Tag Archives: baby
#438798 This AI Thrashes the Hardest Atari Games ...
Learning from rewards seems like the simplest thing. I make coffee, I sip coffee, I’m happy. My brain registers “brewing coffee” as an action that leads to a reward.
That’s the guiding insight behind deep reinforcement learning, a family of algorithms that famously smashed most of Atari’s gaming catalog and triumphed over humans in strategy games like Go. Here, an AI “agent” explores the game, trying out different actions and registering ones that let it win.
Except it’s not that simple. “Brewing coffee” isn’t one action; it’s a series of actions spanning several minutes, where you’re only rewarded at the very end. By just tasting the final product, how do you learn to fine-tune grind coarseness, water to coffee ratio, brewing temperature, and a gazillion other factors that result in the reward—tasty, perk-me-up coffee?
That’s the problem with “sparse rewards,” which are ironically very abundant in our messy, complex world. We don’t immediately get feedback from our actions—no video-game-style dings or points for just grinding coffee beans—yet somehow we’re able to learn and perform an entire sequence of arm and hand movements while half-asleep.
This week, researchers from UberAI and OpenAI teamed up to bestow this talent on AI.
The trick is to encourage AI agents to “return” to a previous step, one that’s promising for a winning solution. The agent then keeps a record of that state, reloads it, and branches out again to intentionally explore other solutions that may have been left behind on the first go-around. Video gamers are likely familiar with this idea: live, die, reload a saved point, try something else, repeat for a perfect run-through.
The new family of algorithms, appropriately dubbed “Go-Explore,” smashed notoriously difficult Atari games like Montezuma’s Revenge that were previously unsolvable by its AI predecessors, while trouncing human performance along the way.
It’s not just games and digital fun. In a computer simulation of a robotic arm, the team found that installing Go-Explore as its “brain” allowed it to solve a challenging series of actions when given very sparse rewards. Because the overarching idea is so simple, the authors say, it can be adapted and expanded to other real-world problems, such as drug design or language learning.
Growing Pains
How do you reward an algorithm?
Rewards are very hard to craft, the authors say. Take the problem of asking a robot to go to a fridge. A sparse reward will only give the robot “happy points” if it reaches its destination, which is similar to asking a baby, with no concept of space and danger, to crawl through a potential minefield of toys and other obstacles towards a fridge.
“In practice, reinforcement learning works very well, if you have very rich feedback, if you can tell, ‘hey, this move is good, that move is bad, this move is good, that move is bad,’” said study author Joost Huinzinga. However, in situations that offer very little feedback, “rewards can intentionally lead to a dead end. Randomly exploring the space just doesn’t cut it.”
The other extreme is providing denser rewards. In the same robot-to-fridge example, you could frequently reward the bot as it goes along its journey, essentially helping “map out” the exact recipe to success. But that’s troubling as well. Over-holding an AI’s hand could result in an extremely rigid robot that ignores new additions to its path—a pet, for example—leading to dangerous situations. It’s a deceptive AI solution that seems effective in a simple environment, but crashes in the real world.
What we need are AI agents that can tackle both problems, the team said.
Intelligent Exploration
The key is to return to the past.
For AI, motivation usually comes from “exploring new or unusual situations,” said Huizinga. It’s efficient, but comes with significant downsides. For one, the AI agent could prematurely stop going back to promising areas because it thinks it had already found a good solution. For another, it could simply forget a previous decision point because of the mechanics of how it probes the next step in a problem.
For a complex task, the end result is an AI that randomly stumbles around towards a solution while ignoring potentially better ones.
“Detaching from a place that was previously visited after collecting a reward doesn’t work in difficult games, because you might leave out important clues,” Huinzinga explained.
Go-Explore solves these problems with a simple principle: first return, then explore. In essence, the algorithm saves different approaches it previously tried and loads promising save points—once more likely to lead to victory—to explore further.
Digging a bit deeper, the AI stores screen caps from a game. It then analyzes saved points and groups images that look alike as a potential promising “save point” to return to. Rinse and repeat. The AI tries to maximize its final score in the game, and updates its save points when it achieves a new record score. Because Atari doesn’t usually allow people to revisit any random point, the team used an emulator, which is a kind of software that mimics the Atari system but with custom abilities such as saving and reloading at any time.
The trick worked like magic. When pitted against 55 Atari games in the OpenAI gym, now commonly used to benchmark reinforcement learning algorithms, Go-Explore knocked out state-of-the-art AI competitors over 85 percent of the time.
It also crushed games previously unbeatable by AI. Montezuma’s Revenge, for example, requires you to move Pedro, the blocky protagonist, through a labyrinth of underground temples while evading obstacles such as traps and enemies and gathering jewels. One bad jump could derail the path to the next level. It’s a perfect example of sparse rewards: you need a series of good actions to get to the reward—advancing onward.
