Machines Learning - Towards a New Synthetic Autobiographical Memory

Abstract

Abstract. Autobiographical memory is the organisation of episodes and contextual information from an individual's experiences into a coherent narrative, which is key to our sense of self. Formation and recall of autobiographical memories is essential for effective, adaptive behaviour in the world, providing contextual information necessary for planning actions and memory functions such as event reconstruction. A synthetic autobiographical memory system would endow intelligent robotic agents with many essential components of cognition through active compression and storage of historical sensorimotor data in an easily addressable manner. Current approaches neither fulfil these functional requirements, nor build upon recent understanding of predictive coding, deep learning, nor the neurobiology of memory. This position paper highlights desiderata for a modern implementation of synthetic autobiographical memory based on human episodic memory, and proposes that a recently developed model of hippocampal memory could be extended as a generalised model of autobiographical memory. Initial implementation will be targeted at social interaction, where current synthetic autobiographical memory systems have had success.

A Top-Down Approach for a Synthetic Autobiographical Memory System

“…In [2] the authors recognise the following requirements enabling a biologically inspired SAM system to match the functionality of AM:…”

“…In [2], unitary coherent perception was also added to the requirements of the generic SAM system, although it was highlighted that this selection is suboptimal in practice and choices stemming from Bayesian brain hypotheses [3] are an alternative. In practice, the unitary coherent perception is usually used as a means of avoiding the costly computational requirements associated with full Bayesian inference.…”

“…Deterministic inference is also desired to achieve transparency. In [2] this requirement was fulfilled indirectly through a deterministic Gibbs sampler variant within the unitary coherent perception framework. Here, the intractable full Bayesian inference is approximated with a deterministic variational approximation.…”

“…However, there is some commonality (specific scenes are associated with specific sounds). Formally, assume that y (1) , y (2) , · · · , y (M ) represent the segmentation of the observables into M different modalities. These can be accounted for in the latent variable framework by maintaining for all modalities a single Q−dimensional latent space (i.e.…”

“…Subsequently, learning which parts of the whole latent space are relevant for which modality is achieved by optimising a set of relevance weights w (m) ∈ ℜ Q for each modality. If, for example, w (2) 5 , w (3) 5 = 0, this means that the latent dimension 5 encodes information for views 2 and 3, thus avoiding redundancy and achieving compression. This idea was developed in [10].…”