A Cognitive Behavioral AI: Novel Conversational Memory Elements for Technical Understanding of Medical Deep Denoisers

Swati Rai, Jignesh S. Bhatt, Sarat Kumar Patra, Tanmay Ambadkar

Abstract: Conversation is a powerful cognitive behavioral mechanism for revealing what is understood. In this paper, we propose conversational memory elements (CMEs) to (technically) comprehend the learning functional by state-of-the-art medical deep denoisers. The gradients of neurons' weights, varying over the epochs, are utilized to build the novel CMEs. These are defined in terms of Jacobian matrices and integrated in between layers of the medical deep denoisers without affecting their neural architectures. A covariance map is then defined using values of CME at the convergence of training a denoiser; while an autocorrelation function is derived considering values of CMEs over the epochs. In turn, these statistical indicators are used to infer gross characterizations of the unknown denoising functional in terms of joint probability function expressed by the hidden weights. The ability of learning is quantified via patterns in the covariance maps, while the autocorrelation function help understanding the stability and convergence of deep denoisers. It is further cross-validated against denoising performance analysis on benchmark synthetic and real medical datasets and found consistent. Besides, the proposed approach assists suggesting optimal convergence as well as a deployment strategy among available medical deep denoisers.

External IDs:dblp:conf/cogmi/RaiBPA23