T. Achler
Optimizing Mind,
United States
Keywords: flexible AI, rehearsal, learning on the fly, flexible robotics
Summary:
AI remains less capable compared to humans at quickly accumulating knowledge without forgetting what they previously learned. Everything that might happen in the robot’s work environment must be in the training set and rehearsed in cumbersome ways. AI based robots are severely limited where and when they can be applied: whether humans can be present in the robotic environment, a very limited amount of autonomy and any changes in function and retraining are costly. Moreover even with severe limits, many hazards remain which safety organizations such as OSHA must carefully monitor. Thus the core bottleneck is not more accurate learning, it is flexibility to learn directly, like a human, unexpected occurrences as they occur within an environment and to understand how the robot is making decisions. Being able to learn without cumbersome rehearsal allows corrections for details that are difficult to capture in a large training set but are always present in the real world environment. The aim of our AI is a solution to be used by computer vision customers to solve their problems right on the spot (without sending data back to retrain the whole network) reducing machine and customer downtime/disruption and increasing productivity. The way this problem is addressed is based on our founder’s Computational Neuroscience research and based on Regulatory Feedback Networks by incorporating predominant top down feedback connections found in the brain as a primary form of computation. Instead of learning distributions from the training set, our approach estimates the distributions during recognition when the object(s) to be recognized are present. This is an iterative optimization process that does not learn weights but finds activations while estimating distributions during recognition when all of the information is available. This step during recognition allows faster, more flexible, and easier to comprehend learning without distributions when the robot needs to learn. Thus learning is more flexible. We offer a solution for robotics where current flexibility limitations do not have viable work-a-round and have an effect on many industries. The proposed AI solution is suited for learning in dynamic changing environments without rehearsal while maintaining scalability even on limited devices as information is encountered. This allows robots to be trained in their environment. Such an advancement would help robots do things better, safer, and with greater efficiency. Success in turn could facilitate disaster response, augment current physical abilities, serve in more environments, and enable exploration beyond the boundaries of Earth. Overcoming this limitation is a key to propel towards the ability to interact with machines in a natural way as is depicted in science fiction. This work has been achieved with the help of NSF support.