Consumer - Open Learning Campus
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now Consumer Trends Zero UIs Consumer-grade AI Applications language processing, along with our data, Modern interfaces are able to do more Low-code and no-code offerings from to anticipate what we want or need to do for us with fewer direct actions—yet still Amazon Web Services (AWS), Azure, next, sometimes before we even know captivate our attention. The average and Google Cloud will start to trickle to ask. Alibaba’s highly advanced DA adult now makes more than 20,000 de- down to everyday people, who will create can not only interact with real humans cisions a day—some big, such as whether their own AI applications and deploy but also deftly handle interruptions and or not to invest in the stock market, and them as easily as they can a website. open-ended answers. Similar to Google’s some small, such as whether to glance at We’re seeing a shift from highly techni- Duplex, Tiān Māo can make calls on your a mobile phone when the screen lights cal AI applications used by professional behalf, but it also understands intent. So up. Zero user interfaces—otherwise researchers to more lightweight, us- if you’re trying to schedule an appoint- known as ambient computing systems— er-friendly apps intended for tech-savvy ment and mention that you’re usually promise to prioritize those decisions, consumers. New automated machine commuting in the morning, the system delegate them on our behalf, and even learning platforms make it possible for infers that you won’t be available then. autonomously answer for us, depending nonexperts to build and deploy predictive In 2017, Future Today Institute’s model on the circumstance. Much of this invisi- models. Platforms hope that in the near correctly projected that nearly half of ble decision-making will happen without future, we’ll use various AI applications Americans would own and use a digital direct supervision or input from people. as part of our daily work, just as we do assistant by 2020. (An estimated 62% of What makes ambient design so tantaliz- Microsoft Office and Google Docs today. Americans use digital assistants today.) ing is that it should require us to make Amazon and Google dominate the smart fewer and fewer decisions in the near speaker market, but digital assistants can Ubiquitous Digital Assistants be found in many places. Thousands of future. Think of it as a sort of autocom- plete for intent. Digital assistants (DAs)—like Siri, Alexa, applications and gadgets now track and Replika is a programmable digital twin that you and their Chinese counterpart Tiān Māo respond to DAs. News organizations, en- can deploy for your friends. from Alibaba—use semantic and natural tertainment companies, marketers, credit 27 © 2021 Future Today Institute
00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now Consumer Trends card companies, banks, local authorities, Personal Digital Twins political campaigns, and many others can A number of startups are building cus- harness DAs to both surface and deliver tomizable, trainable platforms capable of critical information. learning from you—and then represent- ing you online via personal digital twins. Deepfakes for Fun In 2021, China’s annual Spring Festival Faceswap is a free and open-source deep- Gala on the country’s state broadcaster fake app powered by TensorFlow, Keras, (CCTV) included performances from and Python. Deep Art Effects offers synthesized celebrities. With an estimat- desktop and mobile apps to turn images ed billion people watching, the AI copies into stylized art. REFACE is a face swap mimicked their human counterparts app that morphs your face onto celebrity without pre-scripted behaviors, speech- bodies and creates GIFs to share on social es, or routines. Meanwhile, Replika is media. Jiggy is a deepfake that makes a programmable digital twin that you anyone dance. For now, they all result can deploy for your friends. Molly, a Y in images and GIFs that look like they’ve Combinator–backed startup, answers been manipulated—but with the technol- questions via text. The near future could ogy becoming so easy to use, how long include digital twins for professionals until we can’t tell real from fake? across a range of fields, including health and education. Alibaba’s voice assistant uses natural language processing. 28 © 2021 Future Today Institute
Research
00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now Research Trends Closed-Source Code cation landscape. In the past four years, Most Active Institutions for AI Research Code is important for reproducibility, Facebook seems to have gained ground. Google accountability, and transparency, and Of the conference papers that mention Stanford University it is a key to driving improvements in the framework the researchers used, 75% Carnegie Mellon University the greater AI community. But when cited PyTorch but not TensorFlow. Of MIT academic researchers publish papers, the 161 researchers who published more Microsoft Research UC Berkley they don’t often include all of their code. TensorFlow papers than PyTorch papers, Columbia University The reason given: The code they used 55% of them switched to PyTorch, while University of Oxford is intermingled with other proprietary only 15% moved in the other direction. Tsinghua University Facebook research, and