Self-driving truck boss: 'Supervised machine learning doesn’t live up to the hype. It isn’t C-3PO, it’s sophisticated pattern matching'
- Reference: 1584970092
- News link: https://www.theregister.co.uk/2020/03/23/ai_roundup_march20/
- Source link:
Starksy Robotics is no more: Self-driving truck startup Starsky Robotics has shut down after running out of money and failing to raise more funds.
CEO Stefan Seltz-Axmacher bid a touching farewell to his upstart, founded in 2016, in a [1]Medium post this month. He was upfront and honest about why Starsky failed: “Supervised machine learning doesn’t live up to the hype,” he declared. “It isn’t actual artificial intelligence akin to C-3PO, it’s a sophisticated pattern-matching tool.”
Neural networks only learn to pick up on certain patterns after they are faced with millions of training examples. But driving is unpredictable, and the same route can differ day to day, depending on the weather or traffic conditions. Trying to model every scenario is not only impossible but expensive.
“In fact, the better your model, the harder it is to find robust data sets of novel edge cases. Additionally, the better your model, the more accurate the data you need to improve it,” Seltz-Axmacher said.
More time and money is needed to provide increasingly incremental improvements. Over time, only the most well funded startups can afford to stay in the game, he said.
“Whenever someone says autonomy is ten years away that’s almost certainly what their thought is. There aren’t many startups that can survive ten years without shipping, which means that almost no current autonomous team will ever ship AI decision makers if this is the case,” he warned.
If Seltz-Axmacher is right, then we should start seeing smaller autonomous driving startups shutting down in the near future too. Watch this space.
Waymo to pause testing during Bay Area lockdown: Waymo, Google’s self-driving car stablemate, [2]announced it was pausing its operations in California to abide by the lockdown orders in place in Bay Area counties, including San Francisco, Santa Clara, San Mateo, Marin, Contra Costa and Alameda. Businesses deemed “non-essential” were advised to close and residents were told to stay at home, only popping out for things like buying groceries.
It will, however, continue to perform rides for deliveries and trucking services for its riders and partners in Phoenix, Arizona. These drives will be entirely driverless, however, to minimise the chance of spreading COVID-19.
Waymo also launched its Open Dataset Challenge. Developers can take part in a contest that looks for solutions to these problems:
2D Detection: Given a set of camera images, produce a set of 2D boxes for the objects in the scene
2D Tracking: Given a temporal sequence of camera images, produce a set of 2D boxes and the correspondences between boxes across frames.
3D Detection: Given one or more lidar range images and the associated camera images, produce a set of 3D upright boxes for the objects in the scene.
3D Tracking: Given a temporal sequence of lidar and camera data, produce a set of 3D upright boxes and the correspondences between boxes across frames.
Domain Adaptation: Similar to the 3D Detection challenge, but we provide additional segments from rainy Kirkland, Washington, 100 of which have 3D box labels.
Cash prizes are up for grabs too. The winner can expect to pocket $15,000, second place will get you $5,000, while third is $2,000.
You can find out more details on the rules of the competition and how to enter [3]here . The challenge is open until 31 May.
More free resources to fight COVID-19 with AI: Tech companies are trying to chip in and do what they can to help quell the coronavirus pandemic. Nvidia and Scale AI both offered free resources to help developers using machine learning to further COVID-19 research.
Nvidia is providing a free 90-day license to Parabricks, a software package that speeds up the process of analyzing genome sequences using GPUs. The rush is on to analyze the genetic information of people that have been infected with COVID-19 to find out how the disease spreads and which communities are most at risk. Sequencing genomes requires a lot of number crunching, Parabricks slashes the time needed to complete the task.
“Given the unprecedented spread of the pandemic, getting results in hours versus days could have an extraordinary impact on understanding the virus’s evolution and the development of vaccines,” it [4]said this week.
Interested customers who have access to Nvidia’s GPUs should fill out a [5]form requesting access to Parabricks.
