machine learning
Auto Added by WPeMatico
No rules, no problem: DeepMind’s MuZero masters games while learning how to play them
DeepMind has made it a mission to show that not only can an AI truly become proficient at a game, it can do so without even being told the rules. Its newest AI agent, called MuZero, accomplishes this not just with visually simple games with complex strategies, like Go, Chess and Shogi, but with visually complex Atari games.
The success of DeepMind’s earlier AIs was at least partly due to a very efficient navigation of the immense decision trees that represent the possible actions in a game. In Go or Chess these trees are governed by very specific rules, like where pieces can move, what happens when this piece does that, and so on.
The AI that beat world champions at Go, AlphaGo, knew these rules and kept them in mind (or perhaps in RAM) while studying games between and against human players, forming a set of best practices and strategies. The sequel, AlphaGo Zero, did this without human data, playing only against itself. AlphaZero did the same with Go, Chess and Shogi in 2018, creating a single AI model that could play all these games proficiently.
But in all these cases the AI was presented with a set of immutable, known rules for the games, creating a framework around which it could build its strategies. Think about it: If you’re told a pawn can become a queen, you plan for it from the beginning, but if you have to find out, you may develop entirely different strategies.
This helpful diagram shows what different models have achieved with different starting knowledge. Image: DeepMind
As the company explains in a blog post about their new research, if AIs are told the rules ahead of time, “this makes it difficult to apply them to messy real world problems which are typically complex and hard to distill into simple rules.”
The company’s latest advance, then, is MuZero, which plays not only the aforementioned games but a variety of Atari games, and it does so without being provided with a rulebook at all. The final model learned to play all of these games not just from experimenting on its own (no human data) but without being told even the most basic rules.
Instead of using the rules to find the best-case scenario (because it can’t), MuZero learns to take into account every aspect of the game environment, observing for itself whether it’s important or not. Over millions of games it learns not just the rules, but the general value of a position, general policies for getting ahead and a way of evaluating its own actions in hindsight.
This latter ability helps it learn from its own mistakes, rewinding and redoing games to try different approaches that further hone the position and policy values.
You may remember Agent57, another DeepMind creation that excelled at a set of 57 Atari games. MuZero takes the best of that AI and combines it with the best of AlphaZero. MuZero differs from the former in that it does not model the entire game environment, but focuses on the parts that affect its decision-making, and from the latter in that it bases its model of the rules purely on its own experimentation and firsthand knowledge.
Understanding the game world lets MuZero effectively plan its actions even when the game world is, like many Atari games, partly randomized and visually complex. That pushes it closer to an AI that can safely and intelligently interact with the real world, learning to understand the world around it without the need to be told every detail (though it’s likely that a few, like “don’t crush humans,” will be etched in stone). As one of the researchers told the BBC, the team is already experimenting with seeing how MuZero could improve video compression — obviously a very different problem than Ms. Pac-Man.
The details of MuZero were published today in the journal Nature.
Powered by WPeMatico
Turing nabs $32M more for an AI-based platform to source and manage engineers remotely
As remote work continues to solidify its place as a critical aspect of how businesses exist these days, a startup that has built a platform to help companies source and bring on one specific category of remote employees — engineers — is taking on some more funding to meet demand.
Turing — which has built an AI-based platform to help evaluate prospective, but far-flung, engineers, bring them together into remote teams, then manage them for the company — has picked up $32 million in a Series B round of funding led by WestBridge Capital. Its plan is as ambitious as the world it is addressing is wide: an AI platform to help define the future of how companies source IT talent to grow.
“They have a ton of experience in investing in global IT services, companies like Cognizant and GlobalLogic,” said co-founder and CEO Jonathan Siddharth of its lead investor in an interview the other day. “We see Turing as the next iteration of that model. Once software ate the IT services industry, what would Accenture look like?”
It currently has a database of some 180,000 engineers covering around 100 or so engineering skills, including React, Node, Python, Agular, Swift, Android, Java, Rails, Golang, PHP, Vue, DevOps, machine learning, data engineering and more.
In addition to WestBridge, other investors in this round included Foundation Capital, Altair Capital, Mindset Ventures, Frontier Ventures and Gaingels. There is also a very long list of high-profile angels participating, underscoring the network that the founders themselves have amassed. It includes unnamed executives from Google, Facebook, Amazon, Twitter, Microsoft, Snap and other companies, as well as Adam D’Angelo (Facebook’s first CTO and CEO at Quora), Gokul Rajaram, Cyan Banister and Scott Banister, and Beerud Sheth (the founder of Upwork), among many others (I’ll run the full list below).
Turing is not disclosing its valuation. But as a measure of its momentum, it was only in August that the company raised a seed round of $14 million, led by Foundation. Siddharth said that the growth has been strong enough in the interim that the valuations it was getting and the level of interest compelled the company to skip a Series A altogether and go straight for its Series B.
The company now has signed up to its platform 180,000 developers from across 10,000 cities (compared to 150,000 developers back in August). Some 50,000 of them have gone through automated vetting on the Turing platform, and the task will now be to bring on more companies to tap into that trove of talent.
