List of Project Mentors (To Be Updated)
Mentor: Laureline Logiaco
Project: Using recurrent (switching) linear dynamical systems to characterize motor cortical activity during naturalistic motor sequencing
Primary motor cortical activity has been extensively characterized in simple delayed reach tasks and reach sequences. These studies have revealed the importance of two phases of neural dynamics: a preparation phase where the activity settles into a pattern that correlates with upcoming sub-movement characteristics, and - after trigger - an execution phase where the activity oscillates and is thought to drive the muscles.
Recurrent (switching) linear dynamical systems offer good methods to analyze motor cortical activity, and provide a good theoretical framework to understand how the motor cortical regimes of preparation and execution could support more complex movement sequences. However, challenges remain for (i) linking the analyses run on neural data (that typically assume that dynamics can be reduced to a low-dimensional set of factors) and the more general theoretical framework; and (ii) applying these analysis methods to data collected during more complex motor sequencing tasks.
In this project, you will:
1. Apply existing recurrent switching linear dynamical analysis methods (here and there) to artificial data generated by a model of motor cortical dynamics during complex motor sequencing
2. If time permits, apply these recurrent switching linear dynamical analysis methods to motor cortical data during a naturalistic and complex motor sequencing task
Mentor: Lakshmi Govindarajan
Project: The role of synaptic plasticity mechanisms in working memory manipulation
Working memory supports computational abilities that form the cornerstone of intelligence. And, identifying the neural substrates underlying these computations has been a problem of longstanding interest. Recent developments in neurophysiology indicate that synaptic dynamics unobservable through just neural activity readouts play a critical role in working memory computations, challenging several popular theories. Thankfully, end-to-end trainable recurrent neural networks provide the perfect scaffold to pit competing hypotheses against each other.
In this project, we will focus on a dynamic visual working memory manipulation task: 3D object rotations. Specifically, our goal is to test the role of various synaptic plasticity mechanisms in supporting memory manipulation.
- Studies have shown that rapid pre-synaptic facilitation/depression can support memory maintenance in an activity-silent manner. But, a potential limitation is that this works only on pre-formed patterns (fixed points of an attractor) and cannot generalize to novel inputs. Can we demonstrate this on image-computable models?
- Develop a heteroassociative plasticity rule (operating on a timescale faster than gradient updates) to support activity-silent maintenance on novel stimuli.
- Move the focus from memory maintenance to manipulation. The basic idea here is to computationally explain how representations in memory can be systematically manipulated to serve test-time decision-making (more on this to follow).
Mentor: Yudi Xie
Project: Multi-task models of ventral visual stream
Deep convolutional neural networks (CNNs) have been successful and widely adopted models in predicting neural and behavioral responses in vision. In the literature, features extracted from different layers in CNNs correspond to neural responses at different level of processing in the ventral visual stream, and there is a general trend that the better a model at performing image classification task, the better the model capture neural and behavioral responses. It is tempting to think that the ventral visual stream can be explained as performing invariant object classification. However, vision is much more than just classification. If you see an apple, you can perceive many properties orthogonal to the abstract category, such as its position, orientation, color, and texture. Additionally, experimental evidence suggested that category orthogonal information can be decoded from the neural representation in higher-level regions in the ventral visual stream.
In this project, we explore the idea that the visual system is not simply performing invariant object categorization. We will test the hypothesis that neural response in the ventral visual stream and behavior can be better captured by deep neural networks trained on a more diverse set of objectives, such as predicting the latent variables used to generate the training images and other widely studied computer vision tasks. Ultimately, we aim to apply these insights to developing better models of the ventral visual stream that can strengthen our understanding and be applied in practice.
What you can learn from doing this project:
- Hands-on experience in building various deep neural network models to understand the brain. (using Pytorch)
- Get a deeper understanding of current methods for evaluating neural network models based on neural and behavioral data, such as Brain-Score.
- Get exposure to various tasks in computer vision and understand their connection to the brain.
- Potentially contribute to a conference paper.
==================================================================
- Experiment: measure sparsityof attention in several available transformers (Dr. Rangamani)
- Experiments with ChatGPT on consciousness. Mirror-like self-awarness tests? Ask impersonation of a Zombie and of the opposite of it? (?)
- Any news on test with autoencoders? I think autoencoders are an interesting topic bridging research on associative memories with research on the structure of matrices induced by training deep nets....(TP)
- Experiment: regularization is needed if random initial conditions...as in linear case (TP)
- Why linear components converge quicker than nonlinear ones. Consider the different netwrks -- 1 layer, 2, 3...- effectively obtained by residual architectures. Assume they are all trained at the sametime (as they are). Depending on L number of layers each has a different \rho_k (see memo 112 august 2020 version). The networks with smaller L will converge faster! Thus the less complex solutions associated with shallower networks will converge faster! (TP)