Conceptor ellipsoids store and recall many attractors in one reservoir

- A conceptor C = R(R+alpha^-2 I)^-1 is a soft-projection ellipsoid derived from the state correlation matrix R of a trained N=100 echo-state reservoir. [jaeger-2014-conceptors] - Computing one conceptor per learned temporal pattern and inserting it into the recurrent update (x(n+1) = C tanh(W x(n))) clips the reservoir into that ellipsoid, so a single network stores and selectively regenerates multiple attractor patterns (four patterns reconstructed at NMSE 3.3e-5 to 4.0e-3). [jaeger-2014-conceptors] - Conceptors compose via Boolean AND/OR/NOT on the ellipsoids (OR = (R1+R2)(R1+R2+I)^-1), obeying De Morgan laws, to morph and combine stored patterns. [jaeger-2014-conceptors] - Selecting a different conceptor causally selects a different attractor pattern in the same fixed network; this is the origin paper for the conceptor operator. [jaeger-2014-conceptors]