Or, “We are demand-constrained,” which is how Siddharth describes it. At the same time, it’s been growing revenues and growing its customer base, jumping from revenues of $9.5 million in October to $12 million in November, increasing 17x since first becoming generally available 14 months ago. Current customers include VillageMD, Plume, Lambda School, Ohi Tech, Proxy and Carta Healthcare.
Remote work = immediate opportunity
A lot of people talk about remote work today in the context of people no longer able to go into their offices as part of the effort to curtail the spread of COVID-19. But in reality, another form of it has been in existence for decades.
Offshoring and outsourcing by way of help from third parties — such as Accenture and other systems integrators — are two ways that companies have been scaling and operating, paying sums to those third parties to run certain functions or build out specific areas instead of shouldering the operating costs of employing, upsizing and sometimes downsizing that labor force itself.
Turing is essentially tapping into both concepts. On one hand, it has built a new way to source and run teams of people, specifically engineers, on behalf of others. On the other, it’s using the opportunity that has presented itself in the last year to open up the minds of engineering managers and others to consider the idea of bringing on people they might have previously insisted work in their offices, to now work for them remotely, and still be effective.
Siddarth and co-founder Vijay Krishnan (who is the CTO) know the other side of the coin all too well. They are both from India, and both relocated to the Valley first for school (post-graduate degrees at Stanford) and then work at a time when moving to the Valley was effectively the only option for ambitious people like them to get employed by large, global tech companies, or build startups — effectively what could become large, global tech companies.
“Talent is universal, but opportunities are not,” Siddarth said to me earlier this year when describing the state of the situation.
A previous startup co-founded by the pair — content discovery app Rover — highlighted to them a gap in the market. They built the startup around a remote and distributed team of engineers, which helped them keep costs down while still recruiting top talent. Meanwhile, rivals were building teams in the Valley. “All our competitors in Palo Alto and the wider area were burning through tons of cash, and it’s only worse now. Salaries have skyrocketed,” he said.
After Rover was acquired by Revcontent, a recommendation platform that competes against the likes of Taboola and Outbrain, they decided to turn their attention to seeing if they could build a startup based on how they had, basically, built their own previous startup.
There are a number of companies that have been tapping into the different aspects of the remote work opportunity, as it pertains to sourcing talent and how to manage it.
They include the likes of Remote (raised $35 million in November), Deel ($30 million raised in September), Papaya Global ($40 million also in September), Lattice ($45 million in July) and Factorial ($16 million in April), among others.
What’s interesting about Turing is how it’s trying to address and provide services for the different stages you go through when finding new talent. It starts with an AI platform to source and vet candidates. That then moves into matching people with opportunities, and onboarding those engineers. Then, Turing helps manage their work and productivity in a secure fashion, and also provides guidance on the best way to manage that worker in the most compliant way, be it as a contractor or potentially as a full-time remote employee.
The company is not freemium, as such, but gives people two weeks to trial people before committing to a project. So unlike an Accenture, Turing itself tries to build in some elasticity into its own product, not unlike the kind of elasticity that it promises its customers.
It all sounds like a great idea now, but interestingly, it was only after remote work really became the norm around March/April of this year that the idea really started to pick up traction.
“It’s amazing what COVID has done. It’s led to a huge boom for Turing,” said Sumir Chadha, managing director for WestBridge Capital, in an interview. For those who are building out tech teams, he added, there is now “No need for to find engineers and match them with customers. All of that is done in the cloud.”
“Turing has a very interesting business model, which today is especially relevant,” said Igor Ryabenkiy, managing partner at Altair Capital, in a statement. “Access to the best talent worldwide and keeping it well-managed and cost-effective make the offering attractive for many corporations. The energy of the founding team provides fast growth for the company, which will be even more accelerated after the B-round.”
PS. I said I’d list the full, longer list of investors in this round. In these COVID times, this is likely the biggest kind of party you’ll see for a while. In addition to those listed above, it included [deep breath] Founders Fund, Chapter One Ventures (Jeff Morris Jr.), Plug and Play Tech Ventures (Saeed Amidi), UpHonest Capital (Wei Guo, Ellen Ma), Ideas & Capital (Xavier Ponce de León), 500 Startups Vietnam (Binh Tran and Eddie Thai), Canvas Ventures (Gary Little), B Capital (Karen Appleton Page, Kabir Narang), Peak State Ventures (Bryan Ciambella, Seva Zakharov), Stanford StartX Fund, Amino Capital, Spike Ventures, Visary Capital (Faizan Khan), Brainstorm Ventures (Ariel Jaduszliwer), Dmitry Chernyak, Lorenzo Thione, Shariq Rizvi, Siqi Chen, Yi Ding, Sunil Rajaraman, Parakram Khandpur, Kintan Brahmbhatt, Cameron Drummond, Kevin Moore, Sundeep Ahuja, Auren Hoffman, Greg Back, Sean Foote, Kelly Graziadei, Bobby Balachandran, Ajith Samuel, Aakash Dhuna, Adam Canady, Steffen Nauman, Sybille Nauman, Eric Cohen, Vlad V, Marat Kichikov, Piyush Prahladka, Manas Joglekar, Vladimir Khristenko, Tim and Melinda Thompson, Alexandr Katalov, Joseph and Lea Anne Ng, Jed Ng, Eric Bunting, Rafael Carmona, Jorge Carmona, Viacheslav Turpanov, James Borow, Ray Carroll, Suzanne Fletcher, Denis Beloglazov, Tigran Nazaretian, Andrew Kamotskiy, Ilya Poz, Natalia Shkirtil, Ludmila Khrapchenko, Ustavshchikov Sergey, Maxim Matcin and Peggy Ferrell.
