2019 was an interesting year in so many ways. I was able to build and solidify research programs, both at NVIDIA and at Caltech. I was able to continue working towards diversity and inclusion in AI, and saw a lot of visible improvements (recent incident I wrote in my previous blog post is a notable exception). Overall, there is a lot of positivity and a great way to end an eventful decade!
Before I list research highlights that I was personally involved in, here’s an overall summary for the AI field from my viewpoint. This was published on KDnuggets.
In 2019, researchers aimed to develop a better understanding of deep learning, its generalization properties, and its failure cases. Reducing dependence on labeled data was a key focus, and methods like self-training gained ground. Simulations became more relevant for AI training and more realistic in visual domains such as autonomous driving and robot learning, including on NVIDIA platforms such as DriveSIM and Isaac. Language models went big, e.g. NVIDIA’s 8 billion Megatron model trained on 512 GPUs, and started producing coherent paragraphs. However, researchers showed spurious correlations and undesirable societal biases in these models. AI regulation went mainstream with many prominent politicians voicing their support for ban of face recognition by Governmental agencies. AI conferences started enforcing a code of conduct and increased their efforts to improve diversity and inclusion, starting with the NeurIPS name change last year. In the coming year, I predict that there will be new algorithmic developments and not just superficial application of deep learning. This will especially impact “AI for science” in many areas such as physics, chemistry, material sciences and biology.
New book on Spectral Learning on Matrices on Tensors
Builds up spectral methods from first principles. Applications to learning latent variable models and deep neural networks. Order your copy here.
Better Optimization Methods
Fixing training in GANs through competitive gradient descent
(Excellent blog post by Florian here). In contrast to standard simultaneous gradient updates, CGD guarantees convergence and is efficient. NeurIPS poster below:
Application of CGD to GAN training and demonstrating its implicit competitive regularization (NeurIPS workshop):
Guaranteed convergence for SignSGD
SignSGD compresses gradient to a single bit but has no significant loss in accuracy in practice. Theoretically, there are convergence guarantees. Paper. Main theorem:
Exciting collaboration between ML and neuroscience with Doris Tsao at Caltech. Adding feedback generative model to convolutional neural networks significantly improves robustness in tasks such as unsupervised denoising. Short paper here.
Robust Learning in Control Systems
Robust regression for safe exploration
We address the following: How to extrapolate robustly from your training data in real world control tasks and achieve end to end stability guarantees in safe exploration? Paper.
Multi-modal learning for UAV navigation
Multi-modal fusion of vision and IMU improves robustness in navigation and landing. Paper.
Generalization in ML
Detecting hard examples through an angular measure
Watch my GTC-DC talk. Angular alignment is a robust measure for hardness: easier examples align more with the target class. We found that correspondence between angular measure and human selection frequency was statistically significant. Improves self training in domain adaptation. Paper.
Regularized learning for domain adaptation
New domain adaptation algorithm to correct for label shifts. Paper
Our ability to fix bias in deep-learning algorithms
Twitter thread here
Recursive neural networks with external memory
Recursive networks have compositionality and can extrapolate to unseen instances.
Extrapolation to harder instances (higher tree depth) is challenging.
We show that augmenting with external memory stacks significantly improves extrapolation. Paper
Open Vocabulary Learning on Source Code with a Graph–Structured Cache
Use of syntax trees in program code to handle unbounded vocabulary. Paper
Reinforcement Learning and Bandits
Robust off-policy evaluation
Robust methods to handle covariate shift in off-policy evaluation. PaperStochastic Linear Bandits with Hidden Low Rank Structure
Low regret algorithms for discovering hidden low rank structures in bandits. Paper
Competitive Differentiation for Lagrangian Problems
Many RL problems have constraints leading to Lagrangian formulation. Paper
Context-based Meta RL with Structured Latent Space
AI4science and Other Applications
Neural ODEs for Turbulence Forecasting
Turbulence modeling is a notoriously hard problem. Exciting initial results presented at NeurIPS workshop.
Tackling Trolling through Dynamic Keyword Selection Methods
Can we find social media trolls and create more trustworthy environment online? We presented our ongoing study of the #meToo movement, in collaboration with Michael Alvarez from Social sciences at Caltech.