A new probabilistic learning machine for pattern recognition is described. The machine operates on the basis of random criteria obtained from special pseudorandom generators. The input, particularly versatile due to the use of an image dissector, allows a random scan for which normal television storage pick-up tubes are not suitable. The computation data section, completely automatic, uses two ferrite core memories with a capacity of 64,000 bits. Consequently it is possible to obtain very quick recognizing cycles for a maximum of 16 classes of 256 learning examples each. Preliminary experiments are reported, some of which, like the recognition of meteorological events by the analysis of absolute baric topography maps, could already be suitable for practical application of the machine.