What is the compelling question or challenge?
Can we build an artificial general intelligence (AGI) - a “system that outperforms humans at most economically valuable work”[1] - which would unleash rapid progress in science, health, education and art?
What do we know now about this Big Idea and what are the key research questions we need to address?
In the last 6 years, artificial neural network techniques have made rapid progress on previously insurmountable challenges in image recognition, image and sound synthesis, language translation, and robotic control.
Two key algorithmic frameworks that were developed in the 1990s finally began to work well due to greatly increased computing power and the appearance of much larger data sets to train the networks on. Convolutional neural networks (CNNs), based on neuroscience discoveries of how low level visual processing is done in the brains of many animals, allow computers to recognize images at beyond human level, enabling diagnosis of eye diseases and cancers better than expert doctors and autonomous vehicles to recognize other cars, signs and pedestrians. Recurrent neural networks (RNNs) have led to advances in language translation and robotic control. Both techniques allow a computer to autonomously discover high level abstract concepts (like faces or word similarity) from raw pixels or letters.
However, our current state of the art roughly corresponds to the lowest levels of our brains’ instinctive processing of the world. The questions that need to be addressed relate to how to approximate our higher level reasoning abilities to form truly abstract concepts and world models, understand other agents' goals, reason by analogy, transfer knowledge between domains, learn from only a few examples, and imagine plans far into the future. This higher level processing is what separates AGI from current techniques.
These questions can only be addressed by combining knowledge and insight from diverse fields like neuroscience, statistics, computer science, psychology, cognitive science, physics, control theory, robotics, and even game theory. Progress will come by giving researchers the computational resources, long term support and interdisciplinary encouragement to work collaboratively for many years on these key research questions. Currently promising approaches include finding ways to augment neural networks with working memories to store earlier learned skills and notice the larger patterns and getting neural networks to predict what the world will look like in the future to better guide its own actions.
Another major point concerns the safety, interpretability and bias of these artificial neural networks. More research is sorely needed to create confidence in the safety of neural networks, especially when deployed in non-training environments, to better understand how they arrive at decisions (especially in medical, employment and justice applications), and to identify when a network has learned a concept that is different from what was desired due to unrepresentative training data or a faulty algorithm.
Current techniques like CNNs and RNNs already enable dramatically improved algorithms for robotic control, medical diagnosis and more. Even if no further algorithmic progress is made, there are many new areas where established techniques can be applied that would be valuable for society. However, future, improved algorithms and architectures will perform at higher levels of reasoning, and promise to be even more valuable by enabling computers to carry out a much wider set of tasks than they can currently. CNNs and RNNs alone are likely not capable of creating AGI. Much more basic research is required. Unlike the incentives of industry, the National Science Foundation (NSF) is the perfect institution to fund explorations of high risk ideas that could be transformative instead of just incremental improvements of current techniques.
Why does it matter? What scientific discoveries, innovations, and desired societal outcomes might result from investment in this area?
Imagine a world where every child has a personal tutor, where every person has a team of the world’s best doctors to treat them, where scientists have access to both a network of experts and can conduct and analyze experiments extremely rapidly, where the elderly and disabled can live free full lives with personal robotic assistants, where anyone who wishes to create has access to software so intuitive and powerful they can quickly and effortlessly make their visions into reality, and where our infrastructure is robust, sustainable, efficient and safe from cyber attack. Artificial general intelligence (AGI) could make such a world possible.
AGI would be one of the most profound developments in history and could lead to an era of rapid progress akin to the exponential growth in living standards caused by the Industrial Revolution. The promise of AGI is that it could enable a level of automation that would break through many of the bottlenecks that currently hold back development in science, technology, and art. Moreover, it could simultaneously enable the distribution of these cutting edge technologies to all of society.
In the healthcare industry, AGI could provide automated diagnosis and personalized treatment as well as enabling robotic caregivers and surgeons and automated drug discovery.
In education, AGI could enable a personal virtual tutor for every student, and empower educators to easily create lessons in many media making it simple to create movies, games and immersive experiences to showcase history, math or any other subject.
In science, AGI could help scientists find new patterns in giant datasets, surface discoveries and insights scientists might not have been aware of, and design (and even robotically build) new equipment. They could also autonomously analyze the results of their own experiments and propose new tests in a virtuous loop.
Similarly, all fields of engineering would have their pace of innovation substantially increased as many of the designing, modelling and testing tasks could be automated. Many of the most complex aspects of our world that are beyond human control or the understanding of individuals may be amenable to control by an AGI. These include control of much of our electrical, logistic and manufacturing infrastructure, defense from cyber attacks, and perhaps even ways of predicting and positively influencing the climate.
Finally, AGI would enable the dream of robotics and smart software that can understand language and expertly act in our real, complex world leading to personal assistants for the elderly and disabled, safe autonomous transport, autonomous manufacturing and farming, and autonomous monitoring and repair of the natural and urban environment. AGI’s transformative potential comes from its generality. It can help people make progress in all fields of human endeavor and could finally make it possible to provide everyone on Earth access to health, education, and a clean environment.
If we invest in this area, what would success look like?
Successfully building an AGI would mean creating a “highly autonomous system that outperforms humans at most economically valuable work.”[1] Building such a system would necessarily require accurate abstract concepts grounded in experience with the world, full understanding of language, and the generation of novel ideas and inventions. The answer above describes some of the amazing impacts such an AGI might have on the world.
An accurate world model for concepts would mean that instead of only seeing many pictures of cats to learn the concept of a cat, the cat concept would combine videos of cats, the touch and feel of a cat, an idea of cat behavior, and the word ‘cat’. Such multisensory concepts are probably essential for an AGI to fully understanding language or have a full understanding of the world.
Such concepts should enable one of the most striking examples of AGI success, which is fully capable robots. AGI will allow robots to function seamlessly in the unstructured real world, working alongside people, responding to verbal instructions, asking for clarification or demonstrations if confused. It will allow robots to enter fields that are too complex at the moment like construction and home care.
Along with robotics, AGI’s most valuable contribution will be the ability to learn abstract human knowledge (including from reading books and scientific papers) and then come up with its own unique ideas and designs for inventions. A useful new invention that a program produced with minimal human guidance is another stark indicator of successfully building AGI.
More generally, successfully building an AGI would mean that the agent could perform work in science, engineering, law, art and most other fields of endeavor.
Why is this the right time to invest in this area?
Recent rapid progress with artificial neural networks (ANNs) suggests a investment now could yield substantial results in the near future. For example, application of ANNs in neuroscience are nearly capable of recreating the functions of various brain regions. When learning to navigate spontaneously, ANNs have firing activity that resembles the grid cells found in rat brains in vivo [2]. Conversely, research on monkey facial recognition indicates that it operates using similar mathematical concepts (e.g. linear combination) as an ANN posed with the same task [3].
Investing in the development of AGI represents an asymmetric bet with a huge upside and no downside. A minimum outcome of this investment would be equipping the next generation of researchers with the building blocks of AGI, which are broadly applicable to the advancement of diverse scientific fields in and of themselves.
References
Reference #1
Reference #2
Reference #3
Show MoreShow Less