Zero-Space In-Weight and In-Bias Protection for Floating-Point-based CNNs

Deploying convolutional neural networks (CNNs) in image classification systems requires balancing conflicting goals, like throughput, power consumption, and silicon area. In safety-critical environments, ensuring acceptable levels of robustness against faults is also of utmost importance. The robustness gains promoted by quantised CNNs entail a loss of accuracy that may be problematic for some applications. Traditional redundancy-based solutions provide high error coverage at the cost of high, and sometimes unaffordable, overheads, especially for resource-constrained solutions. This paper proposes using error correction codes (ECC) to protect the tensors of CNNs from potential inadvertent corruption. Fault injection is used to locate all bits in tensors that, even if corrupted, do not affect the network inference process. These bits are then replaced by computed parity bits. By exploiting the intrinsic robustness of CNNs, no additional memory bits are required to store the parity bits while preserving both the ECC protection guarantees and the CNN inference accuracy. The proposal applies conventional, conservative, and aggressive policies depending on the required degree of protection and the overhead the system can afford. The usefulness of these alternatives is exemplified through a floating-point-based CNN that is prototyped on a programmable logic device. Unlike existing solutions, the approach can be deployed without retraining, using well-known and proven ECCs and at an in-memory zero-space cost.