Learning and memory are inextricable

Abstract

The authors' aim is to build “more biologically plausible learning algorithms” that work in naturalistic environments. Given that, first, human learning and memory are inextricable, and, second, that much human learning is unconscious, can the authors' first research question of how people improve their learning abilities over time be answered without addressing these two issues? I argue that it cannot.

Learning is the process of acquiring a memory; learning and memory both depend, fundamentally, on association (Fuster, 1999). The inextricability of learning and memory originates because any to-be-learned information is encoded in memory through association, the same associative encoding also enables retrieval from memory (Brown & Craik, 2000; Tulving & Thomson, 1973). Once retrieved, the encoded memory will drive expectations in the same (or similar) environments to those in which the learning took place.

The authors identify their meta-learned models as ones that “acquire their inductive biases from experience, i.e. by repeatedly interacting with an environment.” This implies that learning is cumulative over time as the model repeatedly samples the environment. For human learning to be cumulative across time, the learned information must be encoded and retained in memory. Later in the paper, the authors, briefly, acknowledge the contribution of episodic memory and the hippocampus but do not spell out how hippocampal memory systems impact their models.

In relation to the authors' first research question, people can improve their learning abilities through elaborate encoding which creates associations between the to-be-learnt episodic material and information already encoded in episodic memory networks (Foer, 2011; Llewellyn, 2013; Yates, 1966). The hippocampus is crucial to both associative encoding and later retrieval of episodes through association (Davachi & Wagner, 2002; Llewellyn, 2013; Thakral, Benoit, & Schacter, 2017). The sequential, associative nature of elements of learned, encoded and retained episodic memories gives rise to rational, step-by-step learning. However, during rapid eye movement (REM) sleep, the hippocampus may take elements of different episodic experiences to identify more elaborative, associative, probabilistic and meaningful patterns in experience which may be expressed, visually, as dreams and instantiated as cortical nodes/junctions during non-rapid eye movement (NREM) sleep (Llewellyn & Hobson, 2015). Cortical episodic memory networks then consist of sequential, associative pathways which converge at nodes/junctions which express patterned, probabilistic, experiential learning. The hidden units/nodes in neural networks may be analogous to these nodes/junctions in cortical episodic memory networks.

Given the authors' recognition, “that episodic memory could be the key to explaining human performance in naturalistic environments” it seems pertinent to expand their algorithms to encompass not only the logical, experiential, “step-by-step” learning patterns that predominate during wake but also the more complex, hyper-associative, experiential patterns identified during REM sleep. Admittedly, this extension may be, somewhat, futuristic. But research has already mimicked slow wave sleep which restored stability to neural networks engaged in unsupervised learning (Watkins, Kim, Sornborger, & Kenyon, 2020). Also, given the adaptivity, speed via massive parallelism, fault tolerance and optimality of neural networks (Bezdek, 1992) and that machine learning programmes trained on probabilistic reasoning are superior to the human brain for visual pattern recognition (Pavlus, 2016) it may be possible that, in the recurrent neural networks mobilized by the authors, REM sleep is not required for complex pattern recognition.

Experiences in natural environments may never be re-enacted in their entirety but they are certainly non-random and do not exclude expectations. Across evolutionary time, many interactions with the environment held dangers because of predators and competitors. To try to avoid dangers, humans needed to identify the probabilistic, behavioural patterns of predators and competitors. Such patterns would only be revealed over several episodes, as humans observed the associations that drove predator behaviour. For example, lions tend to visit waterholes at night when prey are abundant (lion presence is associated with nighttime) but, in the dry season, lions get so thirsty that they may be at the waterhole during the day (in the dry season, lion presence can be associated with daytime). Also, elephants tend to drive lions away from the waterhole, so the presence of elephants offers some safety (elephant presence is associated with lion absence). When many associations are at stake and/or actions must be fast, probabilistic associative patterns, derived from multiple episodes, drive expectations and learning during future interactions in the environment.

Much, probably the majority, of learning occurs unconsciously. Contemporarily, dangers may differ but unconscious expectations, formed through learning and retained as unconscious memories, still drive fast responses to threats (Öhman, Carlsson, Lundqvist, & Ingvar, 2007). Concomitantly, it would be expected that the elaborative, complex, associative, probabilistic and meaningful experiential patterns formed in REM sleep would be unconscious. Arzi et al. (2012) showed that new associations, learned during REM sleep, were retained as unconscious memories, then functioning, during wake, as unconscious expectations which drove actions.

Does the human unconscious have any parallels in machine learning? We know that machine learning algorithms inherit unconscious human biases, indeed they can amplify them (Thomas, 2018). Moreover, humans can continue to reproduce machine learning bias, even after they no longer use the algorithm (Vicente & Matute, 2023). Clearly, unconscious learning and memory are not subject to voluntary control. Although unconscious biases can be detrimental (e.g., to marginalized groups) inductive, unconscious (or implicit) biases are essential for faster information processing both in humans and machine learning. Inductive biases depend on the associations formed through experience, in machine learning such associations occur during unsupervised learning. Specifically, in the authors' meta-learned models, associations will be experiential.

In sum, the authors assert that their meta-learned models are “invaluable tools to study, analyse and understand the human mind.” In humans, experiential, associative learning becomes associative memories that, then, drive expectations in subsequent learning. In the memory network architecture of the brain/mind these associative memories may take the form of the serial, associative pathways which underlie conscious learning. Equally, where these serial pathways cross at nodes/junctions, complex, hyper-associative, experiential patterns arise from elements of the different intersecting pathways. These patterns are retained as unconscious memories which associate elements of different experiences.

Association is fundamental to both learning and memory, whether conscious or unconscious. The authors' first research question of how people improve their learning abilities over time can be better addressed through acknowledging, first, the inextricability of learning and memory and, second, the role of conscious and unconscious association in each.