can anyone provide a case study on both the Combinations of Genetic Algorithms and Neural Networks with Deep learning methods can used for classification and forecasting?
I am not sure regarding the combination of these techniques. Deep networks give good classification results on their own, However, Take a look at the following links regarding genetic algo+DNN
https://www.lri.fr/~hansen/proceedings/2014/GECCO/.../late.../p1451.pdf
This video is interesting https://www.youtube.com/watch?v=7mNSY86tEFw
I understand that by 'Neural Networks' you mean shallow network (Perceptron). Deep neural networks are already a combination of these. Take a look at this tutorial on deep learning
https://www.toptal.com/machine-learning/an-introduction-to-deep-learning-from-perceptrons-to-deep-networks
Before a recent success in deep learning, Neural Networks and Genetic Algorithms are used complementary cooperation for pre-training and parameter optimization.
As the power of deep learning for classification and forecasting is getting stronger, the genetic algorithms are also used as hyper-parameter optimization as reliable dropout and parameter initialization.
I would argue for the use of ES (Evolutionary Strategy) instead of GA.
Most researchers when they design a robot brain, program all capabilities. I think that is wrong. Human beings learn all their capabilities. This makes them very flexible. I propose a robot brain that also learns everything. For my proposal see www.intelligent-systems.com.ar/intsyst/robotBrain.htm
Do you know of anybody who also designs along these lines?
In the past days, when the vanish gradient was a bigger problem then it is today, many researchers trained NN using Genetic Algorithm, as a "gradient free" method to train these networks, however, GAs are very slow and also requires a lot of memory, and this is why they aren't used anymore today, you can't beat performance of SGD using GAs nowadays, so the research that I see nowadays is more focused on optimizing the graph/architecture of the networks instead of the weights of the network, or even for hyperparameter optimization.
PS: I'm author of Pyevolve, a evolutionary computation framework for Python and I'm also working with Deep Learning.
- Badges
- Science topic