Go-Explore didn’t just beat all levels of the game, a first for AI. It also scored higher than any previous record for reinforcement learning algorithms at lower levels while toppling the human world record.
Outside a gaming environment, Go-Explore was also able to boost the performance of a simulated robot arm. While it’s easy for humans to follow high-level guidance like “put the cup on this shelf in a cupboard,” robots often need explicit training—from grasping the cup to recognizing a cupboard, moving towards it while avoiding obstacles, and learning motions to not smash the cup when putting it down.
Here, similar to the real world, the digital robot arm was only rewarded when it placed the cup onto the correct shelf, out of four possible shelves. When pitted against another algorithm, Go-Explore quickly figured out the movements needed to place the cup, while its competitor struggled with even reliably picking the cup up.
Combining Forces
By itself, the “first return, then explore” idea behind Go-Explore is already powerful. The team thinks it can do even better.
One idea is to change the mechanics of save points. Rather than reloading saved states through the emulator, it’s possible to train a neural network to do the same, without needing to relaunch a saved state. It’s a potential way to make the AI even smarter, the team said, because it can “learn” to overcome one obstacle once, instead of solving the same problem again and again. The downside? It’s much more computationally intensive.
Another idea is to combine Go-Explore with an alternative form of learning, called “imitation learning.” Here, an AI observes human behavior and mimics it through a series of actions. Combined with Go-Explore, said study author Adrien Ecoffet, this could make more robust robots capable of handling all the complexity and messiness in the real world.
To the team, the implications go far beyond Go-Explore. The concept of “first return, then explore” seems to be especially powerful, suggesting “it may be a fundamental feature of learning in general.” The team said, “Harnessing these insights…may be essential…to create generally intelligent agents.”
Image Credit: Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune Continue reading
#437974 China Wants to Be the World’s AI ...
China’s star has been steadily rising for decades. Besides slashing extreme poverty rates from 88 percent to under 2 percent in just 30 years, the country has become a global powerhouse in manufacturing and technology. Its pace of growth may slow due to an aging population, but China is nonetheless one of the world’s biggest players in multiple cutting-edge tech fields.
One of these fields, and perhaps the most significant, is artificial intelligence. The Chinese government announced a plan in 2017 to become the world leader in AI by 2030, and has since poured billions of dollars into AI projects and research across academia, government, and private industry. The government’s venture capital fund is investing over $30 billion in AI; the northeastern city of Tianjin budgeted $16 billion for advancing AI; and a $2 billion AI research park is being built in Beijing.
On top of these huge investments, the government and private companies in China have access to an unprecedented quantity of data, on everything from citizens’ health to their smartphone use. WeChat, a multi-functional app where people can chat, date, send payments, hail rides, read news, and more, gives the CCP full access to user data upon request; as one BBC journalist put it, WeChat “was ahead of the game on the global stage and it has found its way into all corners of people’s existence. It could deliver to the Communist Party a life map of pretty much everybody in this country, citizens and foreigners alike.” And that’s just one (albeit big) source of data.
Many believe these factors are giving China a serious leg up in AI development, even providing enough of a boost that its progress will surpass that of the US.
But there’s more to AI than data, and there’s more to progress than investing billions of dollars. Analyzing China’s potential to become a world leader in AI—or in any technology that requires consistent innovation—from multiple angles provides a more nuanced picture of its strengths and limitations. In a June 2020 article in Foreign Affairs, Oxford fellows Carl Benedikt Frey and Michael Osborne argued that China’s big advantages may not actually be that advantageous in the long run—and its limitations may be very limiting.
Moving the AI Needle
To get an idea of who’s likely to take the lead in AI, it could help to first consider how the technology will advance beyond its current state.
To put it plainly, AI is somewhat stuck at the moment. Algorithms and neural networks continue to achieve new and impressive feats—like DeepMind’s AlphaFold accurately predicting protein structures or OpenAI’s GPT-3 writing convincing articles based on short prompts—but for the most part these systems’ capabilities are still defined as narrow intelligence: completing a specific task for which the system was painstakingly trained on loads of data.
(It’s worth noting here that some have speculated OpenAI’s GPT-3 may be an exception, the first example of machine intelligence that, while not “general,” has surpassed the definition of “narrow”; the algorithm was trained to write text, but ended up being able to translate between languages, write code, autocomplete images, do math, and perform other language-related tasks it wasn’t specifically trained for. However, all of GPT-3’s capabilities are limited to skills it learned in the language domain, whether spoken, written, or programming language).