“Nvidia is inviting our family of partners to join us in matching this urgent effort to assist the research community. We’re in discussions with cloud service providers and supercomputing centers to provide compute resources and access to Parabricks on their platforms.”
Next up is Scale AI, the San Francisco based startup focused on annotating data for machine learning models. It is offering its labeling services for free to any researcher working on a potential vaccine, or on tracking, containing, or diagnosing COVID-19.
“Given the scale of the pandemic, researchers should have every tool at their disposal as they try to track and counter this virus,” it [6]said in a statement.
“Researchers have already shown how new machine learning techniques can help shed new light on this virus. But as with all new diseases, this work is much harder when there is so little existing data to go on.”
“In those situations, the role of well-annotated data to train models o diagnostic tools is even more critical.” If you have a lot of data to analyse and think Scale AI could help then apply for their help [7]here .
PyTorch users, AWS has finally integrated the framework: Amazon has finally integrated PyTorch support into Amazon Elastic Inference, its service that allows users to select the right amount of GPU resources on top of CPUs rented out in its cloud services Amazon SageMaker and Amazon EC2, in order to run inference operations on machine learning models.
Amazon Elastic Inference works like this: instead of paying for expensive GPUs, users select the “right amount of GPU-powered inference acceleration” on top of cheaper CPUs to zip through the inference process.
In order to use the service, however, users will have to convert their PyTorch code into TorchScript, another framework. “You can run your models in any production environment by converting PyTorch models into TorchScript,” Amazon said this week. That code is then processed by an API in order to use Amazon Elastic Inference.
The instructions to convert PyTorch models into the right format for the service have been described [8]here . ®
[1] https://medium.com/starsky-robotics-blog/the-end-of-starsky-robotics-acb8a6a8a5f5
[2] https://twitter.com/Waymo/status/1239970834527092742
[3] https://blog.waymo.com/2020/03/announcing-waymos-open-dataset-challenges.html
[4] https://blogs.nvidia.com/blog/2020/03/19/coronavirus-research-parabricks/
[5] https://www.nvidia.com/en-us/docs/nvidia-parabricks-researchers/
[6] https://scale.com/blog/scale-ai-covid-19-research
[7] https://events.scale.com/accelerating-covid-19-research
[8] https://aws.amazon.com/blogs/machine-learning/reduce-ml-inference-costs-on-amazon-sagemaker-for-pytorch-models-using-amazon-elastic-inference/
Re: Finally, a proper description of what the media dubs "AI" actually is
"AI" is a whole lot better to flog to venture capitalists than "pattern matching" which sounds a lot like comparing integers. And the worst thing about pattern matching in AI currently is that the resulting models turn out to be black boxes we neither understand, nor can tune to avoid misidentification. In short I expect a new AI winter sooner rather than later unless we have a major breakthrough in the intelligence part of AI which doesn't seem likely.
Re: Finally, a proper description of what the media dubs "AI" actually is
Trying to model every scenario is not only impossible but expensive. “In fact, the better your model, the harder it is to find robust data sets of novel [new] edge cases. Additionally, the better your model, the more accurate the data you need to improve it,” Seltz-Axmacher said.
In other words, AI is really G.I.G.O., proving once again you are only as good as your data set.
Re: Finally, a proper description of what the media dubs "AI" actually is
But why is this news? All along "AI" has been pattern matching, and has always been shown to be pattern matching. I remember early demonstrations of AI telling the difference between headshots of men and women, and then being confused when given headshots of the Beatles.
The real question is, can we use pattern matching to reliably navigate a multi-ton object without human intervention? Sure, if the object is on rails, and isn't subject to major random interference.
IOW, the current crop of artificial "intelligence" does neither see, nor understand, nor reason about the world around us and it makes it extremely dangerous to rely on it.
We have yet to even begin to crack the intelligence which exists even in worms and ... plants.
Such honesty
“Supervised machine learning doesn’t live up to the hype,” he declared. “It isn’t actual artificial intelligence akin to C-3PO, it’s a sophisticated pattern-matching tool.”