Powered by WPeMatico
Boast.ai raises $23M to help businesses get their R&D tax credits
Nobody likes dealing with taxes — until the system works in your favor. In many countries, startups can receive tax credits for their R&D work and related employee cost, but as with all things bureaucracy, that’s often a slow and onerous task. Boast.ai aims to make this process far easier, by using a mix of AI and tax experts. The company, which currently has about 1,000 customers, today announced that it has raised a $23 million Series A round led by Radian Capital.
Launched in 2012 by co-founders Alex Popa (CEO) and Lloyed Lobo (president), Boast focuses on helping companies — and especially startups — in the U.S. and Canada claim their R&D tax credits.
“Globally, over $200 billion has been given in R&D incentives to fund businesses, not only in the U.S. and Canada, but the U.K., Australia, France, New Zealand, Ireland give out these incentives,” Lobo explained. “But there’s huge red tape. It’s a cumbersome process. You got to dive in and figure out work that qualifies and what doesn’t. Then you’ve got to file it with your taxes. Then if the government audits you, it’s like a long, laborious process.”
Image Credits: Boast.ai
After working on a few other startup ideas, the co-founders decided to go all-in on Boast. And in the process of working on other ideas, they also realized that AI wasn’t going to be able to do it all, but that it was getting good enough to augment humans to make a complex process like dealing with R&D tax credits scalable.
“The way I think to bootstrap a company is three things,” Lobo explained. “One, customers are looking for an outcome. Get them that outcome in the fastest, cheapest way possible. Two, when you’re doing that, you may have to do a lot of manual work. Figure out what those manual touch points are and then build the workflow to automate that. And once you have those two things, then you’ll have enough data to start working on artificial intelligence and machine learning. Those are the key learnings that we learned the hard way.”
So after doing some of that manual work, Boast can now automatically pull in data using tech tools like JIRA and GitHub and a company’s financial tools like QuickBooks, Gusto and (soon) ADP. It then uses its algorithms to cluster this data, figure out how much time employees spend on projects that would qualify for a tax credit and automate the tax filing process. Throughout the process — and to interact with the government if necessary — the company keeps humans in the loop.
“So all our [customer success] team is engineers,” Lobo noted. “Because if you don’t have engineers they can’t inform the decision-making process. They help figure out if there are any loose ends and then they deal with the audits, communicating with the government and whatnot. That’s how we’re able to effectively get SaaS-like margins or more.”
Ideally, a tool like Boast pays for itself and the company says it has secured more than $150 million in R&D tax credits since launch. Currently, it’s also doubling growth year over year, and that’s what made the founders decide to raise outside money for the first time. That funding will go toward increasing the sales team (which is currently only four people strong) and improving the platform, but Lobo was clear that he doesn’t want to be too aggressive. The goal, he said, is not to have to raise again until Boast can hit the $30 to $50 million revenue mark.
Once fully implemented, Boast also effectively becomes a system of record for all R&D and engineering data. And indeed, that’s the company’s overall vision, with the tax credits being somewhat of a Trojan horse to get to this point. By the middle of next year, the team plans to offer a new product around R&D-based financing, Lobo tells me.
Over the years, the Boast team also focused on not just growing its customer base but also the overall startup ecosystem in the markets in which it operates, with a special focus on Canada. The Boast team, for example, is also the team behind the popular annual Traction conference in Vancouver, Canada (Disclosure: I’ve moderated sessions at the event since its inception). A thriving startup ecosystem creates a larger client base for Boast, too, after all — and coincidently, the team met its investors at the event, too.
Powered by WPeMatico
Arthur.ai snags $15M Series A to grow machine learning monitoring tool
At a time when more companies are building machine learning models, Arthur.ai wants to help by ensuring the model accuracy doesn’t begin slipping over time, thereby losing its ability to precisely measure what it was supposed to. As demand for this type of tool has increased this year, in spite of the pandemic, the startup announced a $15 million Series A today.
The investment was led by Index Ventures with help from newcomers Acrew and Plexo Capital, along with previous investors Homebrew, AME Ventures and Work-Bench. The round comes almost exactly a year after its $3.3 million seed round.
As CEO and co-founder Adam Wenchel explains, data scientists build and test machine learning models in the lab under ideal conditions, but as these models are put into production, the performance can begin to deteriorate under real-world scrutiny. Arthur.ai is designed to root out when that happens.
Even as COVID has wreaked havoc throughout much of this year, the company has grown revenue 300% in the last six months smack dab in the middle of all that. “Over the course of 2020, we have begun to open up more and talk to [more] customers. And so we are starting to get some really nice initial customer traction, both in traditional enterprises as well as digital tech companies,” Wenchel told me. With 15 customers, the company is finding that the solution is resonating with companies.
It’s interesting to note that AWS announced a similar tool yesterday at re:Invent called SageMaker Clarify, but Wenchel sees this as more of a validation of what his startup has been trying to do, rather than an existential threat. “I think it helps create awareness, and because this is our 100% focus, our tools go well beyond what the major cloud providers provide,” he said.
