Figure 3 shows potential avenues and directions for this research, followed by a table defining the avenues of research.

Project Terminology

Online Learning

General Ideas

When considering a single learning agent in a virtualised world, research has shown that in social contexts, humans seeing or experiencing other people’s reactions or emotions, i.e observing stimuli and resultant responses in a social environment, can cause the same reactions to recur, for example, shared disgust (Sowden, Khemka and Catmur, 2022).

This phenomenon is called Mirroring, and appears to be a mechanism that humans and animals, and possibly agents, could use to begin learning in specifically unexpected, new and unknown situations.

While individual exploration may be an antidote to being able to generalise from variances in experiences, a step up would be incorporating the variances introduced by others. This is essentially social exploration to bring about social learning.

Social contexts appear to provide more variety than individual experiences, which would otherwise be singular and stable (biased) for each individual, and one might therefore reasonably expect or assume that a group’s response (consisting of more collective experiences) would be better than a personal, single and therefore limited experience, particularly when lacking personal experience in that particular circumstance.

This might be a particularly useful social adaptation to reduce the cognitive overhead of determining appropriate behaviour for an unfamiliar situation for ourselves. It may also be an optimisation strategy for learning about unknown or unfamiliar situations generally, for example, exposing oneself to new (and therefore unfamiliar) topics taught in a classroom, in the same way as an optimisation). Unmet expectations and emotions might be used as an optimisation for detecting a lack of knowledge, and therefore the need for more learning. Research has also shown that we tend to abide by the group, and this might be a reason why this is an attractive strategy to avoid the rigorous process of learning individually.

Simulating forms of social knowledge acquisition (social exploration) could help inform how a naive agent might respond in new social situations, such as a group of agents experiencing a specific situation purely through observation and collection of S-R links and CSIs. In this way, social collaboration likely yields more opportunities to establish knowledge to inform behaviour in a more timely manner, particularly if the situation is unknown.

Indeed, from a behavioural standpoint, if others’ behaviours are remotely similar to how we might begin to react (perhaps determined by a measure of experience codified as confidence), then we - or an agent- are unlikely to need much convincing (or perform more contextual processing) to accept/learn others’ behavioural reactions as our own.

Teaching and mentoring, i.e the transfer of knowledge (and identification of lack of situational knowledge), might thus be possible by allowing naive agents to experience or observe other agents’ responses in simulated social situations, and teaching might be simulated through the transmission of the aforementioned externalised agent knowledge.

In the context of simulation, one might be able to introduce other agents’ reactions through a shared social context (a meeting of agents within the same situation) to simulate the sharing of experience and resulting behaviour. This might then be used as the basis of new agent learning. There is also an opportunity to incorporate multi-player agents.

Generally, it might be possible to simulate social interactions between distributed agents to try and improve their own experience, and being social and being in a social context, may provide important variance in circumstantial experience available in the situational environment and increase the amount of information that is available for learning, in a similar way that variance in training data in ML is an an important way to improve learning.

A reason why contemporary AI approaches such as Artificial Neural Networks (ANNs) often fail is because, “…new context powerfully reveals human cognitive biases in the selection of the training data.”, however integrating training data from disparate sources might alleviate this bias in training data and provide the contextual variance which was not selected (or had not been exposed to) by the individual, and therefore not catered for by their independent learning (Denning and Arquilla, 2022).

Early learning

In collaborating environments in nature, initial intellectual growth (beginning to learn) about unknown circumstances appears to be modelled through specific kinds of learning behaviour, which appears to assist early learning. These behaviours include imitation, role-modelling and imprinting. The first two are obvious, while imprinting refers to a form of rapid learning that occurs just after being born (McCabe, 2013). For example, newly hatched chicks physically copy/replicate everything their mothers do, which might be among the first types of learning that occurs in response to having no experience with a given situation.

Some of these social models, such as imprinting and imitation appears to represent a way to form an initial level of knowledge from a clean slate (tabula rasa), so to speak, as proposed by Locke, and this appears to be safer than the alternative, which would be to randomly stimulates all ones physical senses or abilities to produce arbitrary behavior, i.e to learn from them. This is especially true in nature, as this would likely draw unwanted attention from predators.

That this learning mechanism appears to exist instinctively after birth, appears to validate that some knowledge is, in fact, innate and so while it is unlikely that the idea proposed by Locke, of initial learning starting from a blank slate so to speak (tabula rasa), is entirely accurate (but it might be), it is likely that learning does accumulate. This type of initial learning might be triggered by the identification that a high-priority goal exists when no base knowledge for it exists. This might suggest that a form of initial behaviour replication could be encoded into agents when they start…or are ‘born’ (Bigelow et al., 2018). In this way, one might begin simulating the teaching of an agent to grow its initial intellect.

Social learning would represent a much later, higher-order task in this research that depends on having first established prior dependencies, including a consistent observation protocol, a simulation environment extended to support collaborating agents and shared experiences and an ability to reason based on observation.

References

Sowden, S., Khemka, D. and Catmur, C. (2022) ‘Regulating mirroring of emotions: A social-specific mechanism?’, Quarterly journal of experimental psychology (2006), 75(7), pp. 1302–1313. doi: 10.1177/17470218211049780.

Denning, P. J. and Arquilla, J. (2022) ‘The context problem in artificial intelligence’, Communications of the ACM, 65(12), pp. 18–21. doi: 10.1145/3567605.

McCabe, B. J. (2013) ‘Imprinting’, Wiley interdisciplinary reviews. Cognitive science, 4(4), pp. 375–390. doi: 10.1002/wcs.1231.

Bigelow, A. E. et al. (2018) ‘The Effect of Maternal Mirroring Behavior on Infants’ Early Social Bidding During the Still‐Face Task’, Infancy, 23(3), pp. 367–385. doi: 10.1111/infa.12221.