**1. Introduction**

Associative memory learning is a ubiquitous online learning paradigm in animals [1–3]. Unlike the data-driven learning schemes of current Artificial Intelligence (AI), animals have the capability of memorizing the events that occur at the same time or within a certain time interval. The underlying memorization mechanism in the nervous system is the synaptic connection that becomes strengthened under the stimulus of the firing neurons evoked by concurrent events. The strengthened synaptic connection enables the response neurons at the conditional pathway to receive a larger amount of the synaptic transmitter. As a result, the response neuron in the conditional signal pathway will fire, even though it originally did not become active. In other

words, the memorization of the relationship between concurrent events is achieved by signal pathway modification rather than backpropagation. The signal pathway modification is accomplished by synaptic plasticity. An AI system with associative memory potentially provides an alternative way of active self-learning by constantly interacting with environments. The signal pathway modification can be accomplished with a few training processes, leading to less dependence on large-size datasets. For data-driven deep learning, for example, Deep Neural Networks (DNNs), the large datasets prolong training time and increase energy demands. Consequently, the application of deep learning is highly reliant on bulky supercomputers that are not feasible and applicable to scenarios that require Size, Weight, and Power (SWaP) constraints [4, 5]. In addition, massive and labeled data are costly to build or even not practical to collect, such as the Lunar and Martian terrain data [5].

Numerous studies have implemented associative memory with neuromorphic systems [2, 6–13]. However, these studies merely complete a small-scale association with a few neurons in simulation environments. It is far away from the capability of associative memory learning to enable animals to self-learn and explore independently in an unknown environment. In addition, pretraining processes with labeled datasets are still required for these studies [9–13]. In order to resolve these limitations of studies on associative learning, we have designed several experiments of associative memory in real-world scenarios using a mobile robot and neuromorphic chips. Our system of associative memory is validated by reproducing one of the classic associative memory learning in rats: *fear conditioning*. In fear conditioning experiments, the rats learn to associate a particular neutral Conditional Stimulus (CS), for example, tone, with an aversive Unconditional Stimulus (US), such as an electrical foot shock, and show a fear response, freezing or running away. The rats learn fear conditioning after several training sessions and exhibit long-lasting behavioral changes. Several brain regions have been proven to be involved in the learning process, including frontotemporal amygdala, hippocampus, and so on. The process of fear conditioning cannot be reproduced by other state-of-the-art associative memory models [2, 6–13] due to their limited neural network sizes. The simple neural network models cannot process informative signals, such as visual signals. These informative signals are processed with large-scale neural assemblies rather than simply a few neurons in the brains [14–19]. To resolve these limitations, in our design, we use large-scale biological plausible neurons to process the visual signals. Specifically, in our system and experimental designs, the mobile robot with sensors serves as the substitute for the rats in fear conditioning experiments. The neuromorphic chip (Intel Loihi) provides a computational platform for the associative memory learning operation. In our experiment, the brightness of a light emulates the visual stimulus, and the vibration signals from the accelerometer mimic the shock signals to the rats. Thus, the vibration signals are the unpleasant stimulus, and light is the neutral stimulus. The movement of the mobile robot emulates the fear response. The perception of the light and the vibration are separately processed within two different neural assemblies. Two neural assemblies connect to the response neuron, which stimulates the movement of the robot, with two signal pathways. One signal pathway with a weak synaptic connection serves as the conditional signal pathway, while another one with a stronger synaptic connection is the unconditional signal pathway. Thanks to the mobile robot providing a platform directly interacting with the environment, we for the first time, to our best knowledge, implement associative memory as real-time online learning with no pretrained procedure. The contributions of this paper are summarized as follows:

*Implementation of Associative Memory Learning in Mobile Robots Using Neuromorphic… DOI: http://dx.doi.org/10.5772/intechopen.110364*