Investor Mike Volpi from Index certainly sees the value proposition of this company. “One of the most critical aspects of the AI stack is in the area of performance monitoring and risk mitigation. Simply put, is the AI system behaving like it’s supposed to?” he wrote in a blog post announcing the funding.
When we spoke a year ago, the company had eight employees. Today it has 17 and it expects to double again by the end of next year. Wenchel says that as a company whose product looks for different types of bias, it’s especially important to have a diverse workforce. He says that starts with having a diverse investment team and board makeup, which he has been able to achieve, and goes from there.
“We’ve sponsored and work with groups that focus on both general sort of coding for different underrepresented groups as well as specifically AI, and that’s something that we’ll continue to do. And actually I think when we can get together for in-person events again, we will really go out there and support great organizations like AI for All and Black Girls Code,” he said. He believes that by working with these groups, it will give the startup a pipeline to underrepresented groups, which they can draw upon for hiring as the needs arise.
Wenchel says that when he can go back to the office, he wants to bring employees back, at least for part of the week for certain kinds of work that will benefit from being in the same space.
Powered by WPeMatico
AWS expands on SageMaker capabilities with end-to-end features for machine learning
Nearly three years after it was first launched, Amazon Web Services’ SageMaker platform has gotten a significant upgrade in the form of new features, making it easier for developers to automate and scale each step of the process to build new automation and machine learning capabilities, the company said.
As machine learning moves into the mainstream, business units across organizations will find applications for automation, and AWS is trying to make the development of those bespoke applications easier for its customers.
“One of the best parts of having such a widely adopted service like SageMaker is that we get lots of customer suggestions which fuel our next set of deliverables,” said AWS vice president of machine learning, Swami Sivasubramanian. “Today, we are announcing a set of tools for Amazon SageMaker that makes it much easier for developers to build end-to-end machine learning pipelines to prepare, build, train, explain, inspect, monitor, debug and run custom machine learning models with greater visibility, explainability and automation at scale.”
Already companies like 3M, ADP, AstraZeneca, Avis, Bayer, Capital One, Cerner, Domino’s Pizza, Fidelity Investments, Lenovo, Lyft, T-Mobile and Thomson Reuters are using SageMaker tools in their own operations, according to AWS.
The company’s new products include Amazon SageMaker Data Wrangler, which the company said was providing a way to normalize data from disparate sources so the data is consistently easy to use. Data Wrangler can also ease the process of grouping disparate data sources into features to highlight certain types of data. The Data Wrangler tool contains more than 300 built-in data transformers that can help customers normalize, transform and combine features without having to write any code.
Amazon also unveiled the Feature Store, which allows customers to create repositories that make it easier to store, update, retrieve and share machine learning features for training and inference.
Another new tool that Amazon Web Services touted was Pipelines, its workflow management and automation toolkit. The Pipelines tech is designed to provide orchestration and automation features not dissimilar from traditional programming. Using pipelines, developers can define each step of an end-to-end machine learning workflow, the company said in a statement. Developers can use the tools to re-run an end-to-end workflow from SageMaker Studio using the same settings to get the same model every time, or they can re-run the workflow with new data to update their models.
To address the longstanding issues with data bias in artificial intelligence and machine learning models, Amazon launched SageMaker Clarify. First announced today, this tool allegedly provides bias detection across the machine learning workflow, so developers can build with an eye toward better transparency on how models were set up. There are open-source tools that can do these tests, Amazon acknowledged, but the tools are manual and require a lot of lifting from developers, according to the company.
Other products designed to simplify the machine learning application development process include SageMaker Debugger, which enables developers to train models faster by monitoring system resource utilization and alerting developers to potential bottlenecks; Distributed Training, which makes it possible to train large, complex, deep learning models faster than current approaches by automatically splitting data across multiple GPUs to accelerate training times; and SageMaker Edge Manager, a machine learning model management tool for edge devices, which allows developers to optimize, secure, monitor and manage models deployed on fleets of edge devices.
Last but not least, Amazon unveiled SageMaker JumpStart, which provides developers with a searchable interface to find algorithms and sample notebooks so they can get started on their machine learning journey. The company said it would give developers new to machine learning the option to select several pre-built machine learning solutions and deploy them into SageMaker environments.
Powered by WPeMatico
AWS announces SageMaker Clarify to help reduce bias in machine learning models
As companies rely increasingly on machine learning models to run their businesses, it’s imperative to include anti-bias measures to ensure these models are not making false or misleading assumptions. Today at AWS re:Invent, AWS introduced Amazon SageMaker Clarify to help reduce bias in machine learning models.
“We are launching Amazon SageMaker Clarify. And what that does is it allows you to have insight into your data and models throughout your machine learning lifecycle,” Bratin Saha, Amazon VP and general manager of machine learning told TechCrunch.
He says that it is designed to analyze the data for bias before you start data prep, so you can find these kinds of problems before you even start building your model.
“Once I have my training data set, I can [look at things like if I have] an equal number of various classes, like do I have equal numbers of males and females or do I have equal numbers of other kinds of classes, and we have a set of several metrics that you can use for the statistical analysis so you get real insight into easier data set balance,” Saha explained.
