Reinforcement Learning Creates Human-Like AI

Dr. Norberto Grzywacz, a very talented and impressive professional from Loyola University Chicago, gave a talk at the Neuroscience Seminar about his research related to reinforcement learning on social polarization. Our values is how we make decisions and these values come from the society we live in. As life goes on we are constantly receiving feedback from others. When we do something others like, we get some kind of positive feedback. While the reverse is also true, when we do or say something others do not agree with we receive negative feedback. This shapes our future decisions and actions. This is called reinforcement learning. Reinforcement learning can lead to polarization because we like the feeling of positive feedback. Therefore, we interact with the people that give us this feeling, and the loop continues until we are only interacting with people that are the same as us, creating a false reality that everyone else is exactly like us or should be. Something striking that Dr. Grzywacz mentioned was that artificial intelligence (AI) learns in this same manner.

This article written by Julian Horsey delves into more insight of how AI learns in the same way and how human input impacts that learning. An already trained AI is used in this process of reinforcement learning from human feedback (RLHF). That means one that already has a large amount of data and understands the foundations of human language. First it is fined tuned to desired behaviors by giving the AI examples to guide it to correct responses. Secondly, humans help the AI learn. This is done as AI completes tasks, someone assigns the response with a corresponding number to label how desired the response was. This is reinforcement learning in the same way humans learn as a higher number equals a better response. Finally, the AI maximizes the desired responses based on the numerical values assigned earlier. However, RLHF has some drawbacks. The main ones are overfitting and bias. These lead to polarization within the machines themselves. As humans who do the numerical values are inherently biased, the machines can begin to give too narrow of responses or perpetuate stereotypes and malicious biases inadvertently. RLHF does have its benefits of generously increasing AI’s ability and scope to perform tasks in a more naturalistic manner with more nuanced communication. However, RLHF must be done carefully, just like in humans RLHF can lead to polarization and harmful outputs, but if done right it can lead to improved AI that can maximize its benefits to society as a whole.

Reinforcement learning is an incredibly interesting topic of how the brain works to belong within society and be “one of the pack”. However, it also highlights how humans can easily fall into the extremes. It is fascinating how AI can emulate this behavior already acting more human-like than ever before, signifying that we must remain cognizant of what AI is becoming and the implications that it can provide for our society both beneficial and detrimental.

 

References:

Grzywacz, N. M. (2025). Comparison of distance and reinforcement-learning rules in social-influence models. Neurocomputing, 649, 130870. https://doi.org/10.1016/j.neucom.2025.130870

Horsey, J. (2024, August 8). AI reinforcement learning from human feedback (RLHF) explained. Geeky Gadgets. https://www.geeky-gadgets.com/reinforcement-learning-from-human-feedback/

When Dreams Talk Back: Rethinking Cognition Across Sleep and Wake

            Dr. Gabriella Torres-Plata's research this semester challenged long-held sleep science assumptions and presented evidence that contradicts them. The most significant assumption she challenged was that dreaming could only be studied after the fact. In place of that, Torres-Plata and colleagues have shown that some people who become aware while they are dreaming (lucid dreamers) can receive information and respond while still in REM sleep - they can answer yes/no questions, do basic arithmetic, and identify different sensory inputs based on coded eye movements and slight facial expressions. Torres-Plata and colleagues have opened the door for studying what happens in the mind while we dream, including memory creation, emotional processing and creative problem solving. The previous method of studying dreaming has been limited to the degree to which a person can remember what happened while they were asleep. This study has the potential to help us understand how the brain works at all times, and to make better use of our ability to dream. 

The concept that Torres-Plata explored has additional support in a recent paper published in the journal Neuron titled "Linking Memory and Imagination Across Sleep and Wake States" (Wamsley et al.), that came out in 2024. While Wamsley and his colleagues did not explore dream dialogue, they did find similar brain activity in REM states (where dreams occur), and in states of daydreaming. They were able to observe the way that the brain changes old experiences into something new, using areas of the brain such as the hippocampus and cortical networks. Both processes involve flexible, creative recombination of past experiences, guided by the hippocampus and association cortices. The Wamsley et al. study suggests that dreaming is not just “random noise,” but part of a broader system for simulating, reorganizing, and sometimes inventing mental content. 

Ultimately, the possibility of having a real-time conversation with someone while they are dreaming may provide the opportunity to study how the process of transforming memories occurs on a moment-by-moment basis. However, it also raises many important issues related to privacy, autonomy and the extent to which researchers should be allowed to intrude upon an individuals internal mental processes. Real-time dialogue might eventually allow researchers to probe how memory transformations unfold moment by moment. At the same time, these possibilities bring up important concerns about privacy, autonomy, and how much access researchers should have to an individual’s inner cognitive world. 

