HTM could assist with self-driving vehicle problem?

Saw this article today, and it makes me think that perhaps, if we might be able to figure out how to successfully combine the two, that HTM could help solve this issue. It’d certainly push it out into the spotlight.

The basic problem is this: Deep Neural Networks, when classifying images, have a problem where noise can cause errant classifications from random noise.

I think there’s good reason to believe that HTM could be thrown into the mix, to assist as perhaps a gating function in a neural network (i.e. “There isn’t any object there, it’s just random noise” vs. “This really is something to pay attention to. Classify it.”).

My theorized, back of an envelope model would be image features coming from a neural network (convolutional neural network or similar) being converted to SDRs, then feeding into an HTM system (SP and/or TM). The HTM system would determine if it was just random noise, or something worth looking at and sending for further processing.

I don’t have time at the moment for a project like this, but if somebody else out there does, feel free to take this and run with it. It apparently is a $7 trillion USD problem, just waiting for somebody to find a solution.

1 Like

AFAIK, That’s what exactly YOLO does.
And it’s very common to use YOLO when you use CNN to classify images.
But apparently, it’s not sufficient to work safely enough?

YOLO is definitely being used in those systems, but setting the threshold of when something counts or not is an arcane art; too low, and you have a bunch of false positives. Too high, and you run over pedestrians. The article mentions that WEMO’s vans are probably set more to the lower side, (probably after having run over a bike that was crossing where it shouldn’t have been).