The goal of the project is to develop an ultra-fast and compact processor with supercomputer performance and optimized for pattern recognition, data reduction, and fast information extraction for high-quality image processing. The proposed hardware prototype provides sufficient flexibility for applications in a wide range of fields, e.g. triggering in High Energy Physics experiments as well as simulating functions of the human brain. We are also studying new applications in the field of vision: fast extraction of the most relevant features from static images as well as from movies. In general, any artificial intelligence process based on massive pattern recognition can largely profit from our devices. The only necessary condition is that data are suitably prepared and formatted.
For a long time, the primary goal of our research team has been online track reconstruction at hadron colliders experiments. This is a crucial task for the success of an experiment. Since only a minimal fraction of the produced collider events can be written to tape to perform physics analysis, online reconstruction of the detector information is vital to select the most interesting events. The left figure gives an adequate idea of the task complexity: thousands of particles are produced in each bunch crossing, that is called "event" in our language. This level of confusion is due to the extremely high luminosity that is necessary to produce rare particles, as the Higgs boson, at an appreciable rate. These conditions are going to worsen in the future experiments, as the very high luminosity LHC data taking condition expected in the next few years. Applications in this field usually require huge computing resources. Nonetheless, a clever use of the modern, flexible, and powerful electronic resources results in smart solutions that save money, energy (lower consumption), and space (more compact systems).
Since the trend of combining CPUs with ASICs or FPGA-based hardware has been expanding in the real-time, massive, computing realm, we implemented an innovative strategy, based on the optimal mapping of a complex algorithm on different technologies. Our purpose is to get the best results by combining the high performance of rigid dedicated hardware with the distinctive flexibility of general-purpose, but lower-performance, CPUs. The architecture's key role is played by high-level field programmable gate arrays (FPGA), while most computing power is provided by cooperating full-custom ASICs named Associative Memories (AM). The AM chip is suitable for massive parallelism in data correlation searches and it is the most ingenious piece of the entire system. It takes full advantage of the intrinsic parallel nature of the combinatorial problem by comparing at once the image under analysis to a set of precalculated "expectations", or patterns. This approach reduces to linear the exponential complexity of CPU-based algorithms and the problem is solved by the time data are loaded into the chip.
The above figure shows the filtering action performed by the AM on a very simple Higgs event made by only 4 muons, but hidden by 30 superimposed, soft, and not interesting interactions. By filtering only energetic particles the Higgs event is clearly observable in the bottom figure, where the 4 four green muons are clearly visible.

The associative memory can provide its filtering function also to images of different nature. The AM-based processor can simulate the preliminary image processing stages performed by brain, such as the identification of shape edges. At this level, convincing models for the validation of hypotheses about the brain functioning are very similar to the internal AM architecture. Therefore, our hardware processor could implement fast pattern selection/filtering of the type studied in these models of human vision or other brain functions. In particular, strong data reduction of external stimuli before higher level processing is essential in (high rate) background filtering and information extraction from time-changing images for movement identification and its simulations are onerously implementable, if at all, in usual supercomputers.
This approach could be essential in brain neurophysiologic studies. Recent advances in ICT and neuroscience allowed to study and model "in silico" a significant part of the human brain. The brain is certainly the most complex, powerful and fast processing engine and its study is very challenging. Understanding how brain processes information as well as how it communicates with the peripheral nervous system provide new potential applications, new computational system that emulate human skills (e.g. by using the directed fusion of diverse sensors information) or exploit underlying principles for new forms of general purpose computing. Significant improvementes could be achieved in terms of performance, fault tolerance, resilience or energy comsumption over traditional ICT approaches.
Using the associative memory processor for brain studies is particularly fascinating. The most convincing models that try to validate brain functioning hypotheses are extremely similar to the real time architectures developed for High Energy Physics experiments. A multilevel model seems appropriate also to describe the brain organization to perform a synthesis certainly much more impressive than what done in HEP triggers. The AM pattern matching has proven to play a key role in high rate filtering/reduction tasks. We can test the AM device capability as th first level of this process, dedicated to external stimuli pre-processing. We follow the conjecture [1] that brain works by dramatically reducing input information by selecting for higher-level processing and long-term storage only the input data that match a particular set of memorized patterns. The double constraint of finite computing power and finite output bandwidth determines to a large extent what type of information is found to be "meaningful" or "relevant" and becomes part of higher level processing and longer-term memory. The AM-based processor will be used for a real-time hardware implementation of fast pattern selection/filtering of the type studied in these models of human vision and other brain functions.
The bottom row of the above figure shows image quality when filtering by accepting only the "good patterns" stored in the AM. We could apply this filtering function to medical images, to extract the relevant features and apply complex but fast reconstrction algorithms. The extracted features could be measured in a very short time. This would allow to provide a quasi-automatized preliminary computer-aided diagnosis. The reduction of execution time of image reconsturction to be applied after the AM filtering function, would exploit the computing power of parallel arrays of Field Programmable Gate Arrays (FPGAs) as we do in FTK to identify clusters of contiguous pixels above a programmable threshold. As images are processed and measurements relative to their shapes are provided, a more accurate analysis of the interesting objects for medical analysis, for example if there are spots and if they are circular of irregulas, can be performed. With further studies these algorithms can be improved to allow 3D imaging.

[1] Del Viva M.M., Punzi G., Benedetti D, "Information and Perception of Meaningful Patterns", PLoS ONE 8(7): e69154.
doi:10.1371/journal.pone.0069154.