Material Demands for Optical Neural Networks
- PDF / 1,231,393 Bytes
- 6 Pages / 604.8 x 806.4 pts Page_size
- 15 Downloads / 198 Views
rain does, but there exists an exciting potential for mimicking some brain functions. A few primitive but important demonstrations of optical neural networks can be found in the literature.211 The purpose of this article is to convey the enthusiasm for this fledgling optical technology by describing some of the research that has been done or is under way. But more importantly, my purpose in writing this article is to articulate a need: The optical Systems that have so far been demonstrated suggest a great potential for these machines. It is the potential for a kind of cognitive processing that has largely eluded conventional approaches with digital electronic Computers and conventional algorithms. Nevertheless, the optical machines with useful and sophisticated processing power will not be built without considerable advances in the technology of the essential materials. My hope is that more attention can be given to the materials needs of these potentially remarkable machines. In the remainder of the article I introduce some of the general principles of the neural network paradigm in the context of an optical associative memory. Much is left out about neural network modeis in general. My main purpose is to focus on issues particularly relevant to optical implementations. The latter part of the paper specifically concerns the materials requirements. Associative Memory Figure 1 is a very simple neural network model that performs the function of an associative memory. Here we see a collection of elementary processing elements (neurons). The Output of each neuron is connected to the input of
every other neuron through an interconnection strength (indicated by the size of the intersection dot). The neuron produces an Output that is a nonlinear (but elementary) function of the sum of its inputs. In the associative memory, and in all neural networks, the interconnection strengths evolve according to a simple rule that takes into account the current Signals that the network experiences. In other words, the interconnects in some way are a record of the history of the network. They represent stored information. We might somehow störe a picture of a cat, a dog, and a mouse in this memory. Recall is established by presenting to the System partial information, say, a cat's tail. The network is then supposed to recall the entire cat from this small, possibly somewhat inaccurate partial input. How it does this is described for its optical counterpart. Optical Associative Memory Figure 2 is a translation of the associative memory into a nonlinear optical circuit; this circuit has been demonstrated. 4'12 Instead of individual neurons and interconnection strengths, the processing elements and interconnects need to be thought of as continuously distributed — the interconnects are contained in a hologram and the nonlinearity of the neurons is produced by an amplifying medium. This System can memorize, recall, forget, daydream, and become obsessive. Daydream? Become obsessive? These are anthropomorphic terms for things that have more re
Data Loading...
