It’s brutal out there for tech startups trying to raise money, particularly those looking to raise later-stage growth rounds. But innovative A.I. Startups are still getting funded. Case in point: V7, a software platform that makes it much easier for companies to train computer vision algorithms and integrate them into their processes.
The company has helped companies such as Merck KGaA develop systems that can spot physical defects in pills in its manufacturing plants and GE Healthcare create algorithms to analyze the scans its medical imaging devices produce. “We are trying to capture every part of the life sciences and healthcare data that manifests itself in the form of images,” Alberto Rizzoli, V7’s co-founder and CEO, tells me. But the company has also expanded from healthcare—which still represents about 40% of V7’s business, according to Rizzoli—to all sorts of other industries too, helping to train algorithms to analyze satellite images, as well as those that can detect corroded parts from photographs or spot stock outages in retail stores.
Today, V7 announced it has raised $33 million in a Series A investment round to help the company expand into the U.S. Market. The round was lead by Radical Ventures, a Canadian venture capital firm that invests in deep tech, and Temasek, the sovereign wealth fund of Singapore, with participation from Air Street Capital, Amadeus Capital, and Partech. The firm also has an impressive lineup of A.I. Researchers who are investing as individuals, including Ashish Vaswami, who helped created the Transformer model while at Google Brain and has gone on to co-found Adept AI, Francois Chollet, the Google Brain researcher who created Python-based deep learning API Keras, and DeepMind researcher Oriol Vinyals.
These investors like V7 because its software is a key foundation piece for many companies that want to use A.I. As part of a digital transformation strategy. And despite concerns about inflation and looming recession, most companies are pushing ahead with those plans because they are seen as a strategic necessity. (And to the extent that these A.I.-driven solutions ultimately save labor, through automation, or capital costs, through better asset utilization, they are seen as offering a good return on investment.) “Our thesis for V7 is that the breadth of applications, and the speed at which new products are expected to be launched in the market, call for a centralized platform that connects AI models, code, and humans in a looped ecosystem,” Pierre Socha, a partner at Amadeus Capital Partners, said in a statement.
Part of what V7 provides is data labeling, much like larger, better known competitors such as Scale AI. But, Rizzoli says, the company has deliberately stayed away from the end of the market that just requires lots of relatively unskilled human eyeballs—a phenomenon that has driven labelling companies to seek out cheap labor in developing countries and led to charges of “Silicon Valley sweatshops.” Those kinds of labels are most needed for applications such as moderating social media content, surveillance and security technology, and labeling roadside scenes to help train self-driving cars. V7 has tried to sidestep this ethical quagmire by focusing on computer vision use cases that require highly skilled labelers—radiologists, structural engineers, metallurgists, manufacturing experts, intelligence analysts, and the like. “We don’t want to be associated with a low cost, low value set of tasks that you don’t need any specialized background or education to do,” Rizzoli says.
He also says that V7’s specialty is less about Big Data, and more about pinpointing the exact data a company most needs to improve the performance of its computer vision models. He says that this often requires a big shift in thinking from the way academic A.I. Researchers often think about the performance of computer vision systems. Academics often focus on a metric called mAP, or mean average precision. Most computer vision benchmarks are based around trying to obtain the highest mAP for a task. But Rizzoli says that in many real world commercial uses, what actually matters is not a high mAP at all. The value is in the sub-set of data where the mAP is lowest but where failure has tremendous consequences. “You need to think about what is the worst possible disaster that could happen in a plant where A.I. Could save the day,” he says. Most businesses want a model that can spot these rare but catastrophic failures 100% of the time, even if the model is slightly worse on average.
