A neural network is a powerful tools for pattern association nowadays used in deep learning. Researchers were claiming that neural network can learn any function and deep learning learns deep concept/relationship. In my experiments I found that the CNN is giving good classification accuracy (i.e. good association) but irrelevant concept (i.e. giving importance to those things which are irrelevant). Thus first statement is true but second statements hold only when for the first statement the association is the concept.
I totally agree. ConvNets and the associated algorithms are good at pattern recognition. Very good. They have been invented in a way like the first airplanes simulated the birds fly. However, we are still a bit away from jetliners. Thus overcoming pure simulation of pattern recognition (like it is supposed to happen in the brain) and replacing it by understanding of concepts via neuronal nets, truely understanding of concepts is still out of reach.
Although algorithms are developing fast, there is a need for taking into account time as a crucial fact for "recognition". That means, as long as we take neuronal nets as statical constructs, we will not be able to get the results of human perception (like a turn-based strategy game in comparison to an ego-shooter).The human mind is built upon the fact that we advance through time and update every time unit that can be called "consciousness". Nets will achieve great results only in very specific areas of cognition until we manage to teach them how to integrate the dimension of being - time.
Deep networks learn the complex and different levels of abstraction from an input image. this results in a powerful extraction of low to high features.
I totally agree. ConvNets and the associated algorithms are good at pattern recognition. Very good. They have been invented in a way like the first airplanes simulated the birds fly. However, we are still a bit away from jetliners. Thus overcoming pure simulation of pattern recognition (like it is supposed to happen in the brain) and replacing it by understanding of concepts via neuronal nets, truely understanding of concepts is still out of reach.
Although algorithms are developing fast, there is a need for taking into account time as a crucial fact for "recognition". That means, as long as we take neuronal nets as statical constructs, we will not be able to get the results of human perception (like a turn-based strategy game in comparison to an ego-shooter).The human mind is built upon the fact that we advance through time and update every time unit that can be called "consciousness". Nets will achieve great results only in very specific areas of cognition until we manage to teach them how to integrate the dimension of being - time.
Deep concepts are understood only by human beings in their "mind frame". While deep neural nets just "associates" inputs with output(s)through some kind of a " formalism" . May be in the form of a stabilized weight matrices between layers.
- Badges
- Science topic
- Similar topics
- Psychology
- Mental Processes
- Learning