Distinguishing a Cat from a Dog and Other AI Challenges
By Gary Feather
Achin Bhowmik from Intel made a Monday seminar presentation outlining the amazing potential of AI for many applications. These include image recognition, speech, email, fraud, drug toxicology, relationships, moods, reactions, and games.
His focus on image recognition in particular was very insightful. Image identification, such as finding a cat through mathematical sequences and screeners, is not practical when the entire world of images is considered. Although today (and only very recently) image recognition accuracies have surpassed human levels, these are for expected scenes in expected environments. Computed solutions must become more "intuitive.” The objective is a high percentage of detection with a low probability of false identification.
AI is inspired by biology. The process mimics human learning from infants to adults. Through a process of weighing and error and gradients coupled with forward and backward propagation, "learning" can occur at the machine level. The result is an input (image) that can yield an output of accurate identification. So we train the system in a manner similar to the way we learned ourselves.
The objective is to "mimic" the human perceptual system for 3D spatial semantic understanding. The potential of this is staggering. There are concerns when we use systems to make decisions and we do not have a clear deterministic set of rules as to why the decision was made. Realizing that even well-trained humans make large judgement errors, we need to understand that AI with Deep Learning can make similar errors. These might be more critical errors than "thinking" a cat is a dog.