He says that this same logic helps explain why adoption of neural networks and deep learning in industry is continuing to lag. Many of V7’s customers, he says, are mostly deploying older kinds of machine learning such as support vector machines, decision trees, and good old linear regression. Why? Because deep learning, Rizzoli says, is often not reliable enough for engineering and manufacturing use cases where you need “five nines.” (In other words, 99.999% reliability.) “A lot of large companies, say chemical companies for instance, are happy to pour $1 million into a classifier model if it is 99% accurate”— and that is something he says neural nets generally can’t deliver. Plus, neural networks are still perceived of as “black boxes,” whose failure modes can’t be reliably predicted or understood. “A lot of these use cases require hard-core mechanical engineering-levels of accuracy and, until we get there, A.I. Will be met with skepticism,” he says.
With that, here’s the rest of this week’s A.I. News.
Jeremy Kahn
@jeremyakahn
jeremy.kahn@ampressman
***
Hope to see you all at Brainstorm A.I. Next week!
I hope to see some of you at the best business A.I. Conference on the planet next week. Just a reminder that Coins2Day’s Brainstorm A.I. Conference is taking place in San Francisco on Monday, December 5th, and Tuesday, December 6th. We have an amazing lineup of big thinkers on A.I. And on how A.I. Is impacting business. Attendees will hear from luminaries such as Stanford University’s Fei-Fei Li, Landing AI’s Andrew Ng, Meta’s Joelle Pineau, Google’s James Manyika, Microsoft’s Kevin Scott, Covariant co-founder and robotics expert Pieter Abbeel, Stable Diffusion’s founder Emad Mostaque, and Greylock partner, Paypal and LinkedIN co-founder, and A.I. Investor Reid Hoffman. We will also hear from Intuit CEO Sasan Goodarzi and top executives from Sam’s Club, Land O Lakes, Capital One, and more. And there’s still a chance to join us. You can apply here to register. (And if you use the code EOAI you’ll get a special discount.) See you there!
A.I. IN THE NEWS
German railroad operator is using A.I. To spot price fixing. The German railroad operator Deutsche Bahn has more than 20,000 suppliers and an annual purchasing budget of more than $41 billion. That could make it an easy mark for vendors looking to collude and fix prices on contract bids. In fact, in the past decade, the railroad operator has recovered more than $600 million as part of out-of-court settlements with vendors for anti-competitive practices. Now DB has turned to machine learning algorithms to try to spot suspicious patterns among contract bids that may point to possible collusion. The system has so far found at least 120 instances of such suspicious behavior, according to an article in German newspaper Frankfurter Allgemeine, flagging it for further investigation by humans. And, in at least one of those cases, the German federal anti-cartel agency has been called in.
Google licenses its breast cancer screening model to med tech company. The tech giant has signed a partnership with New Hampshire-based iCAD, which makes cancer diagnostic and radiation therapy equipment, to integrate a computer vision algorithm that Google developed to detect breast cancer into its products. ICAD has agreed to license the algorithm for five years with the hope of bringing a product to market, subject to regulatory approval, by 2024, tech publication The Register reports. The medical device company has also agreed to use Google’s cloud computing infrastructure to store patient data securely. Google’s breast cancer detection algorithm reportedly outperformed a panel of six human radiologists, with lower rates of both false positives and false negatives.
U.S. Department of Justice investigating A.I. Rent-pricing software YieldStar. The DoJ’s antitrust division has opened an investigation into A.I. Software YieldStar, which is produced by a Texas-based technology company called RealPage, according to a story from investigative news site ProPublica that cited an anonymous source familiar with the matter. The reported investigation comes after several Congressional lawmakers urged such action following an October ProPublica story that raised the possibility that increasing use of YieldStar was allowing large landlords to tacitly collude in pushing rents higher.
EYE ON A.I. RESEARCH
NeurIPS, one of the preeminent academic A.I. Conferences, is currently underway in New Orleans. And while I’m not there in person this year, I have been trying to follow some of the developments there from afar.
Nvidia A.I. Researcher Jim Fan is at the conference and wrote a great Twitter thread giving a very short summary of the 15 papers at the conference that won top honors this year. It is worth checking out. Among them is his own work (and those of his colleagues) on creating MineDojo, a way of creating a GPT3-like model that is trained in a Minecraft world where it can take any action, but also learn from action-sequences that are video recorded. The Nvidia team thinks this may point the way towards embodied general purpose agents.
