I've recently read about a Reinforcement Learning (RL) agent with an LSTM controller overseeing an LSTM path integration module receiving occassional visual input from a CNN (Banino et al., 2018).
Does the functionality gain of combining different NNs eventually flat out? Is model standardization, bringing an air of component-based development (CBD) into NN architectures, for the best? Or are end-to-end implementations with higher integration values to be preferred?
For me it would have to be a reinforcement network which included convolutions, fully connected layers, lstm and a concatenation procedure. It was a complicated network considering it was doing something rather simple
It honestly comes down to how you design the network and how you utilize the different components to extract certain information. Then the way you combine the extracted information, is also critical. In my opinion it's more of how creative you can be and how well you understand the manipulations being applied to the signals which really determine the fucntionality and productivity of your network.
Before it was 'feature engineering', now it seems it's 'architecture engineering'. I kind of feel that's somewhat counter to the spirit of the NN revolution. Intuition hand-crafting creeps back in. Yes, it instills some much needed human creativity -which is not learnable yet, but on the other hand, it sort of undermines the data-driven paradigm it is designed to usher in.
I agree handcrafting components is going to be critical in the future. I think creating hybrid networks with both hand crafted as well as learner components is going to be the next step. Since neural nets are too rigid post training this to truly instill some dynamic component hand crafting would be the most ideal way.
- Badges
- Science topic