UMBC - Related Work

Figure 1: The associated cluster-dependent sequence (ACDC) network consists of somatosensory cortex-as input layer, primary motor cortex -as RNN layer, Go-not go gate as basal ganglia, and A neuron as thalamus.

As mentioned before, we adopt ACDC architecture for our project, which consists of four major modules as illustrated in Fig. 1. To mimic the biological plausible architecture, these modules are termed as somatosensory cortex (SSC)-consider as input layer, primary motor cortex (PMC)-consider as recurrent neural network (RNN), a basal-ganglia (GN) module with artificial neurons, and thalamus (A) unit with artificial neurons[1-3]. Except for the somatosensory unit, the RNN-GN-A-RNN forms a cortical-basal ganglia-thalamus loop, which acts as working memory and responsible for learning sequential dynamics of motor actions. The input somatosensory layer contains an array of motor sequences which need to be learned by the RNN-GN-A-RNN loop. In this array of input motor sequence only one activates at a particular time. This sequence starts with the activation of an excitatory RNN cluster, which contains only excitatory RNN neurons. It is to be noted that RNN representing the PMC consists of excitatory and inhibitory clusters in order to bring the complementary flex function and winner takes all dynamics into the modules[4-5]. As illustrated in Fig. 1, the i^th input motor sequence activates the i^th cluster of the RNN module. The adaptation of a cluster of excitatory neurons instead of a single unit provides the biological activation plausibility in the RNN module.

Later, this excitatory RNN cluster is mapped to prefrontal cortical–BG models in order to get the corresponding GN neuron. In ACDC architecture, this RNN-GN mapping is learned rather than hardware, which makes it biologically plausible compared to the existing architecture[4]. Moreover, to ensure proper mapping the inhibitory cluster present in RNN module provides inhibition to other adjacent clusters in order to form the attractor states indicating the ordinal position corresponding to that particular motor sequence[1-2]. It is to be noted that as long as the ratio of excitatory to inhibitory inputs is not perturbed by another input, activation in the cluster persists and the RNN continues representing the i^th order of input motor sequence. Additionally, in RNN-GN connection, we adopt biologically plausible hebbian learning and each excitatory RNN cluster maps to its corresponding G neuron of the GN module. These G neurons are responsible for remembering the sequence in which the RNN clusters are activated. Later, these G neurons project to the corresponding A neurons of the thalamus unit[5]. It is to be noted that the corresponding projection of G to A neurons depends on non-biological delta learning also known as reinforcement learning. The A neurons present in the thalamus unit determine the action of the behavioral sequence. These A neurons map to three distinct parts of the networks through excitatory connections simultaneously. First A neuron map to the (i+1)^ih the excitatory RNN clusters for generating the next action sequence. Second, A neuron projects to the inhibitory cluster of the RNN module in order to inhibit all the excitatory clusters in the RNN. Thirdly, A neuron maps the excitatory connections back to their corresponding N neurons of the GN unit, which strongly inhibits their corresponding G neurons, thereby shutting down evidence in favor of the next sequence.

References:

O'Reilly, R.C. and Frank, M.J., 2006. Making working memory work: a computational model of learning in the prefrontal cortex and basal ganglia. Neural computation, 18(2), pp.283-328.
Ratcliff, R. and Frank, M.J., 2012. Reinforcement-based decision making in corticostriatal circuits: mutual constraints by neurocomputational and diffusion models. Neural computation, 24(5), pp.1186-1229.
Doi, T., Fan, Y., Gold, J.I. and Ding, L., 2020. The caudate nucleus contributes causally to decisions that balance reward and uncertain visual information. Elife, 9, p.e56694.
Mink, J.W., 1996. The basal ganglia: focused selection and inhibition of competing motor programs. Progress in neurobiology, 50(4), pp.381-425.
Alexander, G.E. and Crutcher, M.D., 1990. Functional architecture of basal ganglia circuits: neural substrates of parallel processing. Trends in neurosciences, 13(7), pp.266-271.