And they aren’t the only ones. This week, OpenAI has also debuted a Minecraft bot that was trained on 70,000 hours of video of people playing the game. According to a story in MIT Technology Review this could point the way towards training more capable general A.I. Agents from watching humans do things in YouTube videos. But the problem was how to label what it is the people are doing. OpenAI did this for Minecraft by first training another agent that, based on a limited set of human labelled data, can then go out and automatically annotate the YouTube videos of people playing Minecraft. This much larger training set is then used to train its Minecraft playing agent.
FORTUNE ON A.I.
You’re a worse driver than a robot: Research shows gaper blocks and looky-loos aren’t an issue with AI—by Travis Loller and the Associated Press
Beauty in the eye of the A.I.: How inherent racial bias has shaped A.I. And what brands are doing to address it—by Gabby Shacknai
Inside Andreessen Horowitz’s grand plans to scale its venture capital firm into a behemoth and conquer the globe— by Eric Newcomer and Jessica Mathews
BRAIN FOOD
Could a Meta achievement at the tabletop game Diplomacy point the way towards bots that can handle negotiations for business?
Researchers at Meta’s A.I. Research lab have pulled off a major breakthrough in creating a A.I. System capable of matching top human opponents in the full tournament version of the strategy game Diplomacy. This is remarkable because much of the game—which is set in Europe in the years leading up to World War I and which sees up to seven players taking on the role of a major European power attempting to dominate the Continent—involves negotiating with other players, in free form natural language, to make and break alliances. It had been thought this negotiation phase of the game was be too complex for an A.I. To master.
But Meta managed to do it, creating an A.I. System it calls Cicero, that essentially yokes two different types of cutting-edge A.I. Together. (The research was published in Science.) One system tries to figure out the most optimal strategy to pursue based on predictions of what all the other players are likely to do. This policy engine was initially trained by looking at thousands of human-played games of Diplomacy, but it was further refined using reinforcement learning—where an agent experiments to try to achieve a goal—to try to find strategies that might have an even higher expected value according to game theory. Then, to help implement that strategy, the system uses a large language model that has been fine-tuned on human dialogues taken from 40,000 Diplomacy games. This allows the model to function like a specialized chatbot, engaging in a negotiation to help it achieve its goal. It also allows it to take information it learns from the other players as part of that negotiation and use it to update the policy model to see whether it needs to adjust its strategy.
Cicero is remarkable in a number of ways. For one thing, it has a kind of “theory of mind”—it has a prediction about the intentions of all the other players and uses that to interpret the information it gets in the negotiation phase. If a player is suggesting an alliance that Cicero thinks is unlikely to be in that player’s interest, it is more likely to think that a player is bluffing, says Noam Brown, one of the Meta AI researchers who worked on the project. It is also remarkable for having mastered the tone and slang that human Diplomacy players use while negotiating, and also for expressing empathy for other players when they have been victims of betrayal or when a strategy hasn’t panned out.
Brown says that Cicero was intended as pure research, without any commercial application in mind. But, he said, it would be easy to see an agent designed like Cicero being used to create much more dynamic and interesting non-player characters for video games—ones whose dialogue and actions would really vary much more widely in response to what a human player does in the game than is the case with most NPCs today. I think though that Brown is aiming too low. It seems to me that the killer use case for Cicero isn’t in video games—it would be in real business.
In any business negotiation—think a customer talking terms with a supplier—a negotiator has to have a sense of its own goals and the best strategy for achieving them, but also has to be able to figure out what the other side is hoping to get out of the transaction. It seems like Cicero would be ideal as a model for a general business negotiation agent. And as Meta has open-sourced Cicero’s model and code, I would be surprised if someone doesn’t try to build one using Cicero and its architecture as a model.
