NuronLabs

Location-specific neural networks

[Elon Musk @elonmusk]

@flcnhvy @wholemars It can simulate anywhere, but we need to focus on one region until FSD is out of beta, then expand geographically to have betas in other countries. Canada is not a lot different from US, but different enough.

[NuronLabs @NuronLabs]

Meet CarLOS (Car localized operating system) running in simulation!

[NuronLabs @NuronLabs]

Interesting real-world example of a location-specific phenomenon. Incorporating local priors can make autonomous driving a much safer experience.

[NuronLabs @NuronLabs]

Left: Waymo coverage map in SF, Right: Uber traffic in SF.
Driving in various locations, even in the same city, presents vastly different challenges.

[NuronLabs @NuronLabs]

Research ideas in AV: Domain Adaptation https://t.co/7jUIUsC2Qg

[NuronLabs @NuronLabs]

@electricfuture5 @greentheonly @elonmusk Yea, there will probably have to be country/location-specific neural networks to solve problems of this kind in the future. We are currently working on this exact problem.

[NuronLabs @NuronLabs]

Announcing https://t.co/V7naD7OUa7 https://t.co/mbm9DLePBD

[NuronLabs @NuronLabs]

@HarshSikka We're certainly working on things along those lines, we have some exciting stuff coming out in the next couple weeks if you would like to build off of our platform, so stay tuned!

[NuronLabs @NuronLabs]

Determining safe Operational Design Domains for autonomous vehicles remains a challenging task that is usually done manually. We just launched a feature on https://t.co/NzSXidXU1Q to automate this + tools to measure & debug the safety of a model. Sign up: https://t.co/YchLK4ma55

[NuronLabs @NuronLabs]

Measuring reliability and reacting in realtime is key to a safe AV system. Location-specific networks enable reliability tracking and local updates in realtime.

[NuronLabs @NuronLabs]

A simple trick to customize location-specific models.

[Zachary Nado @zacharynado]

Ever left batch norm in train mode at test time? We did, then realized it is shockingly effective at improving calibration on dataset shift! In our note "Evaluating Prediction-Time Batch Normalization for Robustness under Covariate Shift" (https://t.co/6XituVtUFR) we explore why

[NuronLabs @NuronLabs]

A first look at a reliability map of a CNN in SF using only unlabeled data!

[NuronLabs @NuronLabs]

Tips to rapidly scale Deep Learning research. https://t.co/GKn8ScRo1B

[NuronLabs @NuronLabs]

@strangecosmos Most of these cases are location-specific and are not the long-tail when you consider the local distribution. The one exception to this is when the vehicle occludes the stop sign.

[NuronLabs @NuronLabs]

Most of these are cases of distribution-shift that our platform https://t.co/zEGw36f9sn is built to handle reliably!

[hardmaru @hardmaru]

The simple task of recognizing a stop sign is harder than we can imagine. @karpathy https://t.co/vo8gqSvJ1Y

[NuronLabs @NuronLabs]

Distribution-Shift — the hidden reason self-driving cars aren’t safe yet.
https://t.co/8tXOGYJuZP

[NuronLabs @NuronLabs]

Deep Learning is a ripe field for innovation with many big ideas left to research and explore. Here are some of the opportunities we see for the decade ahead
https://t.co/2umQdMmUTj

[NuronLabs @NuronLabs]

Deep learning papers are known to be difficult to reproduce. This post covers a potential solution to the Deep Learning reproducibility crisis.
https://t.co/nBUAsVPiYg

[NuronLabs @NuronLabs]

A new blog post on one of the most useful metrics used in deep learning, Precision-Recall https://t.co/J2nIZhqyEd

[NuronLabs @NuronLabs]

We recently presented our work on https://t.co/C3ipK8M8eI at the Pioneer presentation day.

[Pioneer @pioneerdotapp]

In safety-critical situations like autonomous driving, neural networks aren't good enough at visual recognition—yet.

@bhavashok shows how Nuron Labs can change that: