Research

Neural basis of episodic learning

Much of our day-to-day learning occurs not through repeated pairings of a stimulus and reward, but after single experiences, or "episodes", like a classroom lecture or a walk through a new part of town. How does the brain store the complex information in these episodes so quickly? Synaptic plasticity is widely accepted as a major contributor to learning and memory, with most models based on the classic Hebbian theory that "neurons that fire together, wire together", which is intuitively well suited for binding different patterns of information together. Most experiments, however, suggest that Hebbian plasticity is actually too slow for learning from a single experience, whereas faster, stronger plasticity rules operating at episodic learning timescales do not appear to be Hebbian. In a recent study with Stefano Recanatesi, we found that non-Hebbian plasticity mechanisms can in fact be extremely effective at binding together complex episodic information. In our model, the associative or sequential information usually thought to require Hebbian plasticity is instead reconstructed by prior knowledge such as the state space underlying a world model. A crucial consequence is that episodic recall may strongly depend on one's pre-existing knowledge, including how one makes predictions about the world, which has wide implications in arenas ranging from education to psychotherapy to information transmission through social networks.

Recently, colleagues at Princeton University and I developed an extended version of the model using a rule we term PRISM plasticity, which unifies rapid plasticity rules thought to operate in several animal species and brain regions. In this model, the PRISM rule rapidly modifies the dynamics of a recurrent neural network (RNN) by shaping the flow of information through a plastic feedback circuit. We find that PRISM plasticity is not only more biological, but also more powerful and flexible than traditional Hebbian rules, and exhibits deep links with both kernel and control theory, suggesting a new basis for a rigorous mathematical foundation for studying episodic learning in the brain. In a further study I found that PRISM-like rules may also be highly energy efficient, which provides a first-principles prediction of their widespread occurrence throughout the nervous system. In ongoing work I investigate how PRISM plasticity can be leveraged for advanced computations and how we can use the model and its kernel formulation to fit and interpret neural data.

Noise-powered computational dynamics in the brain

The brain is extremely noisy yet performs remarkable computations. Neurons emit action potentials, or spikes, with highly irregular timing, yet somehow at a collective level this instantiates a robust, flexible information processing system. How does the brain manage in the face of such noisy, irregular spike timing? Does all this noise just interfere with things, or could it be leveraged for computation? I recently developed a new spiking network model in which information in the network flows through changes in the noisiness (variance) of synaptic inputs, rather than their average values. In this study I found that networks of spiking winner-take-all units can produce smooth, controllable collective patterns yet which emerge from and are supported by noisy, irregular spiking. One can train these networks to produce highly flexible nonlinear dynamical patterns, including multistability and slow sequences (thought to support working memory), all powered by noise. Thus, noisy spike times in the brain could in principle not only co-exist with, but be essential to general-purpose dynamics and computation. In ongoing work I investigate how these networks can instantiate specific computations such as pattern completion or time-series transformations, and how to they can be fit to data to test theories about information processing in the brain.

Neuro-mechanistic modeling of natural behavior

With modern recording and computer vision technology, collecting and pre-processing natural animal behavior data has never been easier. But how can we best use this data to gain deeper insights into brain function? In a study with colleagues at Princeton we showed how to build mechanistic models to link together high-throughput neural recordings with separate natural behavior data to gain insights into complex neural codes for natural stimuli. By formalizing neural models representing distinct hypotheses, fitting these to neural data, then using them to simulate artificial neural activity alongside natural behavior data, we found that certain models' simulated activity predicts behavior much better than others, sometimes even outperforming state-of-the-art deep networks. We tested our approach to understand neural population codes for fruit fly social communication signals during courtship, which revealed a novel code for the female's memory of the male's courtship song. The inferred code leveraged three key, highly biological principles—adaptation, integration, and parameter heterogeneity—and exhibited computational advantages lacking in many artificial models, such as the ability to remember fine-scale timing information for long periods. This work thus shows how natural behavior data can be leveraged to generate insights into biologically plausible neural computations even without collecting simultaneous neural data—an important advance toward minimizing invasive animal surgeries and accelerating the design and interpretation of experiments.