After you build your model, you can run SageMaker Clarify again to look for similar factors that might have crept into your model as you built it. “So you start off by doing statistical bias analysis on your data, and then post training you can again do analysis on the model,” he said.
There are multiple types of bias that can enter a model due to the background of the data scientists building the model, the nature of the data and how they data scientists interpret that data through the model they built. While this can be problematic in general it can also lead to racial stereotypes being extended to algorithms. As an example, facial recognition systems have proven quite accurate at identifying white faces, but much less so when it comes to recognizing people of color.
It may be difficult to identify these kinds of biases with software as it often has to do with team makeup and other factors outside the purview of a software analysis tool, but Saha says they are trying to make that software approach as comprehensive as possible.
“If you look at SageMaker Clarify it gives you data bias analysis, it gives you model bias analysis, it gives you model explainability it gives you per inference explainability it gives you a global explainability,” Saha said.
Saha says that Amazon is aware of the bias problem and that is why it created this tool to help, but he recognizes that this tool alone won’t eliminate all of the bias issues that can crop up in machine learning models, and they offer other ways to help too.
“We are also working with our customers in various ways. So we have documentation, best practices, and we point our customers to how to be able to architect their systems and work with the system so they get the desired results,” he said.
SageMaker Clarify is available starting to day in multiple regions.
Powered by WPeMatico
SingleStore, formerly MemSQL, raises $80M to integrate and leverage companies’ disparate data silos
While the enterprise world likes to talk about “big data”, that term belies the real state of how data exists for many organizations: the truth of the matter is that it’s often very fragmented, living in different places and on different systems, making the concept of analysing and using it in a single, effective way a huge challenge.
Today, one of the big up-and-coming startups that has built a platform to get around that predicament is announcing a significant round of funding, a sign of the demand for its services and its success so far in executing on that.
SingleStore, which provides a SQL-based platform to help enterprises manage, parse and use data that lives in silos across multiple cloud and on-premise environments — a key piece of work needed to run applications in risk, fraud prevention, customer user experience, real-time reporting and real-time insights, fast dashboards, data warehouse augmentation, modernization for data warehouses and data architectures and faster insights — has picked up $80 million in funding, a Series E round that brings in new strategic investors alongside its existing list of backers.
The round is being led by Insight Partners, with new backers Dell Technologies Capital, Hercules Capital; and previous backers Accel, Anchorage, Glynn Capital, GV (formerly Google Ventures) and Rev IV also participating.
Alongside the investment, SingleStore is formally announcing a new partnership with analytics powerhouse SAS. I say “formally” because they two have been working together already and it’s resulted in “tremendous uptake,” CEO Raj Verma said in an interview over email.
Verma added that the round came out of inbound interest, not its own fundraising efforts, and as such, it brings the total amount of cash it has on hand to $140 million. The gives the startup money to play with not only to invest in hiring, R&D and business development, but potentially also M&A, given that the market right now seems to be in a period of consolidation.
Verma said the valuation is a “significant upround” compared to its Series D in 2018 but didn’t disclose the figure. PitchBook notes that at the time it was valued at $270 million post-money.
When I last spoke with the startup in May of this year — when it announced a debt facility of $50 million — it was not called SingleStore; it was MemSQL. The company rebranded at the end of October to the new name, but Verma said that the change was a long time in the planning.
“The name change is one of the first conversations I had when I got here,” he said about when he joined the company in 2019 (he’s been there for about 16 months). “The [former] name didn’t exactly flow off the tongue and we found that it no longer suited us, we found ourselves in a tiny shoebox of an offering, in saying our name is MemSQL we were telling our prospects to think of us as in-memory and SQL. SQL we didn’t have a problem with but we had outgrown in-memory years ago. That was really only 5% of our current revenues.”
He also mentioned the hang up many have with in-memory database implementations: they tend to be expensive. “So this implied high TCO, which couldn’t have been further from the truth,” he said. “Typically we are ⅕-⅛ the cost of what a competitive product would be to implement. We were doing ourselves a disservice with prospects and buyers.”
The company liked the name SingleStore because it is based a conceptual idea of its proprietary technology. “We wanted a name that could be a verb. Down the road we hope that when someone asks large enterprises what they do with their data, they will say that they ‘SingleStore It!’ That is the vision. The north star is that we can do all types of data without workload segmentation,” he said.
That effort is being done at a time when there is more competition than ever before in the space. Others also providing tools to manage and run analytics and other work on big data sets include Amazon, Microsoft, Snowflake, PostgreSQL, MySQL and more.
SingleStore is not disclosing any metrics on its growth at the moment but says it has thousands of enterprise customers. Some of the more recent names it’s disclosed include GE, IEX Cloud, Go Guardian, Palo Alto Networks, EOG Resources, SiriusXM + Pandora, with partners including Infosys, HCL and NextGen.
“As industry after industry reinvents itself using software, there will be accelerating market demand for predictive applications that can only be powered by fast, scalable, cloud-native database systems like SingleStore’s,” said Lonne Jaffe, managing director at Insight Partners, in a statement. “Insight Partners has spent the past 25 years helping transformational software companies rapidly scale-up, and we’re looking forward to working with Raj and his management team as they bring SingleStore’s highly differentiated technology to customers and partners across the world.”
