Theses Doctoral

Biological models of prediction and memory formation in the hippocampus

Fang, Ching

The hippocampus is a brain region known to support several cognitive functions. In particular, it is necessary for the formation of episodic memories. These are memories of personal experiences, and they are formed in an one-shot manner. The hippocampus has also been suggested to support the formation of cognitive maps. Cognitive maps are mental representations of how concepts or locations are connected to each other. These maps may be formed through predictive learning in the hippocampus. However, it is unclear how either of these processes are supported on the level of neural circuits. This thesis aims to understand, through theoretical modeling, how networks of neurons are able to organize and learn to function as episodic memory systems and predictive centers.

In chapter 2, we first explore how predictive cognitive maps may be formed in a biological circuit. We begin by analyzing a popular description of hippocampal activity called the successor representation (SR). The SR is an algorithm that models population activity in the hippocampus as a rollout of a transition probability matrix estimated from the animal's experience. We explore how this algorithm can be learned by neurons in the brain by deriving an equivalent neural circuit that can learn the SR using neurally plausible learning rules. We simulate activity from this neural network and show that it matches experimental predictions of hippocampal activity. A key component of the model we construct is the ability to use the strength of recurrent connectivity as a means to control the time horizon of prediction from the model. This feature of the model can support flexible use of prediction across different cognitive tasks. Overall, this work suggests a biological mechanism for how predictive activity may arise in the hippocampus.

In chapter 3, we investigate how a predictive region (similar to the model of the hippocampus discussed in chapter 2) may influence representations found in other brain regions. Specifically, we take inspiration from deep reinforcement learning (RL) to construct a multi-region model. In deep RL, the state space of the agent must be inferred from high-dimensional and complex sensory inputs. Thus, deep RL systems are imbued with sensory encoders that estimate the state of the agent. In a model-free setting, this state estimate is then passed to a value learning system. To improve the representations learned by the encoder, it is standard practice to add auxiliary objectives to the model. A common auxiliary objective is predictive learning via an additional predictive network. The use of a sensory encoder, value learning system, and predictive network parallels the suggested functions of the sensory system, striatum, and hippocampus, respectively. Keeping in mind this biological analogy, we explore how predictive objectives shape representations across a deep RL network. We discuss how this may suggest a role for the hippocampus as a representation learning system to support other brain regions.

In chapter 4, we return to a crucial function of the hippocampus-- that of an episodic memory store. We aim to develop a biological model of memory storage and retrieval. Typically, neural network models of the hippocampus are based off autoassociative networks like the Hopfield network. Here, we consider an alternative instantiation of episodic memory inspired by models used in machine learning. In machine learning, memory is often stored in key-value, or slot-based, systems. In these systems, memory is stored in slots addressable by ``keys'' that may be unrelated to the memory content, or ``value''. However, there is not a clear biological interpretation of these types of memory systems. In this chapter, we suggest a biological implementation of a key-value memory network using a feedforward network with neural learning rules. This network is capable of faithfully storing many memories, even when correlations are present across memories. We also discuss how key-value memory networks are reminiscent of classic theories of hippocampal memory that describes hippocampal activity as an ``index'' into cortical memories. Beyond suggesting a specific model of biological key-value memory, we propose an alternative view on hippocampal memory.

In chapter 5, we develop the ideas from chapter 4 further by taking inspiration from recent experimental findings in the hippocampus of black-capped chickadees. Black-capped chickadees are memory-specialist birds that are model organisms for the study of long-term memory in the hippocampus. Specifically, they engage in food-caching behavior that requires the ability to precisely recall the location of many food caches. To gain insight into memory formation, Chettih 𝑒𝑡 𝑎𝑙. [1] recorded hippocampal activity from these birds while they cache and retrieve seeds. The authors discovered neural activity encoding the location of the animal and activity encoding the presence of seeds. In addition, they also discovered --in the same neural population-- sparse, high-dimensional activity patterns that were unique to each cache and highly uncorrelated. These ``barcodes'' are a suggestion of index-like activity in the hippocampus. In this chapter, we design a recurrent neural network that replicates experimental findings from Chettih 𝑒𝑡 𝑎𝑙. [1]. We show how an indexing-based memory system is functionally advantageous as it allows for precise storage of potentially correlated memories. This work, which unites experimental and theoretical discoveries, suggests a re-imagining of classic theories of hippocampal memory.

In this thesis, we have sought to understand the system-level mechanisms that support hippocampal function. While it is understood that the hippocampus is important for cognition, much is still unknown about the biological processes underlying this region. We believe our findings here have deepened our understanding of the hippocampus and suggested new avenues of research to further the field.

Files

  • thumnail for Fang_columbia_0054D_18881.pdf Fang_columbia_0054D_18881.pdf application/pdf 6.72 MB Download File

More About This Work

Academic Units
Neurobiology and Behavior
Thesis Advisors
Aronov, Dmitriy
Abbott, Larry
Degree
Ph.D., Columbia University
Published Here
October 30, 2024