View in #questions-forum on Slack
@Suri_Bhasker_Sri_Harsha: Hello everyone, I have a small question:
Self driving car companies like Tesla and Comma.ai are claiming that their models get updated as new data arrives over time. How are companies like Tesla and comma.ai updating their neural networks with new data without catastrophic forgetting? Does anyone have any idea? The last time I saw, Tesla was using neural networks for its self driving system. It is not like they are using some other kind of model. Any ideas?
@Martin_Mundt: Hey @Suri_Bhasker_Sri_Harsha. I don’t work at any of these companies or am in any form affiliated. But what I can imagine is that the Tesla scenario that’s currently being employed doesn’t require continual learning in a fashion that needs to solve catastrophic forgetting. To the best of my knowledge, there is no actual update online or in any other form while a car is driving. So I imagine it could just be -> gather and accumualte data -> retrain/fine-tune model -> upload/deploy as it basically done in traditional engineering workflows.
Someone correct me if all of this is wrong 
@Suri_Bhasker_Sri_Harsha: @Martin_Mundt But dont you think that retraining the model on the entire dataset everytime is very expensive? And the training time only keeps on increasing with every passing session…
@Martin_Mundt Do you know any companies that have deployed neural networks and are facing the “forgetting problem”?
@Martin_Mundt: I don’t think companies like Tesla are facing the problem of running out of compute resources yet. If you asked me personally, then If I were Tesla I would rather re-train the model from scratch a bunch of times (I dont except them to this daily anyway because the improvement would be insignificant) than run into a variety of safety concerns (or rather even more of them).
I’m not aware of any companies where CL is actively included into the production stage in the way that many papers analyze the catastrophic forgetting problem, i.e. changing number of classes, types etc. I think the real-world scenarios are much different than what is analyzed academically. In a car driving through a city you wouldn’t really expect it to suddenly forget what a building is because you no longer present it with buildings etc… I would expect this to be maybe a little bit more important in RL based stuff like Robotics, where there is more of a need to quickly adapt to something. I’m not an expert on robotics however.
@Suri_Bhasker_Sri_Harsha: @Martin_Mundt Nice points. Thanks.
@vlomonaco: I’ve written a blog post on the topic:
https://medium.com/continual-ai/continual-learning-for-production-systems-304cc9f60603
There are many companies using some sort of basic CL algorithms now, I will update this wiki section as soon as I can: https://wiki.continualai.org/industry.html
Medium: Continual Learning for Production Systems