Both AlphaFold’s and GPT-3’s success was due largely to the massive datasets they were trained on; no revolutionary new training methods or architectures were involved. If all it was going to take to advance AI was a continuation or scaling-up of this paradigm—more input data yields increased capability—China could well have an advantage.
But one of the biggest hurdles AI needs to clear to advance in leaps and bounds rather than baby steps is precisely this reliance on extensive, task-specific data. Other significant challenges include the technology’s fast approach to the limits of current computing power and its immense energy consumption.
Thus, while China’s trove of data may give it an advantage now, it may not be much of a long-term foothold on the climb to AI dominance. It’s useful for building products that incorporate or rely on today’s AI, but not for pushing the needle on how artificially intelligent systems learn. WeChat data on users’ spending habits, for example, would be valuable in building an AI that helps people save money or suggests items they might want to purchase. It will enable (and already has enabled) highly tailored products that will earn their creators and the companies that use them a lot of money.
But data quantity isn’t what’s going to advance AI. As Frey and Osborne put it, “Data efficiency is the holy grail of further progress in artificial intelligence.”
To that end, research teams in academia and private industry are working on ways to make AI less data-hungry. New training methods like one-shot learning and less-than-one-shot learning have begun to emerge, along with myriad efforts to make AI that learns more like the human brain.
While not insignificant, these advancements still fall into the “baby steps” category. No one knows how AI is going to progress beyond these small steps—and that uncertainty, in Frey and Osborne’s opinion, is a major speed bump on China’s fast-track to AI dominance.
How Innovation Happens
A lot of great inventions have happened by accident, and some of the world’s most successful companies started in garages, dorm rooms, or similarly low-budget, nondescript circumstances (including Google, Facebook, Amazon, and Apple, to name a few). Innovation, the authors point out, often happens “through serendipity and recombination, as inventors and entrepreneurs interact and exchange ideas.”
Frey and Osborne argue that although China has great reserves of talent and a history of building on technologies conceived elsewhere, it doesn’t yet have a glowing track record in terms of innovation. They note that of the 100 most-cited patents from 2003 to present, none came from China. Giants Tencent, Alibaba, and Baidu are all wildly successful in the Chinese market, but they’re rooted in technologies or business models that came out of the US and were tweaked for the Chinese population.
“The most innovative societies have always been those that allowed people to pursue controversial ideas,” Frey and Osborne write. China’s heavy censorship of the internet and surveillance of citizens don’t quite encourage the pursuit of controversial ideas. The country’s social credit system rewards people who follow the rules and punishes those who step out of line. Frey adds that top-down execution of problem-solving is effective when the problem at hand is clearly defined—and the next big leaps in AI are not.
It’s debatable how strongly a culture of social conformism can impact technological innovation, and of course there can be exceptions. But a relevant historical example is the Soviet Union, which, despite heavy investment in science and technology that briefly rivaled the US in fields like nuclear energy and space exploration, ended up lagging far behind primarily due to political and cultural factors.
Similarly, China’s focus on computer science in its education system could give it an edge—but, as Frey told me in an email, “The best students are not necessarily the best researchers. Being a good researcher also requires coming up with new ideas.”
Winner Take All?
Beyond the question of whether China will achieve AI dominance is the issue of how it will use the powerful technology. Several of the ways China has already implemented AI could be considered morally questionable, from facial recognition systems used aggressively against ethnic minorities to smart glasses for policemen that can pull up information about whoever the wearer looks at.
This isn’t to say the US would use AI for purely ethical purposes. The military’s Project Maven, for example, used artificially intelligent algorithms to identify insurgent targets in Iraq and Syria, and American law enforcement agencies are also using (mostly unregulated) facial recognition systems.
It’s conceivable that “dominance” in AI won’t go to one country; each nation could meet milestones in different ways, or meet different milestones. Researchers from both countries, at least in the academic sphere, could (and likely will) continue to collaborate and share their work, as they’ve done on many projects to date.
If one country does take the lead, it will certainly see some major advantages as a result. Brookings Institute fellow Indermit Gill goes so far as to say that whoever leads in AI in 2030 will “rule the world” until 2100. But Gill points out that in addition to considering each country’s strengths, we should consider how willing they are to improve upon their weaknesses.
While China leads in investment and the US in innovation, both nations are grappling with huge economic inequalities that could negatively impact technological uptake. “Attitudes toward the social change that accompanies new technologies matter as much as the technologies, pointing to the need for complementary policies that shape the economy and society,” Gill writes.
Will China’s leadership be willing to relax its grip to foster innovation? Will the US business environment be enough to compete with China’s data, investment, and education advantages? And can both countries find a way to distribute technology’s economic benefits more equitably?
Time will tell, but it seems we’ve got our work cut out for us—and China does too.
Image Credit: Adam Birkett on Unsplash Continue reading