In order to grow the CNSE into a dedicated research facility that continues its groundbreaking work beyond the time frame set by the NSF seed funding, a Strategic Plan has been developed.

Initial Marketing Strategy

The primary goal of the initial marketing strategy is to collaborate with mature companies in demonstrating the applicability of neuromorphic technologies to existing commercial applications. This goal is driven by the premise that demonstrated successes will bring further industry investment in the development and application of neuromorphic technologies.

Modern high-tech marketing experts base their strategies on the technology adoption life-cycle model. In this model, new technologies are championed within companies by technology innovators. Thus, the first step of our initial strategy is to find these technology innovators and introduce neuromorphic technologies to them using the methods described below. The next step is to provide expertise that will help these innovators identify promising applications in their company, and hopefully, develop a demonstration prototype.

Large, mature companies tend to be leery of unproven technologies. Despite this, we are well on our way towards demonstrating the commercial relevance of neuromorphic technologies with three of our industrial member companies. In one case, a new product using neuromorphic technology is already on the market.

Rockwell International has used neuromorphic vision chip technology to develop and market TraffiCamTM, a traffic monitoring system that uses a neuromorphic vision chip to provide a lower cost, lower maintenance, and higher capability alternative to inductive loop detection systems. We are also collaborating with Rockwell International on the development of other neuromorphic vision systems, such as a vehicle based vision sensor for tracking lane markers.

We are working with General Motors to develop neural network based diagnostic and control systems for automobile engine applications. The specific goal of the initial collaborative effort is to develop and demonstrate neural network based systems that can detect soft failures of sensors and actuators. During the first phase of this project, a Caltech post-doctoral researcher spent three months at a General Motors facility collecting data from test vehicles with the lead researcher from General Motors. During the second phase of the project, the lead researcher from General Motors spent 1996 at Caltech as an Industrial Research Fellow collaborating with ERC researchers in developing and comparing various neural network based approaches. The effort is focusing on recurrent neural network architectures, that are well suited for modeling dynamical systems such as engines. The third phase of the project involved implementing and testing the approaches at a General Motors facility in a testbed vehicle. It is likely that the resulting diagnostic system will be implemented for one of GM’s 1998 lines of cars.

In collaboration with Honeywell, we are investigating various neural network based approaches for recognition of signatures in time series data. Honeywell is interested in using these approaches for embedding automatic diagnostics into their systems and is providing the application domain knowledge needed to compare the various approaches.

In a specific project that combines efforts in several ERC research areas, we are working with Honeywell to define concepts for an integrated diagnostics system based on analyzing acoustic emissions. Such a system would have broad applications, such as for detecting imminent machine failures. The system would consist of integrated micromachined sensors and neuromorphic VLSI processing circuitry. This project leverages the ongoing ERC efforts to develop technologies for neuromorphic system integration. In particular, it benefits from synergistic collaboration with other efforts to integrate micro-sensors and neuromorphic VLSI, such as the Active Skin project and the Artificial Nose project.

As an example of collaboration with a smaller company, we are working with International Remote Imaging Systems (IRIS), Inc. to develop a neural network based classification system that can sort white blood cells into 14 classes with better than 95% agreement with human experts. A successful system was developed which provides a significant improvement over the currently used approach. The system was incorporated into a product currently (1997) being deployed by the company.

An objective for the next year is to find industrial collaborators to participate in the development of an integrated artificial nose system. Towards this objective, we have met with a major U.S. engine manufacturer and discussed ideas for using an artificial nose system to detect fuel contaminants. We have also made initial contacts with a major water bottling company, located locally, and proposed an artificial nose system for detecting contaminants in water bottles returned to be refilled. Both these applications represent chemical vapor detection problems that require a broadly tuned and inexpensive sensor, such as the envisioned artificial nose. We will continue to pursue these and other possible collaborations.