Life-long Semi-supervised Learning: Continuation of Both Learning and Recognition
Youki Kamiya, Toshiaki Ishii, Osamu Hasegawa · 2007
This paper presents a new method for continuous and incremental learning and recognition based on self-organized incremental neural networks. It is available in the fluctuating environment where the number of recognition classes cannot be defined. In this method, the learning process and recognition process are not separated. This method can acquire concept when multiple feature vectors of new input object come, and then can recognize it using previously acquired concept. We experiment an examples of life-long semi-supervised learning tasks in real world. In the result, the proposed method was able to learn and recognize 104 objects incrementally, non-stop, and in real time