3 ms·

When it comes to machine learning, the subtlety here is that there are at least two sides or facets to machine learning: (1) training and (2) inference. It's

When it comes to machine learning, the subtlety here is that there are at least two sides or facets to machine learning: (1) training and (2) inference. It's fair to say that there are multiple areas for AI leadership. It is generally believed that: Model Creation (1) Those with the access to the best (which is not necessarily the most, but often believed to be) data have a strong starting point for training models; because of this Google, Facebook, and Microsoft have often been attributed to have this advantage due to the nature of their businesses. Model Application (2) Inference/prediction at the edge, e.g. on-device, is believed to be the best point for applying those models; this can be for a variety of reasons, including latency and other costs associated with sending model input data from edge sensors/devices. Some applications are entirely impractical or likely impossible to achieve without conducting inference on-device. Privacy-preservation is also a property of this approach. Depending on how you want to view this, this property could be a core design principle or a side-effect. Apple's hardware ecosystem approach and marketshare (i.e. iPhones) provide a strong starting point for making the technology ubiquitous for consumer experiences.

llampx 6y ago
Re: Prediction at the edge, I would think that it's better if there aren't going to be any updates to the model. Or if internet access is limited. Correct me if I'm wrong, but most of the ML inference actually takes place on the cloud nowadays, not on-device.
- the_gipsy 6y ago
  IIRC, for example phone keyboards use federated learning, where the model is further trained locally. You don't want to send every word the user types, for privacy reasons and others. Some kind of "diff" of the local model can then be sent to the cloud at some time to add up to the base model. There is still an attack vector, you can infer a bit from the "diff", but you probably can't tell exactly what the user wrote.
- tialaramex 6y ago
  Here's a nice example that like a lot of the best things is always passively present. My Pixel 2 knows what music it hears. There's a pop song playing, I kinda like it. I could pay attention to the lyrics and try to Google them or ask somebody that might know what it is... no need, I just look at my phone, "Break My Heart by Dua Lipa" it says on the lock screen. The phone will remember it heard this, so if I get home this evening and check what was that... oh, "Break My Heart by Dua Lipa". Google builds a model and sends it to phones that opted in to enable this service. It's not large, and I actually don't know how often it's updated - every day? Every week? Every month? No clue. But the actual matching happens on the device, where it's most useful and least privacy invading.
antipaul 6y ago
But can you lead in (2) without being good at (1)? If you train on limited data, then your inferences will be of poor quality, even if they have low latency. So in sum, it seems you can't be a leader in ML without both.
- moandcompany 6y ago
  You can lead in (2) with good enough (1). Being a leader in (1) does not mean you'll be good at (2), and vice versa. There's also a difference between limited data and good enough data. If you train on good enough data, you can have good enough models. If people believe the focus of AI/ML should just be precision/recall, or other measures of accuracy, and having tons of data, they're missing many other areas and elements for what make AI/ML successful in application.

llampx 6y ago

Re: Prediction at the edge, I would think that it's better if there aren't going to be any updates to the model. Or if internet access is limited. Correct me if I'm wrong, but most of the ML inference actually takes place on the cloud nowadays, not on-device.

the_gipsy 6y ago
IIRC, for example phone keyboards use federated learning, where the model is further trained locally. You don't want to send every word the user types, for privacy reasons and others. Some kind of "diff" of the local model can then be sent to the cloud at some time to add up to the base model. There is still an attack vector, you can infer a bit from the "diff", but you probably can't tell exactly what the user wrote.
tialaramex 6y ago
Here's a nice example that like a lot of the best things is always passively present. My Pixel 2 knows what music it hears. There's a pop song playing, I kinda like it. I could pay attention to the lyrics and try to Google them or ask somebody that might know what it is... no need, I just look at my phone, "Break My Heart by Dua Lipa" it says on the lock screen. The phone will remember it heard this, so if I get home this evening and check what was that... oh, "Break My Heart by Dua Lipa". Google builds a model and sends it to phones that opted in to enable this service. It's not large, and I actually don't know how often it's updated - every day? Every week? Every month? No clue. But the actual matching happens on the device, where it's most useful and least privacy invading.

antipaul 6y ago

But can you lead in (2) without being good at (1)? If you train on limited data, then your inferences will be of poor quality, even if they have low latency. So in sum, it seems you can't be a leader in ML without both.

moandcompany 6y ago
You can lead in (2) with good enough (1). Being a leader in (1) does not mean you'll be good at (2), and vice versa. There's also a difference between limited data and good enough data. If you train on good enough data, you can have good enough models. If people believe the focus of AI/ML should just be precision/recall, or other measures of accuracy, and having tons of data, they're missing many other areas and elements for what make AI/ML successful in application.