Cell phones would drop calls 1,000x more often without Bayesian probability bounds
Ben Vigoda · A New Approach to Machine Intelligence // Ben Vigoda, Gamalon (FirstMark's Data Driven) · May 22, 2018 · at 21:46
Ben Vigoda, CEO of Gamalon, draws an analogy between probabilistic natural language understanding and Qualcomm cell phone receiver chipsets at Data Driven NYC.
0:00 / 0:06exact quote · 6.2s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“You would drop calls a thousand times more often if there were, if cell phone receivers didn't found uncertainty in one patient.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Ben Vigoda
Ben Vigoda: Neurons and synapses are the wrong abstraction for AI
“Neurons and synapses is not the right abstraction. Models should be programs, specifically programs that simulate the system that generated the data.”
Ben Vigoda: Machine learning variables must include uncertainty and error bars
“Every variable that you're trying to infer in a program should come with an uncertainty. Neural networks today are just, they're just a number in the neuron. It's like an activation level. But you need error bars around those numbers.”
Ben Vigoda: Deep learning is excellent at instinct but poor at thought
“AI is really good at instinct. I mean deep learning, it's not so great at thought.”
Changing deep learning target categories requires retraining models from scratch
“But if you want to change your columns, your categories, then you have to redo all your multiple choice tests and then retrain the system. And if you want to change the categories, you have to start over.”
A major New York bank receives 1.2 billion text messages annually
“One big bank in, in New York City that we've been talking to gets 1.2 billion of these little text messages a year.”
Gamalon reduced an automaker's text processing costs from $1.25 to 10 cents
“We're able to work with them for a few weeks and generate a model, and they're actually, we're paying a dollar 25 per utterance to get them read. It took months to do it every year, and now it takes 25 milliseconds, and they pay 10 cents.”