Papers and whatnot
I am an assistant professor of psychiatry at Columbia University who studies computational neuroscience. I also am co-founder of: Fauna Robotics, a consumer-facing embodied AI company; Herophilus, a drug discovery company; and Neuromatch, an online neuroscience research and education community.
Sean Escola (PI)
Manueal Beiran Amigo (postdoc)
Andrew Chen (postdoc and resident in psychiatry)
Kaushik Lakshminarasimhan (postdoc)
Salomon Muller (postdoc)
Yuriy Shymkiv (postdoc)
James Murray (former postdoc, now assistant professor at the University of Oregon)
Laureline Logiaco (former postdoc, now a postdoc at MIT in Ila Fiete’s lab)
Jack Lindsey (now at Anthropic)
To investigate this, we have developed a model of motor cortex, the basal ganglia, and thalamus which has offered insights into the computational roles that each of these structures play in sequence generation. Specifically, we show that the activation and inactivation of different cortical-thalamic loops by the basal ganglia can control the dynamics of motor cortical activity in order to produce multiple behavioral outputs when the projection to the spinal cord is fixed. This is a novel hypothesis for the role of the neuroanatomy in motor computation.
We are applying biologically constrained learning rules to the thalamo-cortical weights while constraining intracortical, readout, and cortico-thalamic weights to be fixed. We show that learning is much more successful when cortico-thalamic weights match the readout, a testable hypothesis. Furthermore, when we restrict the readout as arising from a subpopulation of the cortical network (analogous to cortical layers 5 and 6 in vivo), we find that learning at synapses between thalamus and the non-readout projecting cortical units is no longer possible unless the intracortical connectivity obeys a specific structure. This work unifies neuroanatomy, biological learning, and computational goals to make specific predictions about motor system synaptic weight structures.
In collaboration with Dr. Bence Ölveczky’s lab at Harvard, we have been building models to understand the result that motor cortex is necessary for the learning but not for the execution of complex behavior, while thalamus and striatum are necessary for both learning and execution. We have shown that if the connections between cortex and striatum learn relatively quickly using a supervised or reinforcement learning rule, while those between thalamus and striatum learn relatively slowly using a simple Hebbian learning rule, then the thalamic inputs to striatum will learn to mimic the cortical inputs during repetition of the same behavior. This results in transfer of control from cortex to thalamus. Recently, Dr. Ölveczky and colleagues have shown that when animals learn multiple similar complex behaviors, transfer of control from motor cortex to thalamus is no longer successful. Our current modeling efforts show that there is an intrinsic tradeoff in motor learning between flexibility and robustness and that interference of control transfer occurs when, for a given task, flexibility is prioritied.
Historically, neuroscience has played a large influence on the development of AI, mostly famously by inspiring the core architectures of the neural networks that underpin the ongoing machine learning revolution. We believe that there is much more that biologically intelligent systems can teach AI and that a path towards elicidating this is through the development of virutal embodied agents that are trained to accurately recapitulate the detailed behaviors of in vivo animals. This line of thinking is well captured by a 1988 quote from AI pioneer Hans Moravec, who said that abstract thought “is a new trick, perhaps less than 100 thousand years old…effective only because it is supported by this much older and much more powerful, though usually unconscious, sensorimotor knowledge.” To this end, we are developing a platform for virtual rodent neuroscience that will allow us to carefully and exhaustively probe the mechanisms by which animal brains produce ethologically adaptive behaviors.
Moreover, this platform also serves to advance basic questions in neuroscience by allowing for virtual experiments to be performed that can compliment ones in the lab. A cycle of virtual neuroscience predictions, experimental testing, and updated models can ultimately result in a robust testbed in which high fidelity experiments can be conducted at scale to vastly accelerate the pace of neuroscience research.
Historically, neural recordings are analyzed by aligning data to experimentally known times (e.g., stimulus onset, movement onset, etc.). However, given that motivational and attentional features are not clearly accessible behaviorally, and that computations may have variable durations from trial to trial, it is possible that there exist multiple internally relevant times that strongly influence neural responses. We seek to develop tools to infer these internally relevant times and thus potentially provide more complete characterizations of neural responses.
I have recently been involved in the founding and development of two external organizations with this as part of their missions. Herophilus, Inc., a for-profit drug development entity, applies systems neuroscience and machine learning to patient-derived human cerebral organoids for the discovery of novel disease phenotypes and drug treatments for complex neuropsychiatric disorders. This effort has led to multiple biological and technological advances thus far. Neuromatch, Inc., founded in response to the Covid pandemic, is a not-for-profit conference and summer school organization that bring high-quality zero-cost access to neuroscience education to students globally. In 2020, our summer school enrolled 2000 students; in 2021, 4000 students.
For a complete list, visit my Google Scholar Profile