Novel non-Von-Neumann solutions have raised based on artificial intelligence (AI) such as the neuromorphic spiking processors in either analog or digital domain. This paper proposes to study the deep neural network feasibility using ultra-low-power eNeuron. The trade-offs in terms of deep learning capabilities and energy efficiency are highlighted. A linear fit model is found in the region of high energy efficiency of neuromorphic components. Thus, deep learning and energy efficiency mutually exclusive if those neuromorphic components are used.

Different from classical artificial neural network which processes digital data, the spiking neural network (SNN) processes spike trains. Indeed, its event-driven property helps to capture the rich dynamics the neurons have within the brain, and the sparsity of collected spikes helps reducing computational power. Novel synthesis framework is proposed and an algorithm is detailed to guide designers into deep learning and energy-efficient analog SNN using MNIST. An analog SNN composed of 86 electronic neurons (eNeuron) and 1238 synapses interacting through two hidden layers is illustrated. Three different models of eNeurons implementations are tested, being (Leaky) Integrate-and-Fire (LIF), Morris Lecar (ML) simplified (simp.) and biomimetic (bio.). The proposed SNN, coupling deep learning and ultra-low power, is trained using a common machine learning system (Tensor- Flow) for the MNIST. LIF eNeurons implementations present some limitations and weakness in terms of dynamic range. Both ML eNeurons achieve robust accuracy which is approximately of 0.82.

HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L'archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Low-cost devices with ultra-low power radio capabilities are a major challenge in smart devices, while a permanently-on receiver is required for smart communication. This paper proposes a wake-up radio with a neuromorphic preprocessing system both biased in weak inversion region. The system is able to receive a 2.4 GHz signal, demodulate it, and recognize bit patterns based on the spiking frequency of a neuron. Significant performance is obtained with 1.2 nW of total power consumption, which is at least three orders of magnitude less than the conventional RF envelope detectors. Further, spiking frequency responsiveness over input power suggests that the proposed system can distinguish different signals at 2.4 GHz. The proposed system achieves an energy efficiency of 1.2 pJ/bit with a minimum detectable signal of -27 dBm.