Learning of hierarchical features with spiking neurons has mostly been investigated in the database framework of standard deep learning systems. However, the properties of neuromorphic systems could be particularly interesting for learning from continuous sensor data in real-world settings. In this work, we introduce a deep spiking convolutional neural network of integrate-and-fire (IF) neurons which performs unsupervised online deep learning with spike-timing dependent plasticity (STDP) from a stream of asynchronous and continuous event-based data. In contrast to previous approaches to unsupervised deep learning with spikes, where layers were trained successively, we introduce a mechanism to train all layers of the network simultaneously. This allows approximate online inference already during the learning process and makes our architecture suitable for online learning and inference. We show that it is possible to train the network without providing implicit information about the database, such as the number of classes and the duration of stimuli presentation. By designing an STDP learning rule which depends only on relative spike timings, we make our network fully event-driven and able to operate without defining an absolute timescale of its dynamics. Our architecture requires only a small number of generic mechanisms and therefore enforces few constraints on a possible neuromorphic hardware implementation. These characteristics make our network one of the few neuromorphic architecture which could directly learn features and perform inference from an event-based vision sensor.

Terahertz (THz) technology has become of large interest over the last 10 years. THz rays are an alternative to X-rays for imaging through thin materials and their non-ionizing character makes them inherently health-safe. The THz domain is also suitable for heterodyne detection and the use of radar techniques to perform 3D imaging. Commercial applications range from non-destructive testing, security screening of objects or persons, and medical imaging to secure communications.Among the multitude of existing THz detectors, silicon field-effect transistors have shown to be suitable for cost-effective video-rate imaging, offering the advantages of room temperature operation, integration of read-out electronics on the same chip, and straightforward array fabrication. The first demonstration of sub-THz and THz detection by CMOS field-effect transistors in silicon was made in 2004 [1] and it was shown shortly later that these devices can reach a noise equivalent power competitive to the best conventional room-temperature THz detectors [2]. The first CMOS focal-plane arrays (FPAs) for imaging at 600 and 645GHz were demonstrated in 2008 and 2009 [3, 4]. Further reduction of the system costs can be achieved by designing a versatile image sensor that operates at a wide range of THz frequencies. This paper presents a prototype THz imager in 0.13μm CMOS with imaging capability from 300GHz to 1THz, lownoise in-pixel amplifiers and multiplexing circuitry for single video output. The 0.13μm CMOS technology enables short enough gate lengths for optimum detection performance up to a few THz, while being significantly cheaper than 90nm or silicon on insulator (SOI) technologies. Furthermore, it is dense enough to be compatible with the THz pixel size. Figure 2.5.1 presents the architecture of the 3×4 pixel imager. Each pixel consists of a differential bow-tie antenna, a single nMOSFET as detecting element, and a single-ended base band amplifier with capacitive feedback. The pixel size is 190×190μm 2 . The pixel outputs are multiplexed to the array output in a standard way via line and column switches controlled by two shift registers. The antenna couples the free-space THz radiation to the detector nMOSFET. The bow-tie type has been chosen in order to achieve a wide detection bandwidth and allow different possible working frequencies from 300GHz to 1THz. The bow-tie shapes are drawn in each of the metal layers M1 to M6 and interconnected by via arrays to reduce conduction losses. In order to decrease substrate losses in the standard resistivity silicon substrate (ρ=10Ωcm), we grind down the dies to a total thickness of 130μm. The detector nMOSFET has a gate length of 130nm and a gate width of 250nm. Even though being far above its cut-off frequency, it effectively rectifies the received THz radiation, leading to a dc detection voltage ΔU between source and drain. The rectification phenomenon is explained by the non-resonant case of the Dyakonov-Shur plasma wave theory [5] or alternatively by distributed resistive self-mixing [4,6]....