“Across industries, SAS is running some of the most demanding and sophisticated machine learning workloads in the world to help organizations make the best decisions. SAS continues to innovate in AI and advanced analytics, and we partner with companies like SingleStore that share our curiosity about how data and analytics can help organizations reimagine their businesses and change the world,” said Oliver Schabenberger, COO and CTO at SAS, added. “Our engineering teams are integrating SingleStore’s scalable SQL-based database platform with the massively parallel analytics engine SAS Viya. We are excited to work with SingleStore to improve performance, reduce cost, and enable our customers to be at the forefront of analytics and decisioning.”
Powered by WPeMatico
Tecton.ai nabs $35M Series B as it releases machine learning feature store
Tecton.ai, the startup founded by three former Uber engineers who wanted to bring the machine learning feature store idea to the masses, announced a $35 million Series B today, just seven months after announcing their $20 million Series A.
When we spoke to the company in April, it was working with early customers in a beta version of the product, but today, in addition to the funding, they are also announcing the general availability of the platform.
As with their Series A, this round has Andreessen Horowitz and Sequoia Capital co-leading the investment. The company has now raised $60 million.
The reason these two firms are so committed to Tecton is the specific problem around machine learning the company is trying to solve. “We help organizations put machine learning into production. That’s the whole goal of our company, helping someone build an operational machine learning application, meaning an application that’s powering their fraud system or something real for them […] and making it easy for them to build and deploy and maintain,” company CEO and co-founder Mike Del Balso explained.
They do this by providing the concept of a feature store, an idea they came up with and which is becoming a machine learning category unto itself. Just last week, AWS announced the Sagemaker Feature store, which the company saw as major validation of their idea.
As Tecton defines it, a feature store is an end-to-end machine learning management system that includes the pipelines to transform the data into what are called feature values, then it stores and manages all of that feature data and finally it serves a consistent set of data.
Del Balso says this works hand-in-hand with the other layers of a machine learning stack. “When you build a machine learning application, you use a machine learning stack that could include a model training system, maybe a model serving system or an MLOps kind of layer that does all the model management, and then you have a feature management layer, a feature store which is us — and so we’re an end-to-end life cycle for the data pipelines,” he said.
With so much money behind the company it is growing fast, going from 17 employees to 26 since we spoke in April, with plans to more than double that number by the end of next year. Del Balso says he and his co-founders are committed to building a diverse and inclusive company, but he acknowledges it’s not easy to do.
“It’s actually something that we have a primary recruiting initiative on. It’s very hard, and it takes a lot of effort, it’s not something that you can just make like a second priority and not take it seriously,” he said. To that end, the company has sponsored and attended diversity hiring conferences and has focused its recruiting efforts on finding a diverse set of candidates, he said.
Unlike a lot of startups we’ve spoken to, Del Balso wants to return to an office setup as soon as it is feasible to do so, seeing it as a way to build more personal connections between employees.
Powered by WPeMatico
VSCO acquires mobile app Trash to expand into AI-powered video editing
VSCO, the popular photo and video editing app, today announced it has acquired AI-powered video editing app Trash, as the company pushes further into the video market. The deal will see Trash’s technology integrated into the VSCO app in the months ahead, with the goal of making it easier for users to creatively edit their videos.
Trash, which was co-founded by Hannah Donovan and Genevieve Patterson, cleverly uses artificial intelligence technology to analyze multiple video clips and identify the most interesting shots. It then stitches your clips together automatically to create a final product. In May, Trash added a feature called Styles that let users pick the type of video they wanted to make — like a recap, a narrative, a music video or something more artsy.
After Trash creates its AI-powered edit, users can opt to further tweak the footage using buttons on the screen that let them change the order of the clips, change filters, adjust the speed or swap the background music.
Image Credits: Trash
With the integration of Trash’s technology, VSCO envisions a way to make video editing even more approachable for newcomers, while still giving advanced users tools to dig in and do more edits, if they choose. As VSCO co-founder and CEO Joel Flory explains, it helps users get from that “point zero of staring at their Camera Roll…to actually putting something together as fast as possible.”
“Trash gets you to the starting point, but then you can dive into it and tweak [your video] to really make it your own,” he says.
The first feature to launch from the acquisition will be support for multi-clip video editing, expected in a few months. Over time, VSCO expects to roll out more of Trash’s technologies to its user base. As users make their video edits, they may also be able to save their collection of tweaks as “recipes,” like VSCO currently supports for photos.
“Trash brings to VSCO a deep level of personalization, machine learning and computer vision capabilities for mobile that we believe can power all aspects of creation on VSCO, both now and for future investments in creativity,” says Flory.
The acquisition is the latest in a series of moves VSCO has made to expand its video capabilities.
At the end of 2019, VSCO picked up video technology startup Rylo. A few months later, it had leveraged the investment to debut Montage, a set of tools that allowed users to tell longer video stories using scenes, where they could also stack and layer videos, photos, colors and shapes to create a collage-like final product. The company also made a change to its app earlier this year to allow users to publish their videos to the main VSCO feed, which had previously only supported photos.
More recently, VSCO has added new video effects, like slowing down, speeding up or reversing clips and new video capture modes.
