiles and have sophisticated models that let them know the best approach to selling to us.
Companies in the speech recognition space that offer cloud-based solutions, usually have business models based around option #2. For example, if I do a search (whether voice or text) the results could contain ads, they could be ranked by who is paying the most, and it could be that whatever I asked about gets stored, so I start seeing ads for the things in my queries (everybody has seen that happen, right?).
Sensory has decided to offer hybrid client/cloud solutions to have faster more accurate solutions without sacrificing privacy, so our plan is to fulfill purpose #1 (as described above) but not succumb to the privacy busting demands of purpose #2.
We will tell you more about our cloud-based AI approach over the coming year, but we are confident it will lead to higher quality products without sacrificing privacy! Better for our customers and consumers alike!