We are interested in the following questions:
- How do we implement traditional processing in a deep learning environment?
- How to maintain >16-bit precision?
- How to constrain the network to find exact SAT like solutions?
- How to map the minimal network to hardware such as ALUs, LUTs and memories?
- How to create a programming environment where the program/code/hw is defined by test vector generators?
- how to optimally reduce an netwrk automatically to fit existing hardware such as FPGAs?
- What are the qualitive differences between NN based CPUs and traditional ALUs
- What can be achieved in terms of error rate if exact solutions exist?
- How to merge and discover logical graph structure into an NN
NNs+SAT as a CAD tool for 1? kbits for FPGA
- Bit level routing
- HiFi computation
- Code compilation
Demonstrations
- NN ALU
- NN memory
- NN CPU with program
- Live video demo on Zynq FPGA