Graduate

Deep learning is a part of daily life, owing to its successes in computer vision and natural language processing. In these applications, the success of the model-free deep learning approach can be attributed to a lack of (mathematical) generative model. In yet other applications, the data is generated by a simple model and performance criterion mathematically precise and training/test samples infinitely abundant, but the space of algorithmic choices is enormous (example: chess). Deep learning has recently shown strong promise in these problems too (example: alphazero). In this talk, we study two canonical problems of great scientific and engineering interest through the lens of deep learning.

The first is reliable communication over noisy media where we successfully revisit classical open problems in information theory; we show that creatively trained and architected neural networks can beat state of the art on the AWGN channel with noisy feedback by a 100 fold improvement in bit error rate.

The second is optimization and classification problems on graphs, where the key algorithmic challenge is scalable performance to arbitrary sized graphs. Representing graphs as randomized nonlinear dynamical systems via recurrent neural networks, we show that creative adversarial training allows one to train on small size graphs and test on much larger sized graphs (100~1000x) with approximation ratios that rival state of the art on a variety of optimization problems across the complexity theoretic hardness spectrum.

Apart from the obvious practical value, this study of mathematically precise problems sheds light on the mysteries of deep learning methods: training example choices, architectural design decisions and loss function/learning methodologies. Our (mostly) empirical research is conducted under the backdrop of a theoretical research program of understanding gated neural networks (eg: attention networks, GRU, LSTM); we show the first provably (globally) consistent algorithms to recover the parameters of a classical gated neural network architecture: mixture of experts (MoE).

Software side channel attacks have become a serious concern with the recent rash of attacks on speculative processor architectures. Most attacks that have been demonstrated exploit the cache tag state as their exfiltration channel. While many existing defense mechanisms that can be implemented solely in software have been proposed, these mechanisms appear to patch specific attacks, and can be circumvented. We propose minimal modifications to hardware to defend against a broad class of attacks, including those based on speculation, with the goal of eliminating the entire attack surface associated with the cache state covert channel.

These modifications are layered on top of the Sanctum secure processor architecture that offers strong provable isolation of software modules running concurrently and sharing resources.

Joint work with Ilia Lebedev, Vladimir Kiriansky, Saman Amarasinghe and Joel Emer.

Stanford University has an amazing talent pool of student engineers, designers, storytellers, and entrepreneurs. Our theme focuses on inclusion and adoption, both in the market and in the industry, during a year of transition from early adopters to the consumer market. The goal of "VR for Everyone" is to increase exposure, awareness, and involvement in VR/AR for Stanford students and enthusiasts in the Bay Area, while promoting accessibility and diversity in industry. Read more

Reinforcement Learning (RL) is concerned with solving sequential decision-making problems in the presence of uncertainty. RL is really about two problems together. The first is the 'Bellman problem': Finding the optimal policy given the model, which may involve large state spaces. Various approximate dynamic programming and RL schemes have been developed, but either there are no guarantees, or not universal, or rather slow. In fact, most RL algorithms have become synonymous with stochastic approximation (SA) schemes that are known to be rather slow. This is an even more difficult problem for MDPs with continuous state (and action) spaces. We present a class of non-SA algorithms for reinforcement learning in continuous state space MDP problems based on 'empirical' ideas, which are simple, effective and yet universal with probabilistic guarantees. The idea involves randomized Kernel-based function fitting combined with 'empirical' updates. The key is the first known "probabilistic contraction analysis" method we have developed for analysis of fairly general stochastic iterative algorithms, wherein we show convergence to a probabilistic fixed point of a sequence of random operators via a stochastic dominance argument.

The second RL problem is the 'online learning (or the Lai-Robbins) problem' when the model itself is unknown. We propose a simple posterior sampling-based regret-minimization reinforcement learning algorithm for MDPs. It achieves O(sqrt{T})-regret which is order-optimal. It not only optimally manages the "exploration versus exploitation tradeoff" but also obviates the need for expensive computation for exploration. The algorithm differs from classical adaptive control in its focus on non-asymptotic regret optimality as opposed to asymptotic stability. This seems to resolve a long standing open problem in Reinforcement Learning.

Ultracold-atom model-systems offer a unique way to investigate a wide range of many-body quantum physics in uncharted regimes. Quantum gas microscopy enables us to "zoom in" both, in space and time, on a single particle level. We can explore many-body quantum physics in regimes that are not computationally accessible. In my talk I will present an overview of recent experiments, including the first observation of an anti-ferromagnetic phase of Fermions in an optical lattice, and the observation of many-body localization.

