The AI market is also booming as companies continue to invest in cognitive software capabilities. The International Data Corporation indicates global spending on AI systems is expected to hit $77.6 billion in 2022, more than tripling the $24 billion forecast for 2018.
But the industry still has a long way to go, and much of its progress could depend on whether academics and industry players will succeed in finding a way to empower computer algorithms with human-like learning capabilities. Systems powered by artificial intelligence, whether youre referring to the algorithms Facebook uses to detect inappropriate content or the virtual assistants made by Google or Amazon that power the smart speakers in your home, still cant infer context like humans can. Such an advancement could be critical for Facebook as it ramps up its efforts to detect online bullying and identify content related to terrorism on its platforms.
There are cases that are very obvious, and AI can be used to filter those out or at least flag for moderators to decide, Yann LeCun, chief AI scientist for Facebook AI Research, said in a recent interview with Business Insider. But there are a large number of cases where something is hate speech but theres no easy way to detect this unless you have broader context ... For that, the current AI tech is just not there yet.
A key element in advancing the field of artificial intelligence, particularly when it comes to deep learning, will be ensuring that theres hardware capable of supporting it. Thats the big topic LeCun is addressing at the International Solid-State Circuits Conference on Monday, where hes discussing a new research paper outlining key trends that will be important for chip vendors and researchers to consider over the next five to 10 years. Whatever it is that they build will influence the progress of AI over the next decade, he said.
Ahead of the conference, LeCun spoke with Business Insider about where the field of artificial intelligence is headed, what it could mean for the devices we use in everyday life, the state of AI today, and the biggest challenges that lie ahead. Below are key takeaways from our conversation.
Machines have to get much better at power consumption in order for AI to improve.
Imagine a vacuum thats not only smart enough to map your living room so that it doesnt clean the same spot twice, but is also capable of detecting obstacles before bumping into them. Or a smart lawnmower that can intelligently avoid flowerbeds and branches as it trims your lawn. For gadgets like these to work and become prevalent in addition to technologies that companies like Facebook and Google parent Alphabet are investing in, like augmented reality and self-driving cars LeCun says more power-efficient hardware is needed. Such an advancement isnt just necessary for technologies like these to thrive, but also for improving the way companies like Facebook identify the content of photos and videos in real time. Understanding whats happening in a video, transcribing that activity into text, and then translating that text into another language so that people around the world can understand it in real time requires enormous" amounts of computing power, LeCun says.
Well continue seeing AI advancements in smartphones in the near term before improvements appear elsewhere.
In the next three years, LeCun believes most smartphones will have AI built directly into the hardware through a dedicated processor, which would make features like real-time speech translation more prevalent on phones. This likely isnt a surprise to those who have been paying close attention to the smartphone industry in recent years, as companies such as Apple, Google, and Huawei have been incorporating AI more closely into their mobile devices, which LeCun says will enable all kinds of new applications.
Giving machines common sense will be a big focus for AI research in the next decade.
While humans often learn about the world through general observations, computers are typically trained to perform specific tasks. If you want to design an algorithm that can detect cats in photos, for example, youd have to help it understand what a cat looks like by exposing it to a large trove of data, which could include thousands of photos labeled as including cats. But the Holy Grail in the next decade to push AI forward lies in perfecting a technique known as self-supervised learning, according to LeCun. In other words, enabling machines to generally learn about how the world works through data rather than just learning how to solve one particular problem like identifying cats.
If we actually train [algorithms] to do this, there is going to be significant progress in the ability of machines to capture context and make decisions that are more complex, says LeCun, who added that this technique currently only works reliably for text but not videos and images. Such a breakthrough could be what companies like Facebook need to improve content moderation on their platforms, although theres no telling when that solution will come, LeCun says: This is not something thats going to happen tomorrow."
