How will our research affect the future of machine intelligence?
Over the last decade, machine learning and artificial intelligence have succeeded in performing many tasks that were previously unimaginable. Such problems as identifying a cat in an image or recognizing someone’s speech have become commonplace in our lives. However, the way in which these technologies solve problems is fundamentally different from how our brain performs the same tasks. For example, a machine learning system that recognizes a cat can be trained with more than a million labeled training images to reliably identify a cat, while a human child will only learn about a cat by a few examples. Furthermore, this child knows that the cat can purr, the cat can scratch and the cat can jump high, whereas all the machine knows is what the cat looks like. The machine doesn’t understand the cat; the child does.
The limitations of modern artificial intelligence systems are generally agreed upon in the machine learning community. Such systems require huge amounts of time and resources for learning, they are fragile to noise, cannot learn continuously and do not summarize, but rather achieve narrow, specific goals. We are still far from a robot that performs tasks that a child can easily do.
We do not believe that these limitations can be overcome simply by following the same path with a large amount of data and more power. A fresh approach is needed, and we believe that it should include what we have learned from the brain.
We have a reliable road map that applies the principles of neurobiology to direct us towards machine intelligence. In the near future, we will apply our model to the existing CNN and AI models to enable systems that are more like the brain, such as systems that are noise-resistant and capable of continuous learning. In the long term, we will continue to create more parts of our model with the ultimate goal of creating intelligent sensor-based systems that can learn, plan and act.
