China
Africa
Explore chapters and articles related to this topic
Implementing DBNS Arithmetic
Published in Vassil Dimitrov, Graham Jullien, Roberto Muscedere, Multiple-Base Number System, 2017
Vassil Dimitrov, Graham Jullien, Roberto Muscedere
Components used in the commercial analog computers of the 1950s to early 1970s were based on high-performance (and very expensive) amplifiers. Such accurate components were necessary in order to provide accurate tracking of solutions to the differential equations being evaluated. We are only interested in the final equilibrium state of the network, when all cells are in one of their two saturated states. The accuracy, in time, of how they relaxed to that state is not important. Based on this fact, the block components in Figure 3.9 can be implemented by the equivalent circuit of Figure 3.11. The integrator function is provided by Cx and Rx; this will act as a “leaky integrator” with losses attributed to Rx. ACyC and BCuC are linear voltage-controlled current sources that are connected to all cells in the neighborhood, and are driven by the state variables from, and the inputs to, the neighborhood cells, respectively. Iy is a current source controlled by the state voltage, xij, where the output voltage, yij, is developed across resistor, Ry; saturation occurs at the power supply voltages. I is an independent bias current source.
Modulation and demodulation
Published in Geoff Lewis, Communications Technology Handbook, 2013
ADPCM (adaptive differential pulse code modulation). After demodulation, the digital signal is separated into audio data bits and step-size control bits. The general principle of the technique is shown in Fig. 22.12. The audio bit stream is converted into bi-polar pulses and clocked into a multiplying stage, the step-size bits being used as the multiplying constant. The resulting signal is then converted into analogue format with a leaky integrator. This device is used to allow the output signal to follow the high frequency components of the original audio.
Industrial Forecasting with Exponentially Smoothed Recurrent Neural Networks
Published in Technometrics, 2022
Leaky integrator neurons are continuous dynamical systems which generalize ESNs with a leaking rate parameter which reduces their excitation and forces the reservoir units to change slowly (Jaeger et al. 2007). The leaking rate parameter can be chosen to behave like ES, although the parameter is not optimized during weight fitting. However, ESNs are designed for short-term memory (STM) capacity and have difficulties with varying time series lengths in the input sequence. Additionally, renormalization of states is needed as an additional step to avoid the network becoming oscillatory, and only a subset of the weights can be fitted if the network is to be stable.