3 ms·
The problem you describe (model architecture will work for which problem) is not lacking rigour, but due to the Turing complete expressive power of neural netwo
by GregarianChild 6y ago
The problem you describe (model architecture will work for which problem) is not lacking rigour, but due to the Turing complete expressive power of neural networks.
What you can do is use the existing rigour and derive stronger properties for restricted classes of NNs (just like you can prove stronger properties for simple subsystems of lambda-calculus) and that is an interesting field of study.
- eli_gottlieb 6y ago> The problem you describe (model architecture will work for which problem) is not lacking rigour, but due to the Turing complete expressive power of neural networks. Non-recurrent neural networks are not Turing-complete in any sense.