This paper offers for the first time a novel method for creating Predictive Associative Memories of Time series (PAMeT), where a full set of time series variables is used to create a machine learning predictive model (global learning), and afterward, only a few temporal variables are used at a shorter time to recall the model on new data (local recall). Inspired by human brain processes, PAMeT-SNN leverages the brain-inspired SNN NeuCube and the concept of spatio-temporal associative memories (STAM). PAMeT-SNN is a 4-dimensional spatio-temporal structure. First, it encodes time series data into spike sequences that reflects on the changes in the data over time and then maps the temporal variables into the SNN using temporal similarity to define the spatial locations of the variables in the SNN. It then learns temporal associations between a full set of time series variables, thereby memorizing these temporal associations as spatio-temporal connections between neurons. These connections are activated when only part of the time series data is used to recall the model on new data. The proposed groundbreaking method is exemplified with PAMeT-SNN for predictive modeling on two distinct case study time series data sets: trade dynamics and commodity prices. In these case studies, the method effectively captures the intricate data dynamics, enabling accurate forecasting of future values using a minimal set of variables. The method has the potential to be applied in diverse domains.

The paper introduces the main principles of spatio-temporal associative memory (STAM) inspired by learning principles in the human brain and implemented on a brain -inspired spiking neural network (SNN) architecture. A STAM is a machine learning model that is trained on a full set of spatio-temporal variables, but can be successfully recalled on only a subset of the variables measured in different time intervals. In addition, a STAM model can be further incrementally trained on a new subset of variables measured at varying times. STAM is based on spatio-temporal learning (STL), where existing spatial or other relevant information from the data is used to structure a SNN model and then temporal information is used to train the model. The model captures evolvable and explainable spatio/specro temporal patterns. A STAM model is validated through the introduced association- and generalization accuracy. Departing from traditional deep neural networks and machine learning methods, where trained models can only be recalled or incrementally trained on the same set of variables using vector-based representation, STAM opens the field of machine intelligence for the development of large-scale global spatio-temporal models that can be recalled and used locally, within the available local spatio-temporal data. Possible applications for STAM include: biological and brain signals; audio-visual data; seismic sensory data; financial and economic data; and other.

The paper explores how deep LSTM and deep spiking neural networks (SNN) can be used to extract meaningful features from spatio-temporal EEG brain data for early, on-line diagnosis. It introduces a new online spike encoding algorithm for Izhikevich neural networks and new methods for learning and diagnostic biomarker discovery for each of the three most popular deep learning neural network models, namely: deep BiLSTM; reservoir SNN; and NeuCube. The methods reveal that hidden neurons in a BiLSTM can capture biological meaning, while a reservoir SNN neuron activities and the spiking activities in a NeuCube model can be used to discover EEG channels that can be utilized as brain biomarkers. The study used EEG data from three datasets related to epileptic-, migraine- and healthy subjects. The problem of discriminating epilepsy from migraine using EEG data is a well known hard problem.  LSTM and a reservoir SNN achieved 90% and 85% classification accuracy correspondingly and indicated that channels F8, T3,T’6 and F7,T6 correspondingly can be explored as potential biomarkers, while a NeuCube model achieved 97% with T6, F7 and also C4 and F8 as potential biomarkers for a better classification. In addition, a 2 class NeuCube model discriminated perfectly well epilepsy from migraine and pointing to two channels C4 and F8 as potential biomarkers for the task. While all models resulted in a good classification and feature selected, the main reason for the superior performance of NeuCube is its brain-inspired architecture, where a brain map is used to structure a 3D SNN model to capture deep spatio-temporal dynamics from EEG data. This research has important implications for the development of more effective algorithms for on-line EEG classification, analysis and early brain state diagnosis, adding new features of explainability and discovery to otherwise effective AI models. The proposed methods and findings can be used for the development of more efficient brain-computer interfaces and for clinical practice.

EEG data collected from several subjects when perfoming complex spatio-temporal tasks