JRJR33, INC.
* Each winner receives mentorship and USD
QIF is an excellence award through which Qualcomm Technologies recognizes and mentors some of the most innovative engineering PhD students across
"Each year we are faced with the difficult decision of choosing between excellent research proposals that have the potential to positively impact the world" said
The twelve finalists are PhD candidates from
After careful review, the following four winners were selected for their outstanding proposals:
Massive Scale Bayesian Neural Networks with Sampling-Based Inference
(Machine Learning)
With machine learning models being deployed in safety-critical fields, many concerns have been raised regarding their safety, and ability to quantify uncertainty in predictions when applied to real data. It is therefore important to be able to recognise when these methods fail; an overly confident wrong decision can result in misleading scientific analysis, ethical violations, and even catastrophic consequences such as fatal accidents. Towards this goal, Shreyas's proposal investigates probabilistic inference techniques and specifically, Bayesian Neural Networks and Gaussian Processes. While these techniques have so far been shown successful primarily in low-dimensional scenarios, the proposal plans to scale them to high-dimensions to enable uncertainty-aware predictions in realistic scenarios.
A unifying perspective of causal and representational identifiability
(Machine Learning)
Tübingen AI Center
Representation learning with deep neural networks aims to convert complex information into a lower-dimensional and easier-to-handle format. For example, the trajectory of an autonomous vehicle from a video might be described by its position, velocity, and acceleration. However, it is difficult to provide guarantees for the representation, which is crucial for safety-critical systems like healthcare or transportation. Identifiable representation learning allows compatibility across different models by aligning representations through simple transformations. However, ensuring identifiability presents many open questions, such as identifying connections between causal relationships and identifiability and collecting appropriate data towards training the models. Patrik's proposal explores a recent insight on how identifiable representation can reveal causal graphs, thus enhancing the efficiency and reliability of interactive agents by bridging theory and practice.
Interaction-aware Learning of 3D Human Motions and Behaviors from Egocentric Views
(
ETH Zürich
AI-powered AR/VR devices are transforming sectors like healthcare, education, and entertainment by understanding surroundings and human actions. However, current techniques are primarily limited to third-person view applications and rarely consider social interactions. Siwei proposes to address these limitations by advancing fundamental research in egocentric human behavior understanding in an interaction-aware manner. The proposal plans to investigate synthesizing human-scene and social interactions and egocentric motion capture and reconstruction.
Effective Lifelong Adaptation of Vision-Language Foundation Models across Domains for Practical Deployment
(Machine Learning)
Tübingen AI Center
Open-vocabulary vision-language models are key for advancements in multimodal applications, but their usefulness diminishes as their training context becomes outdated. It is thus important to develop and understand mechanisms that allow us to continuously update these models without costly retraining. As a result, Karsten first proposes to benchmark current continual learning techniques and obtain insights on impact of various design choices considered in such techniques. The proposal then plans to leverage these insights and introduce extensions to improve performance of continual learning strategies. This work is expected to provide valuable insights into transferring continual learning research from small data and model regime to large-scale continuous adaptation in foundation models.
Learning Equivariances from Data
(Machine Learning)
Large neural networks are increasingly used to solve real-world problems outperforming more classical machine learning models on many tasks. Inductive biases, such as invariance and symmetries in these models, play a key role in their overall performance and generalization abilities. Although recent works have allowed extensions to various symmetry groups and domains, the neural network architectures and group structures they use are fixed and need to be selected manually or through expensive cross-validation. Tycho proposes to instead learn equivariances and corresponding neural structures from training data through a combination of flexible parameterizations and an amenable objective function capable of learning symmetries. Solving this problem would make finding the architecture as simple as learning weights and pave the way for the automatic discovery of compute/parameter efficient architectures.
