What do you think of the pig face recognition competition in JDD contest?
At first glance, it's fun to think about it, but it's a unique task of pig face recognition. I also clicked in and looked at the data set for a while. I have to say that this is a very difficult task for human beings. Cover your face. I won't say much about the surprise, as everyone knows. And that's where I say "it's fun to think about it"-it's a difficult visual recognition task for human beings. In the usual machine learning competition, those tasks are basically easy for human beings. In common visual competitions (such as ImageNet), human performance is considered as a benchmark. The mistakes made by human beings in the usual perspective tasks are considered to be close to the theoretical minimum (Bayesian error). We usually think that human beings are good at perspective tasks. It's not hard to understand. The human brain must evolve to be very good at survival in the long process of natural selection. For example, when hunting, you must know the location, size, speed and so on of the prey. For example, when socializing, you must be able to distinguish who looks like (face recognition). In many cases, visual information is the main source of information. Therefore, the human brain must be good at very visual tasks. However, things are not that simple. Although (because socialization is important) we are very good at face recognition, we are not very good at identifying individuals of other animals. We are not good at recognizing pig face, sheep face, cow face and wolf face. And we also know that pigs are at least good at identifying pigs. This recognition ability is likely to be acquired. If a baby grows up in a wolf's den, he/she may also become better at recognizing the wolf's face because of the small "training data". The feature recognition of wolf face is depicted in the nervous system of "Wolf Child". In the process of obtaining the task of wolf face recognition, how much does human recognition ability depend on hard coding (the innate biological basis and gene-determined part of the human brain) and how much depends on soft coding (data training in the process of brain development and the experience of wolf children)? This is also an interesting question. We already know that although the human brain has evolved to be very good at visual tasks, human beings may not be good at those rare visual tasks if they lack training data from childhood. It is no longer reliable to approximate Bayesian error by human error, and human performance is no longer the benchmark. In this type of visual task represented by pig face recognition (even whether CNN can identify delicious pork), to what extent can machine learning achieve and how much better than human beings? Isn't it good to train directly on the pig map with the deep network of face recognition? If you can train well without changing the network structure, it may indicate that machine learning can easily surpass human performance in many "dead ends of human vision". The mature brain of nervous system may not be so good at learning to deal with the "dead angle problem". However, it is different for the deep web. Face and pig face are the same, both feeding data and extracting features.