The dichotomy between machines and living things is narrowing. Today, artificial intelligence (AI) is embedded in all kinds of technology, from robots to social networks. AI has not only had its impact on people from all walks of life, but has also captured the hearts of all age groups.
Kids are introduced to “Internet of Toys” at a very young age, teenagers spend a significant amount of time interacting on social networks, while adults seek comfort by conversing frequently with AI agents like Siri and Alexa to get their music requests fulfilled, shopping wishes granted among other things. We also observe that AI has sneaked into the daily lives of our older generation, as they regularly deploy roomba for cleaning their apartments, and appoint AI personal assistants for managing their health needs.
Although the features of AI that have pervaded our lives help us navigate our complex lives, these applications do not capture the grandeur of artificial intelligence as proposed by AI visionaries of the past. Marvin Minsky, Frank Rosenblatt and Seymour Papert, AI proponents of the 60s, dreamt of a world dominated by “organisms” that would grow, divide and function just like humans, equipped with the added component of super-human intelligence.
The struggle to endow artificial systems with super-human intelligence began in the early 80s, where researchers collaborated with large firms (like IBM) to build computers that could defeat humans in strategy games like Chess or Go. The first instance of AI catching up with human intelligence was in the late 90s when IBM built Deep-Blue, a super computer that defeated Garry Kasparov (world chess champion) in a game of chess. 6 sets of games were played and Deep-blue won 4 of them.
Having defeated world chess champion, Garry Kasparov, AI researchers had every incentive to design smarter AI systems that could play a host of strategic games. However, this endeavor did not culminate in “smarter” AI systems for a very long time. This was partly due to the internet bubble burst and the exponential rise of personal computers in the early 21st century. The technological developments in the early 2000’s gave AI the right opportunity to sneak into the lives of common men, as a smart house assistant.
The next breakthrough in designing AI with super-human intelligence was recorded in 2016 when Alpha-Go, a supercomputer built by DeepMind, defeated Lee Sedol, 18 time world champion, in a game of Go. This was a momentous occasion for AI fans and critics as it convinced everyone that the field of AI still possesses large reserves of untapped potential.
Although AI has made a mark on our society by demonstrating its utility as well as its majestic powers, AI researchers still believe that AI systems have a lot to learn from humans. Yoshua Bengio, a Canadian computer scientist, world renowned for his work on building neural networks and AI systems, mentioned that human babies are far better than present-day AI systems in identifying objects. Object classification and identification are extremely important tasks that need to be performed by systems deployed in self-driving cars. As driver-less cars autonomously move around the city, they need to observe traffic rules, expect reckless fellow drivers and react promptly to jay-walking pedestrians.
His research group developed a simple, yet elegant example to demonstrate how complex AI systems that can beat Go world-champions, cannot identify simple objects in an image, a task that human babies can perform reliably. They trained an AI system to label images that have a car in the foreground as ‘car’. This was done by constructing a large artificial neural network (ANN) and feeding it images that have a car in the foreground. One of the images used for training the AI system is shown below in figure-3a.
The AI system was able to label images that have a car in the foreground as ‘car’ for all images in the training set (set of images that were used to train the AI system), and performed well on a test set (a set of images that the AI system hasn’t encountered before). Following this, the training image was perturbed by altering certain spatial frequencies of the image. This resulted in an image as shown in figure-3b.
Any human would immediately recognize that the perturbed image should be classified as a ‘car’ as the modified image contains a car in the foreground. However, complex AI systems fail miserably at this task! Premature deployment of such AI systems in self-driving cars would cause more chaos than convenience, as the world witnessed in Arizona, where a pedestrian was struck and killed by a self-driving uber in the night.
This brings us to an important set of questions: “Are current day AI-systems really “intelligent” or is their performance just a function of how much they’ve been taught explicitly?”; “Can AI systems ever autonomously grow and learn like humans do?” and “What should AI researchers do differently to build truly intelligent AI systems?”.
Researchers at Caltech believe that in order to develop truly intelligent AI systems, the field needs a total revamp as today’s AI systems are designed and built with the ancient mindset that proponents in the 80s harbored. On speaking to one of the graduate students, Guruprasad Raghavan, we learn that his work focuses on the second question posed above. He endows immature AI systems with a minimal set of rules to grow, connect and build complex architectures that subsequently self-organize their intelligence. He demonstrates that by equipping AI systems with strategies and algorithms adopted by human brains to grow themselves from a single cell, one can autonomously grow functioning ‘brains’ with human intelligence.
His latest work demonstrates that artificial systems can grow their own brains, akin to how a baby’s brain grows from a small set of cells as a fetus to the 8 billion neurons present as an adult. Figure-4 has been borrowed from his paper titled “Neural networks grown and self-organized by noise”, where he shows a single cell seeded on a scaffold, divides, migrates and locally communicates with other cells to form complex architectures.
The functionality of artificial systems growing their own `brains’ and self-organizing their intelligence holds the key to fulfilling the vision laid down by the founding fathers of AI in the early 60’s. We have observed a gradual evolution of AI systems over the past 50 years, as we have moved away from human-designed artificial systems, towards systems that can autonomously grow and function like humans do!
William Gibson’s statement “The future is already here” is no exaggeration for AI as our era is one in which AI systems seamlessly interact with humans on a daily basis providing some people comfort, while others distress. The ability to function at human intelligence levels might mitigate the distress it currently causes, and this would be made possible by allowing artificial systems to have dynamic brains like ours.
- Jo, Jason, and Yoshua Bengio. “Measuring the tendency of CNNs to learn surface statistical regularities.” arXiv preprint arXiv:1711.11561(2017).
- Guruprasad R, M. Thomson, “Neural networks grown and self-organized by noise”, under review