As with its other video features, the new technology integrations from Trash will be subscriber-only features.
Today, VSCO’s subscription plan costs $19.99 per year, and provides users with access to the app’s video editing capabilities. Currently, more than 2 million of VSCO’s 100 million+ registered users are paid subscribers. And, as a result of the cost-cutting measures and layoffs VSCO announced earlier this year, the company has now turned things around to become EBITDA positive in the second half of 2020. The company says it’s on the path to profitability, and additional video features like those from Trash will help.
Image Credits: Trash
VSCO’s newer focus on video isn’t just about supporting VSCO’s business model, however, it’s also about positioning the company for the future. While the app grew popular during the Instagram era, today’s younger users are more often posting videos to TikTok instead. According to Apple, TikTok was the No. 2 most downloaded free app of the year — ahead of Instagram, Facebook and Snapchat.
Though VSCO doesn’t necessarily envision itself as only a TikTok video prep tool, it does have to consider that growing market. Similar to TikTok, VSCO’s user base consists of a younger, Gen Z demographic; 75% of VSCO’s user base is under 25, for example, and 55% of its subscribers are also under 25. Combined, its user base creates more than 8 million photos and videos per day, VSCO says.
As a result of the acquisition, Trash’s standalone app will shut down on December 18.
Donovan will join VSCO as Director of Product and Patterson as Head of Applied Research. Other Trash team members, including Karina Bernacki, Chihyu Chang and Drew Olbrich, will join as Chief of Staff, Engineering Manager and Sr. Software Engineer for iOS, respectively.
“We both believe in the power of creativity to have a healthy and positive impact on people’s lives,” said Donovan, in Trash’s announcement. “Additionally, we have similar audiences of Gen Z casual creators; and are focused on giving people ways to express themselves and share their version of the world while feeling seen, safe, and supported,” she said.
Trash had raised a total of $3.3 million — a combination of venture capital and $500,000 in grants — from BBG, Betaworks, Precursor and Dream Machine, as well as the National Science Foundation. (Multiple TechCrunch connections here: BBG is backed by our owner Verizon Media, while Dream Machine is the fund created by former TechCrunch editor Alexia Bonatsos.)
“Han and Gen and the Trash team have always paid attention to the needs of creators first and foremost. My hope is that the VSCO and Trash partnership will turn all of us into creators, and turn the gigabytes of latent videos on our phones from trash to treasures,” said Bonatsos, in a statement about the deal.
Flory declined to speak to the deal price, but characterized the acquisition as a “win-win for both the Trash team and for VSCO.”
Updated 12/3/20, 11:27 AM ET: VSCO alerted us that Patterson’s title is being updated to “Head of Applied Research.” We’ve updated the article accordingly.
Powered by WPeMatico
WaveOne aims to make video AI-native and turn streaming upside down
Video has worked the same way for a long, long time. And because of its unique qualities, video has been largely immune to the machine learning explosion upending industry after industry. WaveOne hopes to change that by taking the decades-old paradigm of video codecs and making them AI-powered — while somehow avoiding the pitfalls that would-be codec revolutionizers and “AI-powered” startups often fall into.
The startup has until recently limited itself to showing its results in papers and presentations, but with a recently raised $6.5M seed round, they are ready to move towards testing and deploying their actual product. It’s no niche: video compression may seem a bit in the weeds to some, but there’s no doubt it’s become one of the most important processes of the modern internet.
Here’s how it’s worked pretty much since the old days when digital video first became possible. Developers create a standard algorithm for compressing and decompressing video, a codec, which can easily be distributed and run on common computing platforms. This is stuff like MPEG-2, H.264, and that sort of thing. The hard work of compressing a video can be done by content providers and servers, while the comparatively lighter work of decompressing is done on the end user’s machines.
This approach is quite effective, and improvements to codecs (which allow more efficient compression) have led to the possibility of sites like YouTube. If videos were 10 times bigger, YouTube would never have been able to launch when it did. The other major change was beginning to rely on hardware acceleration of said codecs — your computer or GPU might have an actual chip in it with the codec baked in, ready to perform decompression tasks with far greater speed than an ordinary general-purpose CPU in a phone. Just one problem: when you get a new codec, you need new hardware.
But consider this: many new phones ship with a chip designed for running machine learning models, which like codecs can be accelerated, but unlike them the hardware is not bespoke for the model. So why aren’t we using this ML-optimized chip for video? Well, that’s exactly what WaveOne intends to do.
I should say that I initially spoke with WaveOne’s cofounders, CEO Lubomir Bourdev and CTO Oren Rippel, from a position of significant skepticism despite their impressive backgrounds. We’ve seen codec companies come and go, but the tech industry has coalesced around a handful of formats and standards that are revised in a painfully slow fashion. H.265, for instance, was introduced in 2013, but years afterwards its predecessor, H.264, was only beginning to achieve ubiquity. It’s more like the 3G, 4G, 5G system than version 7, version 7.1, etc. So smaller options, even superior ones that are free and open source, tend to get ground beneath the wheels of the industry-spanning standards.
This track record for codecs, plus the fact that startups like to describe practically everything is “AI-powered,” had me expecting something at best misguided, at worst scammy. But I was more than pleasantly surprised: In fact WaveOne is the kind of thing that seems obvious in retrospect and appears to have a first-mover advantage.