it therefore can’t be re- Cornell University leased. Fewer than 15% of all academic Cost of Training Models University of Texas at Austin Princeton University papers on AI publish their full code, and It costs a lot to train a model. Several UCLA some big names—DeepMind and Ope- variables influence those costs, all of University of Illinois at Urbana-Champaign nAI—notoriously leave theirs out, citing which have increased in the past few INRIA Georgia Tech proprietary concerns. years. For example, it costs an average Peking University of $1 per 1000 parameters. OpenAI’s IBM University of Toronto Framework Consolidation 175 billion parameter, GPT-3, likely University of Washington Google’s TensorFlow and Facebook’s cost more than $10 million to train. For ETH Zurich PyTorch are two popular frameworks smaller research groups and companies, EPFL New York University used by researchers, and the relative pop- the costs are out of reach. Some in the AI Duke University ularity of different frameworks typically community are instead allowing the big 0 10 20 30 40 50 60 70 80 90 mirrors trends in the commercial appli- tech companies to pre-train and publish big models. USA UK China Europe Canada Papers published at NeurlIPS 2019 (fractional count) Source: https://macropolo.org/digital-projects/the-global-ai-talent-tracker/ 30 © 2021 Future Today Institute
00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now Research Trends NLP Benchmarks resources, and a new public leaderboard. AI Summarizing Itself Graph Neural Networks 2016, it is a new framework that makes The General Language Understanding We predict that by the end of 2021, this A new AI model can summarize scientific Because we perceive scents using mil- it possible for algorithms to use data Evaluation (GLUE) benchmark is a new benchmark will also be surpassed. literature, including research about itself. lions of sensory neurons in our brains, on devices—such as mobile phones and collection of resources for training, eval- The Allen Institute for Artificial Intelli- and because scents are multifaceted, smart watches—without compromising uating, and analyzing natural language Machine Reading Comprehension gence (AI2) used the model in Semantic predicting the way something will smell user privacy. Research in this space has understanding systems. It includes a Scholar, an AI-powered scientific paper is incredibly complex. For example, dramatically increased. For AI researchers, machine reading benchmark of nine sentence- or sen- comprehension (MRC) has been a chal- search engine to provide a short sum- how would you describe the smell of an tence-pair language-understanding tasks lenging goal, but an important one. MRC mary of papers on AI. What makes this orange? Sweet? Bright? Grassy? Each GP Models built on existing datasets and selected makes it possible for systems to read, work impressive—and ultimately so use- descriptor is unique. Classifying smell is Gaussian processes (GP) are the gold to cover a diverse range of dataset sizes, infer meaning, and immediately deliver ful—is that it is capable of compressing tricky because it requires a multi-label standard for many real-world modeling text genres, and degrees of difficulty. It answers while sifting through enor- long papers with accuracy and efficiency. system. Graph neural networks (GNNs) problems, especially in cases where a includes a diagnostic dataset designed to mous datasets. In 2019, China’s Alibaba constitute a particular type of deep model’s success hinges on its ability to evaluate and analyze model performance outperformed humans when tested by No Retraining Required neural network that operates on graphs faithfully represent predictive uncertain- with respect to a wide range of linguistic the Microsoft Machine Reading Com- as inputs. GNNs are being used to detect ty. GPs are becoming more accurate and phenomena found in natural language. Training robots to do more than one smell—to predict odors at a molecular prehension dataset (or MS MARCO for thing is difficult, but a new model pits easier to train, benefiting from neural And it includes a public leaderboard so short), which assessed its ability to use level—and for a wide array of chemical network improvements. that researchers can track their perfor- identical robot arms against one another and biological processes. For example, natural language to answer real questions in a game (moving objects on a virtual mance. The human baseline score is 87, posed by humans. Alibaba’s system deliv- researchers at the Broad Institute used and between May 2018 and August 2020, table in specific ways) in which one robot them to discover antibiotic compounds GPT-3’s Influence ered answers to search queries posted by challenges the other with increasingly natural language processing systems people to Microsoft’s Bing, such as “How that don’t have toxic side effects. The enormous AI that generates hu- increased from 60 to 90.6, surpassing difficult tasks. It’s an example of multi- man-like language, GPT-3, was released many carbs are in an English muffin?” task learning, a deep learning model in humans. The SuperGLUE benchmark and “How do you grow hops?” Federated Learning by OpenAI last year. The text generator is a new measurement of more difficult which machines learn different skills as has written blog posts and code. It was language understanding tasks, improved they progress. OpenAI’s model allows Federated learning is a technique that pitted against college students in an essay a bot to solve new kinds of problems distributes machine learning to the edge. writing contest, and the anonymized without requiring retraining. Introduced by Google researchers in 31 © 2021 Future Today Institute
