Analog weight adaptation hardware

Abstract

In the previous chapter, some circuits that can be used in the forward-part of analog on-chip learning neural networks were presented. For on-chip learning neural networks, also on-chip weight adaptation hardware is required. In this chapter, a circuit for weight adaptation in on-chip learning feed-forward neural nets is described. This circuit must be very accurate because, as shown in chapter 6, the demands on the maximum parasitic charge injection during adaptation are hard to satisfy. Other demands on the weight adaptation circuit include accurate weight adaptation and a small chip area. Furthermore, the circuit must be fast enough to insure an acceptable speed of operation. The design of the weight adaptation circuit is presented in this chapter.