Projects past & present

Meta RL as a model of cognition

pretty picture Reinforcement learning has proven a succesful paradigm for explaining and understanding human and animal behavior in many contexts. However, while many canonical RL algorithms rely on slow parameter updates, humans and animals are able to learn from their interactions with the environment on a timescale of seconds through recurrent network dynamics. Recently, ‘black box’ meta-RL with recurrent networks has been proposed as a model of such rapid reinforcement learning in the brain. In this work with Marcelo Mattar at UCSD & Guillaume Hennequin in the CBL, we further investigate the similarities between such artifical meta-RL circuits and biological networks with a focus on planning and inference in environments with complex and changing transition structures. In particular, we analyze how model-based behavior arises from the internal dynamics of the artificial recurrent networks and compare both the dynamics and emergent behavior with biological circuits.

Natural continual learning

pretty picture
Humans and animals are known to learn many different tasks over the course of their lives, while artificial agents are prone to ‘catastrophic forgetting’ whereby performance on previous tasks deteriorates rapidly as new ones are acquired. This shortcoming of AI systems has been addressed using Bayesian weight regularization which constructs an approximate posterior loss function over all tasks, and alternatively using projected gradient descent which limits parameter updates to directions of state space that will not interfere with previous tasks. In this work with Calvin Kao & Guillaume Hennequin in the CBL and Alberto Bernacchia at MediaTek Research, we argue that it is important in continual learning to have both good stationary points as in weight regularization and a stable optimization path as in projected gradient descent. We then develop ‘natural continual learning’ which combines continual learning using a Laplace posterior with projected gradient descent based on an approximate posterior over the parameters from previous tasks.

Bayesian GPFA with automatic relevance determination

pretty picture
Methods for dimensionality reduction such as PCA, tSNE, GPFA, LFADS etc. are frequently used in neuroscience to try to decipher what low-dimensional quantities are represented by high-dimensional neural activity. Some of the challenges associated with these methods include high computational costs, difficulty in optimizing hyperparameters to avoid overfitting, and deciding on an appropriate latent dimensionality. In this work with Calvin Kao, Jasmine Stone and Guillaume Hennequin in the CBL, we address these challenges by developing a scalable Bayesian version of Gaussian Process Factor Analysis which infers the appropriate latent dimensionality from the training data together with the remaining hyperparameters. The Bayesian approach ensures appropriate regularization to avoid overfitting, and our scalable implementation supports non-Gaussian noise models and can be used for long time series spanning many minutes or even several hours. Code

Manifold Gaussian process latent variable models

pretty picture
Most latent variable models used in neuroscience and machine learning implicitly assume Euclidean latent states. However, this is at odds with how the brain is thought to represent quantities such as head direction and certain motor plans. With Calvin Kao and Guillaume Hennequin in the CBL, we build on work by Anqi Wu in Jonathan Pillow’s lab to extend the Gaussian Process Latent Variable Model originally developed by Neil Lawrence to cases where the latent space is no longer Euclidean. We expect this framework to be useful for querying spatial representations in the brain ranging from navigation circuits to cognitive processes and motor control. Current work-in-progress includes extending the probabilistic mGPLVM framework to a Poisson noise model, adding a prior over the temporal smoothness of neural activity, and further applying the framework to experimental data. Code

Stability of motor memories

pretty picture Motor memories such as the ability to serve in tennis can last for many months and years even without further practice, but it remains unclear how such memories are stored at the circuit level. The literature on this topic has seen conflicting results, but most of these suffer partly from an inability to record the same neurons for long periods of time during complex motor behaviors. In this work with Bence Ölveczky at the Harvard Center for Brain Science, we investigate the stability of the neural correlates of two motor behaviors, one learned and one innate. We record from both motor cortex and dorslateral striatum and find stable correlates of behavior in both brain regions. This suggests that stable single-neuron activity can underlie persistent memories which in turn suggests an underlying circuit with a stable core that is largely unperturbed after learning.

Head direction circuits in the Drosophila central complex

pretty picture Head direction circuits have been characterized in organisms ranging from insects to mammals and are important for navigation. The head direction circuit in Drosophila melanogaster has been characterized in an effort led by Vivek Jayaraman in a series of papers between 2015 and now. This circuit exists in the so-called central complex of the fly which consists of multiple recurrently connected neuropils. In this work with Daniel Turner-Evans and Vivek Jayaraman at Janelia Reseaerch Campus, we characterize the fly head direction at molecular, synaptic, functional and behavioral levels in the context of Ben-Yishai’s canonical ring-attractor model. We verify that the features of the circuit are consistent with such a ring-attractor model as previously hypothesized. However, we also find additional local connectivity and distributed functions which we hypothesize might help increase the robustness of the circuit.

ab initio modelling of electron transfer reactions

pretty picture Electron transfers are ubiquitous in the world around us, taking place in proceesses ranging from solar cells to photosynthesis and respiration. However, such electron transfer reactions are notoriously hard to model computationally, in part due to the relative difficulty of modelling excited states of molecular systems. In this work with Alex Thom in the Cambridge Department of Chemistry, we used Hartree-Fock based methods to formulate electron transfer reactions as paths on the HF energy surface. This builds on work by Alex Thom and Martin Head-Gordon showing that multiple Hartree-Fock solutions of the same molecule resemble adiabatic electronic states. We can therefore model an electron transfer as a trajectory between local minima on the HF energy surface, where we find the optimal reaction trajectory in geometry space using a constrained optimization on the hyperplane where the reactant and product states have the same energy.

Efficiency of CRISPR/Cas9-mediated genome editing

pretty picture CRISPR/Cas9 has revolutionized the biomedical sciences since it was first introduced for reprogrammable genome editing in 2012. However, much remains to be understood about the system and how it operates in mammalian cells. In this research project with Yonglun Luo at Aarhus University, we used assays in HEK293 cells to investigate which biochemical factors affect the efficiency with which we can edit the human genome. We found two major contributors to the gene editing efficiency. First, we found that the chromatin state at the target site was important with editing being less efficient in heterochromatin were the DNA is more tightly bound to histones. Secondly, we found the capacity for RNA secondary structure formation of the guide sequence to affect editing, which we expect to be due to its effect on the strength of binding to the target DNA.