Funded Seed Projects

Morozov/Zhang:
Generative models for predicting causal protein-signaling networks

Many objects in science and technology have composite nature – they consist of basic building blocks that can be added one-by-one to obtain a collection of final solutions with some useful characteristics. For example, we can imagine a molecule being built one functional group at a time. The resulting collection of molecules with perhaps four or more functional groups can be tested as an inhibitor, to prevent protein synthesis in harmful bacteria. In practice, cheaper computational tests will be run first, and only a few top candidates will be tested in the lab on live bacterial cultures. As a result, each tested molecule will be assigned an award based on its efficiency as an inhibitor. In this proposal, we describe an efficient AI approach to creating compositional objects such as graphs or strings out of their constituent components (graph nodes or letters) by starting from an empty state and adding components sequentially. This creates a unidirectional network in which each node represents the current state of the object, and the edges correspond to adding the next component. We will create deep-learning models of probability flows on this network, fitting them to sets of observed or predicted rewards. Our framework will enable building sequential pathways towards high-scoring solutions of various complex systems. In particular, we will predict protein-signaling networks directly from simultaneous observations of protein levels in single cells. Overall, we expect our AI-driven sampler to yield high-quality solutions to various hard optimization problems. Thus, our models will be useful to scientists and engineers in diverse fields of research.

 

Ramanathan:
Dendritic processing unit (DPU): neuromorphic software and hardware for continual learning at the edge

Design of hardware based on biological principles of neuronal computation and plasticity in the brain is a leading approach to realizing energy- and sample-efficient artificial intelligence and learning machines. An important factor in selection of the hardware building blocks is the identification of candidate materials with physical properties suitable to emulate the large dynamic ranges and varied timescales of neuronal signaling. This inter-disciplinary project aims to design a device that utilizes semiconducting quantum materials to emulate recently discovered synaptic plasticity in animal brains. The project brings together two vastly different disciplines namely electrical engineering and computational neuroscience to propose disruptive advances in AI hardware and algorithms. Specifically, the goals of this proposal are to develop a mathematical model for behavioral timescale synaptic plasticity and emulate the principal characteristics in an oxide semiconductor device capable of multiple timescale relaxations. We will extract the hardware behavior into a device model that can be used to implement larger scale network training algorithms to test efficacy in solving problems such as an agent navigating two-dimensional mazes.

 

Woojin Jung:
Improving distribution of food aid via trustworthy multimodal poverty

For Artificial Intelligence (AI)/machine learning (ML) models to be adopted for aid allocation decisions, the accuracy, trustworthiness, and explainability of need assessments require improvement. Our research creates an intuitive poverty map by combining explainable deep learning with a multimodal approach to identify vulnerable communities in need of food assistance. To capture the multifaceted human experience of poverty, we apply attention mechanisms to extract features from high-resolution/frequency satellite imagery (visual), social media features (verbal), and geographic attributes (spatial). For instance, we pair convolutional neural networks (CNN)- extracted satellite features with transformer-based embeddings to generate a joint representation for poverty at a small spatial unit. We augment these features and use them as input for a deep learning model to predict sub-national wealth using various ML algorithms. Then, we reverse-engineer this process to identify the regions of satellite imagery or specific social media topics that may contribute the most to the prediction of poverty. As a result, our models can generate and visualize poverty prediction at a granular level with justification, transparency, and evidence improving the overall trustworthiness. The detailed poverty maps generated by our prediction will be shared with development stakeholders to help them optimize the geographic distribution of food aid.