Hearables: Interpretable real-time detection of arrhythmias via in-ear ECG/PPG continuous monitoring device
(Machine Learning)
Cardiovascular diseases (CVDs) are the number one cause of death globally. An Electrocardiogram (ECG) is a gold standard technique to diagnose possible heart-related diseases such as arrhythmias (ARs). However, ARs can go unnoticed in ECGs. To tackle this, in-ear "Hearables" continuously monitor heart signals but face issues like low signal-to-noise ratios and interpretation difficulty. Edoardo's proposal plans to address these limitations by developing novel AI/ML models to analyze in-ear ECG data and make them robust to noise and highlight the information used to diagnose abnormal heart signals. The goal is to minimize consultations, lower costs, and provide real-time warnings of potential heart rhythm abnormalities.
Image-based 6D Pose Estimation for Previously-Unseen Objects
(
EPF Lausanne
Object pose estimation is at the cornerstone of some real-world applications of computer vision and robotics, e.g., robotic manipulation and grasping, augmented reality, and autonomous driving. The objective is recognizing the object from the captured image and predicting the 6D object pose, i.e., 3D rotation and 3D translation. Current models assume the objects they encounter during deployment are the same as the ones used for training, limiting their practical use in environments with new objects. Chen aims to overcome this by creating an object pose estimation method that can adapt to new objects, even from unseen categories, without retraining, making it easier to deploy robotics systems in the real world.
Constructive Certification
(Machine Learning)
Machine learning (ML) models are almost as celebrated for their near-perfect accuracy as they are infamous for being unreliable. For example, autonomous vehicles can gracefully navigate busy intersections and yet in some cases fail to detect pedestrians or emergency vehicles. Ensuring these systems' safety in unfamiliar situations remains a major challenge in ML. While past research has focused on adversarial robustness, practical applications require prediction correctness at deployment. To this end, Aleksandar proposes a novel concept of a certificate, called constructive certificate, that offers prediction correctness guarantees for novel inputs at deployment time. A successful implementation will open new avenues for reliable ML research and enable certifiably correct safety-critical applications in robotics, autonomous driving, finance, and healthcare.
SecureCells: A Secure Compartmentalized Architecture
(Security Systems)
EPF Lausanne
In the world of software security, vulnerabilities like log4J and Heartbleed pose significant risks to the safety of our programs, particularly browsers. Existing protection methods have fallen short in meeting the demands of modern performance and security requirements. Atri proposes SecureCells to address this issue. By revolutionizing the way computer systems manage and protect their memory, SecureCells offers enhanced security for a wide range of programs, including those used on mobile devices, desktop computers, and servers. With minimal performance overhead, SecureCells ensures the safety of critical programs in today's rapidly evolving technological landscape.
Advanced Digital Signatures for Privacy in Electronic Payments
(Security Systems)
CISPA
Digital signatures play a crucial role in various applications, but their basic functionality may not be sufficient for modern needs. To address this, advanced signature variants with enhanced features are sought after, particularly to improve scalability and privacy in digital currencies. Benedikt's proposal aims to develop efficient constructions of these signature variants, building upon establish cryptographic assumptions. The goal is to identify the most effective way to incorporate these components into applications such as private digital payment systems. By enhancing digital signatures, this proposal can unlock new possibilities for secure and efficient transactions in the digital realm.
Analyzing the Architectural and Microarchitectural Security of the CPU Instruction Path
(Security Systems)
CISPA
Daniel's research proposal focuses on analysing the path that instructions follow inside modern CPUs - an area that has not received as much attention compared to the data path. The aim is to understand how vulnerabilities in certain components of the processor's instruction path can affect the CPU's overall security. The research will explore potential new ways that attacks could exploit these vulnerabilities. Findings from this study will contribute to a better understanding of CPU security and help in developing stronger protection mechanism against attacks that target the processors.
About Qualcomm
Qualcomm is enabling a world where everyone and everything can be intelligently connected. Our one technology roadmap allows us to efficiently scale the technologies that launched the mobile revolution - including advanced connectivity, high-performance, low-power compute, on-device intelligence and more - to the next generation of connected smart devices across industries. Innovations from Qualcomm and our family of Snapdragon platforms will help enable cloud-edge convergence, transform industries, accelerate the digital economy, and revolutionize how we experience the world, for the greater good.
.
(C) 2023 M2 COMMUNICATIONS, source
