SNR: <u>S</u> queezing <u>N</u> umerical <u>R</u> ange Defuses Bit Error Vulnerability Surface in Deep Neural Networks

As deep learning algorithms are widely adopted, an increasing number of them positioned in embedded application domains with strict reliability constraints. The expenditure significant resources to satisfy performance requirements neural network accelerators has thinned out the margins for delivering safety applications, thus precluding adoption conventional fault tolerance methods. potential exploiting inherent resilience characteristics networks remains though unexplored, offering a promising low-cost path towards applications. This work demonstrates possibility such exploitation by juxtaposing reduction vulnerability surface through proper design quantization schemes shaping parameter distributions at each layer guidance offered appropriate training methods, high merely algorithmic modifications. Unequaled error can be injected into safety-critical applications tolerate bit rates up absolutely zero hardware, energy, and costs while improving error-free model accuracy even further.