00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now Research Trends papers were graded by professors. GPT-3 language-only data by contextually deeply culturally embedded racism and earned mostly B’s—the same as human mapping language “tokens,” or the words sexism. A few years ago, if you typed students. But the AI has demonstrated used to train language models, to related “CEO” into Google Images, the first a strong anti-Muslim bias. Researchers images, or “vokens.” For example, au- result of a woman was CEO Barbie. In from Stanford University and McMaster to-generated image captions often can’t an experiment, researchers at Carnegie University probed the neural network infer context. Vokenization would enable Mellon University trained a system to on tasks including prompt completion, machines not just to recognize objects autocomplete images of men and wom- analogical reasoning, and story genera- but to really “see” what’s in them. en cropped below the neck. In pictures tion. They found that a Muslim-violence of men, the system autocompleted him bias appears consistently and creatively Machine Image Completion wearing a suit. The system autocomplet- in many use cases of the model. It’s yet ed women—including U.S. Rep. Alexan- another example of how bias creeps into If a computer system has access to dria Ocasio-Cortez (D-N.Y.)—wearing a our automated systems. Left unchecked, enough images—say, millions and low-cut top or bikini 53% of the time. it will cause problems throughout society millions—it can patch and fill in holes as AI matures. in pictures. This capability has practical applications for professional photog- Predictive Models Using Single raphers, as well as for everyone who Images Vokenization wants to take a better selfie. Soon, if Computer vision systems are getting Models like GPT-3 are trained on syntax the foreground of a mountain is out of smarter. Neural networks can predict and grammar, not creativity or common focus, or if your skin has an unsightly geometry from a single color image. In sense. So researchers at the University of blemish, another version can be swapped 2019, the DeepMind team developed a The SuperGLUE benchmark will be broken by the end of 2021. North Carolina–Chapel Hill are com- in to generate the perfect picture. As generative adversarial network (GAN) bining language models with computer such technology becomes commonplace, that creates videos from images. For vision. Humans learn in a multilayered, there will be significant biases and other example: Imagine a photo of a person multidimensional way, so a new tech- pitfalls to navigate. For example, image holding a basketball. Based on his pos- nique called vokenization extrapolates generation algorithms routinely reflect ture, face, and other data within the 32 © 2021 Future Today Institute
00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now Research Trends picture, the GAN figures out what likely Model-free Approaches to RL a $10 million grant program to catalyze The U.S. Army Research Laboratory Classic AI is the former, because it more happened next and generates a video Dreamer is a reinforcement learning research in this area, although all of the has a system that uses a brain-computer closely represents how we understand clip of the action. Earlier, researchers at (RL) agent that uses a world model to big tech companies are working closely interface armed with computer vision human thought—and the original intent MIT’s Computer Science and Artificial learn long-sighted predictions, employ- to advance RTML too. technology and allows a person to rapidly was to teach machines to think like us. Intelligence Laboratory (CSAIL) trained ing backpropagation through model see and sort images within her line of Researchers are working on new ways to computers to predict what humans predictions. It can create models from Automated Machine Learning sight. CloudSight, a technology company combine both learning and logic using would do next using YouTube videos raw images and learn from thousands (AutoML) specializing in image captioning, is work- neural networks, which would under- and TV shows such as “The Office” and of predicted sequences in parallel using ing on a hybrid crowdsourced computer stand data through symbols rather than “Desperate Housewives.” CSAIL’s system Some organizations want to move away vision system. Microsoft researchers always relying on human programmers a graphics processing unit (GPU). This from traditional machine learning meth- predicts whether two people are likely to new approach solves long-horizon tasks have proposed Pandora, a set of hybrid to sort, tag, and catalog data for them. hug, kiss, shake hands, or slap a high five. ods, which are time-consuming and diffi- human-machine methods and tools for Symbolic algorithms will aid the process, using an imagined world. cult and require data scientists, specialists SinGAN is an unconditional generative understanding