XNOR-Net is an optimized binary neural network architecture — that approximates full-precision convolutions using XNOR (exclusive-NOR) operations and popcount, achieving ~58x computational speedup with a carefully designed scaling factor to reduce accuracy loss.

What Is XNOR-Net?

• Innovation: Introduces a real-valued scaling factor $alpha$ per filter. $Conv approx alpha cdot XNOR(sign(W), sign(X))$.
• Reason: Pure binary ($pm 1$) loses magnitude information. The scaling factor $alpha$ (computed analytically from the filter) restores some of this information.
• Result: Significantly better accuracy than naive BNNs, closer to full-precision.

Why It Matters

• Practical BNNs: Made binary networks accurate enough to be taken seriously for real deployment.
• Speed: XNOR + popcount is natively supported on all modern CPUs (SSE, AVX instructions).
• Memory: 32x compression of both weights AND activations.

XNOR-Net is logic-gate deep learning — reducing the multiply-accumulate heart of neural networks to simple bitwise boolean operations.