FPGA for Machine Learning

graph TD;
A[Classification Algorithms] --> b[Decision Trees]
A --> c[k Nearest Neighbours]
A --> d[Neural Network]

Any ML Program uses one of the above algorithm.
These Algorithms can be run on CPU or a specialized hardware,
- On a CPU, every instruction must be processed sequentially using generalized hardware, which requires a lot of time
- On a logical circuit (hardware), it would be faster, as the processing is replaced by simple logical circuits on specialized hardware, that can run simultaneously.
The special hardware can be pre designed or dynamically built on FPGA which makes it application specific.

Decision Trees

Decision tree is a hierarchical model that can be implemented using switch or if conditions
Decision trees can be easily implemented using multiplexer, a logic circuit, which reduces code, and increases the speed.

For classification this k Nearest Neighbors algorithm is used.
The nearest neighbors are decided based on certain parameters like Euclidian Space, or City block
- Calculating these distance, would require mathematical operations, which can be quickened using an adder / subtractor and other mathematical calculators on FPGA.

Depend on this \(Y = w^t x + b\)
Where \(W'\) is the transpose weight matrix that is multiplied with the input vector \(x\) which can be handled by the FPGA for faster matrix multiplication
Once again this should be added with the bias \(b\), which can be done on FPGA for faster computation.