We have not advanced much, if at all, since 1980s "Expert Systems". I argued in the 1970s that if we knew how to do it we would simply have slow AI, if it was a question of needing more computing power. We have very much more CPU speed, RAM, storage and the Internet now. Still rubbish translation (hint it uses pattern matching and known "Rosetta translations", not real grammar and parsing as envisaged in the 1960s and 1970s). Hardly better spelling and grammar checkers than 35 years ago. Image Recognition that is really easily fooled pattern matching using big curated databases.
AI is more about hype and marketing than research or real products. A few big "tech" companies that didn't exist in the Mainframe era are more interested in selling overpriced gadgets or stealing usage data to help sell adverts. Not doing real AI research.
I learnt programming because I wanted to help develop AI. I think I read too much SF in the 1960s and 1970s. Modern SF with AI now seems very self-indulgent, egotistical and pointless. We are still no closer to defining what natural intelligence is, or why we have language but the smartest animals only seem to have vocabularies. Nor why it seems there is little correlation between brain size and tool using / problem solving / vocabulary in animals (compare corvids, dogs, horses, goats, dolphins, chimps, whales). There is none in humans.
Note IQ tests and Psychometric tests only compare people with the same background and training. Not at all intelligence, though we don't quite know what intelligence is. It's not purely about language, though that's important. A vocabulary IS NOT language. One is simply recall and matching and the other is still a bit mysterious.
Tesla exists
You know Tesla exists right?
These things that you're claiming are not coming to pass and never will, they're like, in showrooms and on your phone and on the net now, right?!
Grab Google translate now and do real-time voice to voice translation like a 'Universal Translator' of 80's SciFi movies.
This idiot didn't know how to break a complex problem down into parts small enough to make the training set for each viable, but others do. His approach needed a ridiculously large training set he could never have delivered.
But a lot of these AI mini companies are really investor or patent troll plays not companies ever intended to make the thing they raised money for. So a lot of them will fail.
Re: Tesla exists
Tesla's "self driving" works until it doesn't and you crash into something.
Google Translate works until it doesn't and the results get laughed at on social media.
There is no "Intelligence" in AI, just a lot of "Artificial".
Re: Tesla exists
Which is buggy but useful pattern recognition. Not AI. It works the same as it did in the 90s, but now fits in your hands (actually it does not, the dataset is often GBs large, and specific data pruned from a server and sent to your phone), instead of a desktop PC (they had speech to text in 1995!).
With 5g
His approach is a perfect hybrid to get self driving vehicles working. A remote driver simultaneously monitoring a few cars for the edge cases. As a more experienced driver they may even be able to handle emergencies better than a driver.
As the Software gradually improves the remote driver can handle increased number of simultaneous drives (each drive may be monitored by more than one remote driver providing redundancy)
A first use might be delivering fast food.
The service might only cost £5 an hour so popular with commuters who would rather play on the internet than drive
Re: With 5g
That's not self driving, that's remote control.
Re: With 5g
Prey tell, how does such a remote handler help when 2 incidents happen at the same time? We don't phone in aircraft pilots for a reason, and the reason is not the lag in video feed or the resolution of the cameras.
The phrase "artificial intelligence" is the real problem.
The actual phrase "artificial intelligence" has caused more problems than anything else. People hear that phrase and they think "oh, we're right around the corner from having a computer equal to a human brain." This is prima facie stupid, of course. Machine learning, expert systems, neural networks, all of these things can be trained (by humans) to automate tasks in specific problem domains. The phrase "artificial intelligence" was stupid when it became popular in the 1980's, and it was stupid again when it became popular in the 2010's. The phrase needs to be retired. Anyone using it should automatically be flagged as a non-expert.
Finally, a proper description of what the media dubs "AI" actually is
Sophisticated pattern matching. Sounds about right. When I did the Google Beginners Course in AI, I followed a dozen lectures in statistics. There wasn't a hint of AI, it was just how to define a slice of dataset to get the desired result.
Now a head of company has finally called it. Good. I'm not expecting that to actually change the media's mind, but I'm glad that somebody is putting "AI" back into its place.