The first thing Rippel and Bourdev made clear was that AI actually has a role to play here. While codecs like H.265 aren’t dumb — they’re very advanced in many ways — they aren’t exactly smart, either. They can tell where to put more bits into encoding color or detail in a general sense, but they can’t, for instance, tell where there’s a face in the shot that should be getting extra love, or a sign or trees that can be done in a special way to save time.
But face and scene detection are practically solved problems in computer vision. Why shouldn’t a video codec understand that there is a face, then dedicate a proportionate amount of resources to it? It’s a perfectly good question. The answer is that the codecs aren’t flexible enough. They don’t take that kind of input. Maybe they will in H.266, whenever that comes out, and a couple years later it’ll be supported on high-end devices.
So how would you do it now? Well, by writing a video compression and decompression algorithm that runs on AI accelerators many phones and computers have or will have very soon, and integrating scene and object detection in it from the get-go. Like Krisp.ai understanding what a voice is and isolating it without hyper-complex spectrum analysis, AI can make determinations like that with visual data incredibly fast and pass that on to the actual video compression part.
Image Credits: WaveOne
Variable and intelligent allocation of data means the compression process can be very efficient without sacrificing image quality. WaveOne claims to reduce the size of files by as much as half, with better gains in more complex scenes. When you’re serving videos hundreds of millions of times (or to a million people at once), even fractions of a percent add up, let alone gains of this size. Bandwidth doesn’t cost as much as it used to, but it still isn’t free.
Understanding the image (or being told) also lets the codec see what kind of content it is; a video call should prioritize faces if possible, of course, but a game streamer may want to prioritize small details, while animation requires yet another approach to minimize artifacts in its large single-color regions. This can all be done on the fly with an AI-powered compression scheme.
There are implications beyond consumer tech as well: A self-driving car, sending video between components or to a central server, could save time and improve video quality by focusing on what the autonomous system designates important — vehicles, pedestrians, animals — and not wasting time and bits on a featureless sky, trees in the distance, and so on.
Content-aware encoding and decoding is probably the most versatile and easy to grasp advantage WaveOne claims to offer, but Bourdev also noted that the method is much more resistant to disruption from bandwidth issues. It’s one of the other failings of traditional video codecs that missing a few bits can throw off the whole operation — that’s why you get frozen frames and glitches. But ML-based decoding can easily make a “best guess” based on whatever bits it has, so when your bandwidth is suddenly restricted you don’t freeze, just get a bit less detailed for the duration.
Example of different codecs compressing the same frame.
These benefits sound great, but as before the question is not “can we improve on the status quo?” (obviously we can) but “can we scale those improvements?”
“The road is littered with failed attempts to create cool new codecs,” admitted Bourdev. “Part of the reason for that is hardware acceleration; even if you came up with the best codec in the world, good luck if you don’t have a hardware accelerator that runs it. You don’t just need better algorithms, you need to be able to run them in a scalable way across a large variety of devices, on the edge and in the cloud.”
That’s why the special AI cores on the latest generation of devices is so important. This is hardware acceleration that can be adapted in milliseconds to a new purpose. And WaveOne happens to have been working for years on video-focused machine learning that will run on those cores, doing the work that H.26X accelerators have been doing for years, but faster and with far more flexibility.
Of course, there’s still the question of “standards.” Is it very likely that anyone is going to sign on to a single company’s proprietary video compression methods? Well, someone’s got to do it! After all, standards don’t come etched on stone tablets. And as Bourdev and Rippel explained, they actually are using standards — just not the way we’ve come to think of them.
Before, a “standard” in video meant adhering to a rigidly defined software method so that your app or device could work with standards-compatible video efficiently and correctly. But that’s not the only kind of standard. Instead of being a soup-to-nuts method, WaveOne is an implementation that adheres to standards on the ML and deployment side.
They’re building the platform to be compatible with all the major ML distribution and development publishers like TensorFlow, ONNX, Apple’s CoreML, and others. Meanwhile the models actually developed for encoding and decoding video will run just like any other accelerated software on edge or cloud devices: deploy it on AWS or Azure, run it locally with ARM or Intel compute modules, and so on.
It feels like WaveOne may be onto something that ticks all the boxes of a major b2b event: it invisibly improves things for customers, runs on existing or upcoming hardware without modification, saves costs immediately (potentially, anyhow) but can be invested in to add value.
Perhaps that’s why they managed to attract such a large seed round: $6.5 million, led by Khosla Ventures, with $1M each from Vela Partners and Incubate Fund, plus $650K from Omega Venture Partners and $350K from Blue Ivy.
Right now WaveOne is sort of in a pre-alpha stage, having demonstrated the technology satisfactorily but not built a full-scale product. The seed round, Rippel said, was to de-risk the technology, and while there’s still lots of R&D yet to be done, they’ve proven that the core offering works — building the infrastructure and API layers comes next and amounts to a totally different phase for the company. Even so, he said, they hope to get testing done and line up a few customers before they raise more money.
The future of the video industry may not look a lot like the last couple decades, and that could be a very good thing. No doubt we’ll be hearing more from WaveOne as it migrates from lab to product.
Powered by WPeMatico