Today, we had the PhD defence of Olgerta Asko at RITMO. Her research is super interesting in itself (check out this feature story for an overview). This blog post is following up on one of the points she made during her trial lecture that I hadn’t thought about before: the difference between generalisation, inference, and deliberation.
Towards deliberation
Olga argued that current AI—here understood as large language models (LLMs)—are based on generalisation. They extract patterns from a lot of data and apply them broadly. As we have seen with recent commercial products and as I have explored in many ways on this blog, LLMs excel at this task.
Inference, on the other hand, builds on generalisation but uses learned patterns to draw conclusions about new situations. This is much harder for an LLM, which will be bound by what it has been trained on.
Deliberation is even harder, representing the conscious, reflective process of weighing options and reasoning through complex decision-making. Interestingly, at times, when I get help from CoPilot solving programming problems, it seems to reason with itself. It presents various solutions and lets me choose one. This appears to be a deliberation process, but, in fact, it is still only based on what is in the model.
Memory for prediction
In her trial lecture, Olga argued that memory is profoundly forward-looking. It is designed, through evolution, to help us simulate, predict, and prepare for the future.
The idea that memory is essential for future thought was pioneered by psychologist Endel Tulving, with Thomas Suddendorf and Michael Corballis later coining the term “mental time travel” to describe our ability to project ourselves backwards in time (retrospection) and forward into the future (prospection). They argued that these are not two separate abilities but two directions of the same underlying brain system.
Evidence for the link between retrospection and prospection was presented in a 2007 neuroimaging study led by Donna Rose Addis. Participants were asked to either recall a detailed past event or imagine a plausible future one. In the trial lecture, Olga tested us: we first had to remember a childhood birthday and later think about a future one.
There was no way to see the effect on us in the audience during her trial lecture, but in the Addis study, participants were in an fMRI scanner, and the results showed that the brain activity was almost identical for the two tasks. The same core regions lit up in the fMRI images, suggesting the brain uses the same “machinery” to reconstruct the past as it does to construct the future.
The Addis study only showed a correlation. However, other studies with patients who have hippocampal damage provide a causal link. These patients cannot recall past personal events and, interestingly, they are also unable to imagine future ones. They can retrieve general facts, such as that a beach has sand, waves, and sunshine, but they cannot bind these elements into a single, coherent scene. This proves that the machinery for re-living the past is necessary for pre-living the future.
The Default Mode Network
I have only heard of it in passing, but the default mode network (DMN) is highly active when we are not focused on a specific external task. This was once thought to be the brain simply “doing nothing.” However, this “do nothing” network is the same brain system that supports remembering the past and envisioning the future. As some say, your brain’s downtime is actually used for “mental time travel”.
Patients with damage to the ventromedial prefrontal cortex, a key node in the default mode network, experience a dramatic reduction in mind-wandering. Their thoughts become “restricted to the present,” and they almost entirely lose the ability to project themselves into the past or future.
Episodic memory
According to the “Constructive Episodic Simulation Hypothesis” by Daniel Schacter, your brain actively reconstructs future scenarios by pulling and recombining details from multiple past experiences. The idea is that your memory is not a passive replay system; it retrieves details from multiple past episodes and recombines them into novel episodes, happenings, and scenarios that haven’t happened yet.
The ability to rebuild the past to simulate the future is critical for shaping daily decisions. Humans have a natural tendency known as “delay discounting,” in which we prefer smaller, immediate rewards to larger, delayed ones. To explain this phenomenon, Olga played a video of the famous “Marshmallow test”:
Episodic future thinking is how we make a future reward more concrete and emotionally real. This increases its perceived value in the present moment, making you more patient and willing to make choices that benefit your long-term goals.
Mixing memory and imagination
The point is that decision-making is based on a combination of memory and imagination. These processes are deeply intertwined. We learn from the past to simulate the future and guide our choices in the present. This ability appears to be uniquely human.
AI has access to vast amounts of “knowledge”, at least data and patterns between them. However, this knowledge is not grounded in lived, embodied experiences. As Olga summarised it, AI has never been anywhere; it has never felt anything, it has no personal episodes to draw from. Ultimately, it has no “self” doing the remembering or imagining.
Humans don’t just remember; we have lived, embodied experience that grounds our episodic memory and helps us in making decisions. Machines are still far from having the same, although our research in embodied AI is going in that direction.
I co-wrote this post with NotebookLM, which summarised Olga’s defense from the video recording. Grammarly improved the grammar.