If you read ‘YMC’, what do you think comes next? You weren’t told to learn it, but you intuitively know it’s ‘A’. Our brains can pick up patterns in the world like this without being taught, rewarded, or corrected. This process, called statistical learning, is well studied in humans and underlies how children learn language – “ba” is more likely to follow certain syllables than others, for example.

In the sensory complexity and uncertainty of the natural world, where guidance is scarce and rewards are rare, many other species use this form of learning, too. But exactly how the brain does this has been unclear. Now, a new preprint from the Akrami lab at the Sainsbury Wellcome Centre (SWC), led by Dr Dammy Onih, has shown how mice, like humans, can quickly learn multiple types of patterns: how often sounds happen, the order in which they appear, and even abstract rules connecting them, purely from incidental experience, without instruction or reward.

Their work shows that the dorsal CA1 (dCA1) region of the hippocampus is not just involved but plays a causal, mechanistic role in forming, updating, and generalising internal models of sound patterns. By combining targeted optogenetic interventions with large-scale neural recordings, the work reveals how the hippocampus performs true unsupervised statistical learning, offering new insight into one of the brain’s core computational abilities.

“Statistical learning is less well studied than reward-based learning, but it's innate to all of us. It’s also efficient, as we don’t need to learn every single detail; we can learn a structure and then quickly map new experiences onto that. Understanding how the brain does this is an important and exciting challenge for neuroscience,” says Dr Onih, Research Fellow at SWC.

A new reinforcement-free paradigm

To study statistical learning, the team developed a new task paradigm that both mice and humans could complete. Subjects undertook a simple auditory target detection task and learned that a noise (X, the target), embedded in a stream of background sound, was linked to reward. They had to respond as soon as they heard X. At the same time, a structured tone sequence (ABCD) was presented at random times, sometimes present, sometimes not. These sequences were irrelevant to the task and never predicted the target sound X. As the sequences appeared at unpredictable times, there was no reason for subjects to pay deliberate attention to them.

To test whether subjects learned the statistical properties of the tone sequence pattern, the team changed its presentation rate (i.e. how often it appeared), sequence identity (pitch of the tones), and the abstract rule governing the sequence structure (e.g. ABAD was presented, instead of ABCD). They monitored pupil dilation as a measure of surprise to the changes. 

Why pupils? Previous work in humans has shown that pupil dilation consistently increases when expectations are violated. Crucially, this surprise can only occur if an expectation has already been formed, and expectations arise through learning. The team reasoned that pupil responses could provide a powerful, non-verbal readout of implicit learning. In this study, they show for the first time in mice that changes in pupil size reliably signal sensory surprise driven by learned patterns. This establishes pupil dilation as a practical indicator for statistical learning in an animal model.

In both mice and humans, responses to X were unaffected by variations to the ABCD tone sequence. But, altering the tone sequence did surprise the subjects, as indicated by pupil dilation. This dilation response was seen when the regularity of the tone sequence changed, the pitch changed, or the sequence rule changed. 

“We see extraction of multiple levels of statistical structure from passive background noise,” says Dr Onih. “The key one is that the mice can generalise across distinct events that have a common rule or a common structure between them.”

The role of the hippocampus

Evidence from multiple lines of research suggests that the hippocampus plays a key role in learning patterns and predicting what’s next. To examine its role in statistical learning of sounds, the team began by pharmacologically silencing the dorsal CA1 (dCA1) region.

“Inactivating the hippocampus showed us that animals are less surprised by things that should have been surprising,” explains Dr Onih, “indicating the hippocampus is central to statistical learning.” 

They then used optogenetic silencing to identify when dCA1 activity is required. Briefly turning off dCA1 during initial exposure to a sound pattern prevented animals from forming normal expectations, so later, when the pattern was violated, their pupils no longer responded. This demonstrates that dCA1 is essential for building internal models that detect the unexpected.

Importantly, hippocampal inactivation did not affect attention, arousal, or the animals’ ability to detect the rewarded sound X. This ruled out simpler explanations and showed that disrupting dCA1 specifically impaired their ability to learn background structure.

Neural representations

To uncover precisely how the hippocampus supports statistical learning, the team recorded dCA1 neuronal activity as mice performed the task.

They found that dCA1 neurons did represent the tones A, B, C, and D. As the animals gained experience, the neural patterns for these tones became more alike, suggesting that the brain was grouping together sounds that belonged to the same sequence.

The neurons also tracked statistical properties: rare or unexpected events caused stronger responses, and the population activity reliably signalled when a sequence order violated expectations. These neural changes closely tracked the animals’ pupil responses, showing that hippocampal population dynamics evolve as expectations are updated.

“Our results show that CA1 simultaneously encodes sensory features and statistical context,” says Dr Onih. 

They also investigated the neural representations of abstract rules. They observed that over time, neural representations of tones in the same pattern (ABCD) but at a different pitch, became more similar.  This was reflected in the pupil responses, which were larger to rule-violating than rule-conforming deviants (e.g. ABAD vs ABCD at a different pitch). This indicates that the brain grouped its coding of tones with the same underlying structures, suggesting the animals had learned and generalised the abstract sequence rule.

“We provide evidence that the hippocampus is well suited for extracting out relationships, understanding connections, and in general understanding the structure in the environment,” said Dr Onih.

Understanding statistical learning

This work offers the clearest evidence yet that the hippocampus is a core engine for discovering hidden structure in the world. By developing a new reinforcement-free learning paradigm in mice and pairing it with targeted optogenetic disruption and large-scale recordings, the team shows that dCA1 is indispensable for rapidly forming, updating, and generalising internal models from experience. 

“I first encountered statistical learning research almost twenty years ago during my PhD, through friends studying language acquisition in human infants. I was deeply inspired by landmark studies such as Saffran et al, showing that even pre-verbal babies can learn structure from passive listening, and can reveal what they’ve learned through surprise and novelty preference, long before they can speak. For years, I wanted to bring this kind of question into animal models. That idea (that learning can be inferred from expectation violation), motivated me to search for a comparable approach in animals: one that would let us probe acquired knowledge in rodents without instructing them, rewarding them, or forcing them to learn specific rules. Being able to finally do that now, thanks to Dammy’s work, and to uncover such a central role for the hippocampus, is incredibly exciting!” said Dr Athena Akrami, Group Leader at SWC.