system failures. Pandora which should eventually lead to robust scheme that can manipulate and enhance in AI fields, and engineers. Automated leverages both human and system-gen- systems that don’t always require a hu- images—sketch a mountain, and it will Real-time Machine Learning machine learning, or AutoML, is a new erated observations to explain malfunc- man for training. produce a realistic-looking synthetic One big challenge in AI is building approach: a process in which raw data tions related to input content and system photograph. This research will someday machines that can proactively collect and and models are matched together to re- architecture. veal the most relevant information. Goo- General Reinforcement Learning enable robots to more easily navigate interpret data, spot patterns and incor- Algorithms human environments—and to interact porate context, and ultimately learn in gle, Amazon, and Microsoft now offer a host of AutoML products and services. Neuro-Symbolic AI Researchers are developing single with us humans by taking cues from our real time. New research into real-time body language. Retail, manufacturing, machine learning (RTML) shows that The development of AI has been on algorithms that can learn multiple tasks. and education settings could be especially it’s possible to use a continual flow of Hybrid Human-Computer Vision two conceptual tracks since the 1950s: DeepMind, the team behind AlphaGo, relevant. data and adjust models in real time. This symbolic (machines that use a base of which learned how to play Go with the AI isn’t yet capable of fully functioning knowledge and rules that represent skill level of a human grandmaster, has signals a big change in how data moves, without human assistance. Hybrid in- and in how we retrieve information. The concepts) and non-symbolic (machines developed an innovative new algorithm: telligence systems combine humans and that use raw data to create their own pat- AlphaZero. It is capable of achieving National Science Foundation launched AI systems to achieve greater accuracy. terns and representations of concepts). superhuman performance not only in 33 © 2021 Future Today Institute
00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now Research Trends Go but in other games as well, including tonomous and incremental skill building independently on different algorithms wrote in Python. It’s a move that signals chess and shogi (Japanese chess). This and development, and researchers will and datasets, and they only see one likely fragmentation in the future of the one algorithm starts with no knowledge continue to push the limits of what’s another’s work once it is deployed. This AI ecosystem, not unlike the current except for the rules of the game and possible in this field. has been the cause of recent stock market iOS/Android rivalry or the long Mac/PC eventually develops its own strategies glitches and e-commerce website wonki- war. Businesses will find it increasingly to beat other players. In January 2020, Proliferation of Franken- ness. It is especially challenging for big cost-prohibitive and difficult to switch DeepMind published new research Algorithms companies like Facebook, which have between AI frameworks and languages. showing how reinforcement learning billions of algorithms working together techniques could be used to improve Algorithms are simply rules that define at any given time. our understanding of mental health and and automate the treatment of data. They motivation. are built using “if this, then that” logic Using AI, researchers automated the task of convert- that a computer can understand and pro- Proprietary, Homegrown AI ing live actor performances (left) to computer game cess. Here’s an easy example: If a website Languages virtual characters (right). Continuous Learning reader’s IP address is based in Baltimore, Python is a leading language with lots of At the moment, deep learning tech- the rules then allow that reader to freely pre-built libraries and frameworks. Julia, niques are helping systems learn to solve access the site; if the IP address is based a language developed by Massachusetts complex tasks in a way that resembles in Belgium, then the rules first show a Institute of Technology, is an open- what humans can do—but those tasks GDPR screen stating privacy and cookie source language that focuses on numer- are still specific, such as beating a human policies. While a single algorithm might ical computing. And of course there’s at a game. And they require a rigid be easily described and deployed as ex- Lisp, created by modern AI’s foreparent sequence: Gather data, determine the pected, systems of algorithms all working John McCarthy in 1958. Companies are goal, deploy an algorithm. This process together can sometimes pose problems. starting to build and release their own requires humans and can be time-con- Developers don’t always know in ad- software packages now, as well as unique suming, especially during early phases vance how one algorithm will function programming languages for AI applica- when supervised training is required. alongside other algorithms. Sometimes, tions. Uber released its own probabilistic Continuous learning is more about au- several teams of developers are working programming language, Pyro, which it 34 © 2021 Future Today Institute
You can also read