Learning sequential behavior is essential to mimic biological brain and hence, authors in [1], proposed a spatio-temporal network, also known as associated cluster-dependent sequence (ACDC) network inspired by the biological brain. For instance, a swimmer, when learning swimming for the first time, scattered sequential activity patterns are observed in the basal ganglia [2-3], hippocampus [4-6] and the cortex [7-8], which provide the working memory [9-11] for this task. This working memory provides a temporal signal for these dynamical neural patterns of swimming to emerge in multiple precise time sequences, where spatial domain will be swimming and temporal domain will be the speed of motor actions involved in swimming. After learning the pattern of swimming, swimmers can later modify the speed of swimming or completely change the style of swimming, which can be done very quickly and with flexibility. Hence, in ACDC architecture, authors proposed a biologically plausible neural computational model of cortico-basal ganglia-thalamus, which provides computational working memory sufficiently powerful to learn arbitrary motor sequences.

Hence, in this project, we choose ACDC architecture to show the correlation and dependency among different motor actions through temporal signature extraction. For this purpose, we employ eight different motor sequences as input to the ACDC network. It is to be noted that these input motor sequences have precisely timed activation functions. Next, these inputs are fed to the recurrent neural network, which is responsible to learn the temporal sequence of these motor actions. Then from recurrent networks, a projection to basal ganglia is trained, which is later mapped to thalamus. Thalamus acts as an activity center, which determines the pattern of the sequence. As we have eight different motor sequences and each thalamus and basal ganglia pair is able to learn only one sequence, we employ eight pairs of thalamus and basal ganglia neurons. It is to be noted that, we use both hebbian (use for the biologically plausible neuron) and delta learning (use for artificial neuron) for updating the weights for these synaptic connections. After training, we extract the temporal information from these ACDC networks, and analyze temporal characteristics through temporal signature extraction, temporal shifting, temporal rescaling and temporal compositionality. Our results show the correlation and temporal matching among the trained sequence.