Semiconductor heterostructures are central for all modern electronic and optoelectronic devices. Traditional semiconductor heterostructures are typically created through a "chemical integration" approach with covalent bonds, and generally limited to the materials with highly similar lattice symmetry and lattice constants (and thus similar electronic structures) due to lattice/processing compatibility requirement. Materials with substantially different structure or lattice parameters can hardly be epitaxially grown together without generating too much defects that could seriously alter their electronic properties. In contrast, van der Waals integration, where pre-formed materials are "physically assembled" together through van der Waals interactions, offers an alternative "low-energy" material integration approach (vs. the more aggressive "chemical integration" strategy). The flexible "physical assembly" approach is not limited to materials that have similar lattice structures or require similar synthetic conditions. It can thus open up vast possibilities for damage-free integration of highly distinct materials beyond the traditional limits posed by lattice matching or process compatibility requirements, as exemplified by the recent blossom in van der Waals integration of a broad range of 2D heterostructures. Here I will discuss van der Waals integration as a general material integration approach beyond 2D materials for creating diverse heterostructures (e.g., dielectric/semiconductor and metal/semiconductor) with minimum integration-induced damage and interface states, enabling high-performing devices (including high speed transistors, diodes, plasmonic devices) difficult to achieve with conventional "chemical integration" approach. A particular highlight is the creation of van der Waals metal/semiconductor contacts free of interfacial disorder and Fermi level pinning, thus for the first time enabling experimental validation of the Schottky-Mott rule first proposed in 1930s.

Compressed sensing refers to the reconstruction of high-dimensional but low-complexity objects from a limited number of measurements. Examples include the recovery of high-dimensional but sparse vectors, and the recovery of high-dimensional but low-rank matrices, which includes the so-called partial realization problem in linear control theory. Much of the work to date focuses on probabilistic methods, which are CPU-intensive and have high computational complexity. In contrast, deterministic methods are far faster in execution and more efficient in terms of storage. Moreover, deterministic methods draw from many branches of mathematics, including graph theory and algebraic coding theory. In this talk a brief overview will be given of such recent developments.

Scenario generation is an important step in the operation and planning of power systems. In this talk, we present a data-driven approach for scenario generation using the popular generative adversarial networks, where to deep neural networks are used in tandem. Compared with existing methods that are often hard to scale or sample from, our method is easy to train, robust, and captures both spatial and temporal patterns in renewable generation. In addition, we show that different conditional information can be embedded in the framework. Because of the feedforward nature of the neural networks, scenarios can be generated extremely efficiently.

Conventional cameras record all light falling onto their sensor regardless of its source or its 3D path to the camera. In this talk I will present a emerging family of coded-exposure video cameras that can be programmed to record just a fraction of the light coming from an artificial source---be it a common street lamp or a programmable projector---based on the light path's geometry or timing. Live video from these cameras offers a very unconventional view of our everyday world in which refraction and scattering can be notice with the naked eye can become apparent, and the flicker of electric lights can be turned into a powerful cue for analyzing the electrical grid from room to city.

I will discuss the unique optical properties and power efficiency of these "transport aware cameras" through three case studies: the ACam for analyzing the electrical grid, EpiScan3D for robust progress toward designing a computational CMOS sensor for coded two-bucket imaging---a novel capability that promises much more flexible and powerful transport-aware cameras compared to existing off-the-shelf solutions.

Diffuse optical spectroscopy and imaging (DOSI) is a class of non-invasive near-infrared imaging techniques based upon measuring the wavelength-dependent absorption and (reduced) scattering optical properties of living tissues. In the far-red to near-infrared optical therapeutic window, these optical properties provide information about deep (several cm) tissue composition, structure, and oxygen metabolism. In particular, DOSI is capable of quantifying tissue concentrations of the physiologically relevant molecules oxyhemoglobin, deoxygenated hemoglobin, lipid, and water, as well as structural parameters including cellular size and density (obtained from scattering spectra). The significance and applicability of these and other DOSI biomarkers collected with research devices have been demonstrated in numerous clinical studies of oncology, cardiovascular assessment, exercise physiology, and neuroscience.

In this presentation, I will discuss how DOSI has shown promise in the field of breast oncology for risk assessment, screening, differential diagnosis of benign and malignant lesions, and predicting and monitoring response to chemotherapy treatment. DOSI biomarkers vary significantly in abundance and molecular state between breast cancer and normal tissue and unique cancer-specific absorption signatures have been observed. Finally, I will demonstrate how we are working to translate this promising technology to clinical practice and my vision for the future.