FQTree: Fine-grained Quantization and Hardware Generation of Boosted Decision Trees
FQTree reports smaller FPGA boosted-tree hardware without giving up accuracy.
The paper pairs fine-grained quantization-aware training for BDTs with QXGB, an automatic hardware-generation flow. Its leaf-value scheme uses a global quantization step plus tree-wise shifts, with clipping, pruning, and bias folding to cut datapath cost. The authors say applying quantization during boosting lets later trees adapt to errors from the already-quantized ensemble. On JSC, MNIST, and NID, they report 26-57% lower LUT use than state-of-the-art FPGA BDT designs while matching or improving accuracy. ArXiv · AI/CL/LG's note
The paper pairs fine-grained quantization-aware training for BDTs with QXGB, an automatic hardware-generation flow. Its leaf-value scheme uses a global quantization step plus tree-wise shifts, with clipping, pruning, and bias folding to cut datapath cost. The authors say applying quantization during boosting lets later trees adapt to errors from the already-quantized ensemble. On JSC, MNIST, and NID, they report 26-57% lower LUT use than state-of-the-art FPGA BDT designs while matching or improving accuracy. ArXiv · AI/CL/LG's note
score 4