Meta-learning modeling and the role of affective-homeostatic states in human cognition

Abstract

The meta-learning framework proposed by Binz et al. would gain significantly from the inclusion of affective and homeostatic elements, currently neglected in their work. These components are crucial as cognition as we know it is profoundly influenced by affective states, which arise as intricate forms of homeostatic regulation in living bodies.

Binz et al. offer a very promising research program based on meta-learning to advance our understanding of cognition. Nonetheless, their proposal could greatly benefit from integrating affective and homeostatic elements, which have been traditionally underrepresented in cognitive science and are totally absent in their article as well. This integration is predicated on the premise that cognitive processes in general are profoundly influenced by affective states (e.g., emotions and moods), which arise as intricate forms of homeostatic regulation in living organisms.

First, there is ample evidence from psychology and affective neuroscience showing the pervasive influence of affective states on cognitive processes (Cea, 2023; Clore & Schiller, 2018). When someone feels affectively moved by new information, she will more likely attend to it and think about it for an extended duration, compared to neutral information (Manns & Bass, 2016). People in positive moods tend to engage in more creative thinking and learn subjects meaningfully (Ormrod, Anderman, & Anderman, 2019), while those feeling sadness or frustration are prone to a shallower learning (Ahmed, Van der Werf, Kuyper, & Minnaert, 2013). This is related to negative emotions being accompanied by cortisol release and the fight-or-flight response that can suppress the prefrontal cortex, thereby hindering higher cognition (Brackett, 2019). Hence, it is reasonable that emotion regulation abilities are the strongest predictors of academic achievement in high-school students (Di Fabio & Palazzeschi, 2009).

Also, our affective states shape what and how we perceive (Barrett, 2017; Cea & Martínez-Pernía, 2023). According to the affective information principle, our feelings signal the value and urgency of any perceived object (Clore & Schiller, 2018). Moreover, by altering people's affects, researchers can sway perceptions, for example, making a beverage seem appealing or distasteful (Berridge & Winkielman, 2003), and people friendly or mean (Li, Moallem, Paller, & Gottfried, 2007). A key brain area involved is the orbitofrontal cortex, which integrates sensory and affective information, ensuring that our perceptions are always imbued with affect (Barrett & Bar, 2009).

Concerning attention, positive emotions generally broaden it, leading to a global focus (Fredrickson & Branigan, 2005), whereas negative emotions narrow it, fostering detail-oriented processing (Bar-Haim, Lamy, Pergamin, Bakermans-Kranenburg, & Van Ijzendoorn, 2007). Also, emotionally charged stimuli are shown to capture attention more effectively than neutral ones, both in terms of speed and focus (Hajcak, Jackson, Ferri, & Weinberg, 2018). This can cause perceptual interference, making subsequent neutral stimuli less noticeable (Wang, Kennedy, & Most, 2012), a process associated with emotional allocation of attentional resources as indicated by the late positive potential in the parietal lobe (Hajcak et al., 2018).

Concerning memory, William James claimed that “an experience may be so exciting emotionally as almost to leave a scar upon the cerebral tissues” (James, 1890, p. 670). People tend to remember information with emotional significance more easily than neutral material (Phelps & Sharot, 2008). Memories associated with strong emotional arousal are recalled more vividly (Schaefer & Philippot, 2005) and with greater detail in certain respects compared to neutral memories (Kensinger & Schacter, 2018). Neurobiologically, the reciprocal interactions between the amygdala and the hippocampus are considered essential for this (McGaugh, 2013).

Concerning the homeostatic roots of affect, Seth and Barrett have independently suggested that affect arises from the brain's inferences about the causes of internal body signals to regulate physiological states (e.g., sugar levels, heart-beat, etc.), ensuring survival based on past experiences (Barrett, 2017; Barrett & Simmons, 2015; Seth, 2021; Seth & Tsakiris, 2018). Hence, according to them, our feelings would then be expressions of our current and future degree of success or failure in staying alive. In this way, moods and emotions would be intimately linked to our bodily nature and homeostatic needs.

This core idea of feelings being rooted in homeostatic regulation in vulnerable systems has been applied to robotics and artificial intelligence. Man and Damasio (2019) propose a novel class of soft robots that incorporate physical vulnerability and self-regulation akin to living organisms. They hypothesize that this would allow them to develop motivations and evaluations reminiscent of feelings in humans, potentially leading to more intelligent interactions with their environments. Similarly, Bronfman, Ginsburg, and Jablonka (2021) propose that feelings may arise in artificial systems constructed with soft materials through the development of homeostatic, self-preservation mechanisms that could allow them to instantiate an open-ended domain-general form of learning, what they call unlimited associative learning. Finally, Yoshida (2017) introduces a reinforcement learning model where agents learn to survive by regulating critical variables like energy levels, using a reward system rooted in homeostatic principles, leading to adaptive behavior. Importantly, all proposals emphasize that the possibility of engineering sentient artificial systems depends on having vulnerable bodies that, akin to ourselves, need to be constantly sensed and regulated to remain integral, and that this would enhance the machines’ cognitive capacities.

To conclude, I would like to suggest some potential benefits of incorporating affective and homeostatic elements into the meta-learning research program: (i) Enhanced adaptability: By incorporating these elements into the meta-learning algorithms, the resulting computational models may better simulate the adaptability of human cognition, like the ability to adjust learning strategies to changing environmental and internal states; (ii) richer contextual understanding: Incorporating affective-homeostatic elements in learning-to-learn processes can result in a deeper understanding of how emotionally salient contexts influence cognition; (iii) improved learning efficiency: Affective-homeostatic signals can guide attention and memory processes, leading to more efficient learning. Meta-learning algorithms that incorporate affective-homeostatic signals or mechanisms may achieve higher efficiency in adapting learning algorithms to new tasks or to new information; (iv) more realistic simulations: By incorporating affect and homeostasis, meta-learned models may more accurately simulate human cognition, which is inherently influenced by affective-bodily states; and (v) cross-domain generalization: The integration of affective-homeostatic states may facilitate better generalization across different cognitive domains, as affect and bodily regulation often play a role in a wide range of cognitive tasks, from decision making to social interactions.

In sum, I encourage Binz et al. to consider the beneficial prospects of incorporating these elements into their proposed research program, so as to acknowledge the intertwined nature of affect, bodily homeostasis, and human cognition.