Reliable and Efficient hardware for Trustworthy Deep Neural Networks

The remarkable capacities of modern DNNs facilitate their use in real-time, edge and mobile applications, as well as in safety-critical systems (e.g., autonomous cars). In these contexts, very opposite goals need to be achieved simultaneously: performance, energy consumption and reliability. Hardware accelerators can be an alternative when targeting the first two goals, however, they usually require redundancy techniques to reach the expected DNN resilience level, negatively impacting development costs in terms of area, power and latency. This work explores the use of approximate computing to increase DNN resilience without resorting to redundancy techniques, thereby avoiding additional costs.