Both the seminar and the Neuron article highlight REM sleep as an active and dynamic state—not a pause in consciousness, but a unique mode of thinking. Together they suggest that dreams may be a window into how we imagine, problem-solve, and reshape our emotional experiences, both during sleep and while awake. 

References 

Konkoly, K. R., Appel, K., Chabani, E., Mangiaruga, A., Gott, J., Mallett, R., Caughran, B., Witkowski, S., Whitmore, N. W., Mazurek, C. Y., Berent, J. B., Weber, F. D., Türker, B., Leu-Semenescu, S., Maranci, J.-B., Pipa, G., Arnulf, I., Oudiette, D., Dresler, M., & Paller, K. A. (2021). Real-time dialogue between experimenters and dreamers during REM sleep. Current Biology, 31(7), 1417–1427.e6. https://doi.org/10.1016/j.cub.2021.01.026 

Wamsley, E. J., Elfassy, N. M., Ambo, L. K., Wangsness, H., Mughieda, N., Jensen, O., & Bhattacharya, M. (2024). Linking memory and imagination across sleep and wake states. Neuron, 112(9), 1501–1518. https://doi.org/10.1016/j.neuron.2024.02.012 

Dreaming Circuits and Directed Memories: How Selective REM Control Sheds Light on Sleep’s Hidden Functions

    From this semester, the one striking idea portrayed was the discovery that the lateral posterior nucleus (LPN), a higher-order thalamic visual nucleus that takes part of a specific role in REM sleep regulation. Guerrero and Cavanaugh’s conduct of this study demonstrates that silencing the LPN selectively leads to the reduction of REM sleep in mice. In a way, initiating this method leaves non-rapid eye movement (NREM) sleep almost intact. Going along the same idea, selective circuits lead to governing sleep function, which is applied to NREM sleep, the critical phase of deep memory consolidation. This proposes the fact that REM may rely on targeted support from sensory-related thalamic circuits instead of whole-brain sleep mechanisms. In another sense, REM isn’t just “dreaming turned on” but appears to be an adapted situation, in which the neural structures are working all around and stepping in at the right time needed. This finding raises the question of whether REM is supported by specialized neural circuits, and could altering REM change memory, emotion, or even dream content?

    A 2021 study from Northwestern University was widely covered in news outlets such as NPR and Science Daily. They examined the occurrence called target memory reactivation (TMR), which is the process of activating the brain with keen sounds during sleep to either strengthen dreams or to modify them. With this study presented, the participants practiced a melody on a keyboard while also listening to a specific sound. The scientists replayed that cue during sleep, which led to the factor the participants performing the melody significantly better. The results acquired from these participants the next morning were better than allowing that cue to be replayed overnight. The key point of this research displayed how specifically REM and light NREM stage sleep are not submissive, therefore is considered an active state. That is because the specific brain circuits present selectively alter what we perceive and how well we remember it. This study resonates with Guerrero and Cavanaugh’s work. If REM sleep needs the LPN to maintain its normal structure,  then damaging that specific circuit could interfere with the neural cycle that TMR relies on. There could be an indication of an “entry point” for comprehending how sensory information, dreams, and memory consolidation could interact at a circuit level. 

    In the TMR study done at Northwestern, memories are reinforced by activating certain cortical pathways during sleep. Looking into the mouse in the REM study, by selectively silencing a visual thalamic structure, it weakened REM, which supports memory consolidation. There is the idea of distinct brain circuits that may play a role, determining what kinds of information get processed during sleep. These studies have opened the door to a new interpretation of sleep, how it's a series of regulated brain modes in which each is linked and supported by distinct neural structures with special functions. Sleep ultimately becomes less like a blanket covering for the brain but more like a series of subtly tuned spotlight beams that illuminate the different neural pathways, all at different times. As the continuation of learning proceeds, on how sleep interacts with memory and perception. The idea that one day we may influence how we dream seems to be realistic, not just being able to know whether we dream or not. The conclusion from the studies of Guerrero and Cavanagh suggests how these possibilities begin in the cortex but also how surprisingly powerful thalamic circuits can shape the landscape of our sleeping minds.

References

Guerrero, J.R., & Cavanaugh, J. (2025). Silencing the lateral posterior nucleus produced a highly selective reduction in mouse REM sleep. Neurobiology of Sleep and Circadian Rhythms. No DOI available. https://www.sciencedirect.com/science/article/pii/S2451994425000136  

Anthony, J. W., Cheng, L. Y., Brooks, P. P., Paller, K. A., & Norman, K. A. (2021). Sleep spindle refractoriness segregates periods of memory reactivation. eLife, 10, e70068. https://doi.org/10.7554/eLife.70068