Convolutional Neural Networks (CNN) have become a popular solution for computer vision problems. However, due to the high data volumes and intensive computation involved in CNNs, deploying CNNs on low-power hardware systems is still challenging.
The power consumption of CNNs
...
Convolutional Neural Networks (CNN) have become a popular solution for computer vision problems. However, due to the high data volumes and intensive computation involved in CNNs, deploying CNNs on low-power hardware systems is still challenging.
The power consumption of CNNs can be prohibitive in the most common implementation platforms: CPUs and GPUs. Therefore, hardware accelerators that can exploit CNN parallelism and methods to reduce the computation burden or memory requirements are still hot research topics. Quantization is one of these methods. One suitable quantization strategy for low-power deployments is logarithmic quantization.
Logarithmic quantization for Convolutional Neural Networks (CNN): a) fits well typical weights and activation distributions, and b) allows the replacement of the multiplication operation by a shift operation that can be implemented with fewer hardware resources.
In this thesis, a new quantization method named Jumping Log Quantization (JLQ) is proposed. The key idea of JLQ is to extend the quantization range, by adding a coefficient parameter "s" in the power of two exponents ($2^{sx+i}$).
This quantization strategy skips some values from the standard logarithmic quantization. In addition, a small hardware-friendly optimization called weight de-zeroing is proposed in this work. Zero-valued weights that cannot be performed by a single shift operation are all replaced with logarithmic weights to reduce hardware resources with little accuracy loss.
To implement the Multiply-And-Accumulate (MAC) operation (needed to compute convolutions) when the weights are JLQ-ed and de-zeroed, a new Processing Element (PE) have been developed. This new PE uses a modified barrel shifter that can efficiently avoid the skipped values.
Resource utilization, area, and power consumption of the new PE standing alone are reported. Resource utilization and power consumption in a systolic-array-based accelerator are also reported.
The results show that JLQ performs better than other state-of-the-art logarithmic quantization methods when the bit width of the operands becomes very small.