Brain memory working. Optimal control behavior for improved Hopfield-like models
Several authors have recently highlighted the need for a new dynamical paradigm in the modelling of brain working and evolution. In particular, the models should include the possibility of non constant and non symmetric synaptic weights $T_{ij}$ in the neuron-neuron interaction matrix, radically overcoming the classical Hopfield setting. Krotov and Hopfield have proposed a non constant, still symmetric model, leading to a vector field describing a gradient type dynamics then including a Lyapunov-like energy function. In this note, we first will detail the general condition to produce a Hopfield like vector field of gradient type obtaining, as a particular case, the Krotov-Hopfield condition. Secondly, we abandon the symmetry because of two relevant physiological facts: (1) the actual neural connections have a marked directional character and (2) the gradient structure deriving from the symmetry forces the dynamics always towards stationary points, prescribing every pattern to be recognized. We propose a novel model including a set of limited but varying controls $|\xi_{ij}|\leq K$ used for correcting a starting constant interaction matrix, $T_{ij}=A_{ij}+\xi_{ij}$. Besides, we introduce a reasonable controlled variational functional to be optimized. This allows us to reproduce the following three possible outcomes when submitting a pattern to the learning system. If (1) the dynamics leads to an already existing stationary point without activating the controls, the system has \emph{recognized} an existing pattern. If (2) a new stationary point is reached by the activation of controls, then the system has \emph{learned} a new pattern. If (3) the dynamics is \emph{wandering} without reaching neither existing or new stationary points, then the system is unable to recognize or learn the pattern submitted. A further feature (4), appears to model \emph{forgetting and restoring} memory.
PDF Abstract