Seminars

Watch the recording here. 

Watch the recording here.

Watch the recording here.

Watch the recording here.

Watch the recording here.

Watch the recording here.

Here, we show that two DNNs trained on non-overlapping subsets of a dataset learn nearly the same score function, and thus the same density, when the number of training images is large enough. 

Watch the recording here.

Watch the recording here.

Find more details here.

You can find the recording here.

Find more details here

You can find the recording here.

This talk will give an overview of a range of emerging challenges at the intersection of AI and law, including some ongoing copyright lawsuits, legal controversies around algorithmic decision making, and a brief overview of major regulatory efforts. (The talk will be aimed at computer scientists; no law background is required.)

In this talk, Prof. Fouhey will show off the systems that he built with members of his research group and collaborators. He will focus primarily on solar physics, where they have developed a series of systems with a primary focus on obtaining better measurements of the Sun's vector magnetic field.

Click here to view the recording.

In this talk Nur Muhammad “Mahi” Shafiullah will talk on the topic "On Bringing Robots Home" and Ulyana Piterbarg will talk on the topic "NetHack is Hard to Hack".

Click here to view the recording

In this talk, Nate Gurver will discuss two of his recent projects on applying generative sequence models to atomistic data, in particular protein design and discovery of stable inorganic materials. Taro Makino will talk about Asymmetric Prior Variational Autoencoder (AP-VAE)

Click here to view the recording

This talk will discuss the massive shift that has come about in the vision and ML community as a result of the large pre-trained language and language and vision models such as Flamingo, GPT-4, and other models.

Click here to view the recording.

In this talk, we will dive into self-supervised learning (SSL) and how to generalise ideas developed for images to other domains.

Click here to view the recording

Subgrid parameterizations, which represent physical processes occurring below the resolution of current climate models, are an important component in producing accurate, long-term predictions for the climate. In this talk Karl will present this research problem and discuss an ongoing project investigating a multiscale, deep learning-based approach to this task. Lei will introduce some of his ongoing research on optimization instability and implicit bias.

Sungmin will introduce some of his past research on continual learning using neural networks. To do so, he will briefly explain the settings of continual learning research and then discuss his research findings related to the novel loss function, layer, and applications for continual learning. Finally, he will briefly discuss the limitations and future works.

Click here to view the recording

Language technology tools such as Google Translate or virtual assistants (Siri, Alexa) were components of collective SciFi-inspired imagination not many years ago. Today, they are an essential driver of the digital AI transformation, used by hundreds of millions of people. I will introduce a range of recent techniques, breakthroughs and lessons learned that aim to deal with large cross-language variations and low-data learning regimes. 

I will introduce my previous efforts extending the scope of meta-learning towards real-world learning scenarios. We will see how to deal with realistic task distributions including class/task imbalance and distributional shift, discuss how to develop practical meta-knowledge applicable to arbitrary CNN architectures, and also how to scale up few-shot learning to many-shot and heterogeneous learning problems.

We have seen machine learning making great strides in understanding visual scenes. Yet, most of its success relies on training models on a massive amount of data in a closed world and evaluating them in a similar environment. I present an alternative paradigm that will allow machines to acquire visual knowledge through an online stream of data in an open world, which entails abilities such as learning visual representations and concepts efficiently with limited and non-iid data. 

Pre-trained language models have fundamentally changed the landscape of NLP.  However, as the pre-trained language models are becoming increasingly large, we have also witnessed that the gain in their generalization performance is becoming marginal, especially when we only have limited labelled data in downstream tasks. To improve their generalization, we propose a new framework for fine-tuning of pretrained models to yield better generalization performance. 

Deep-learning models for image processing achieve impressive results when trained on standard natural-image datasets in a supervised fashion. However, unleashing their potential in practice will require developing unsupervised or semi-supervised approaches capable of learning from real data, as well as understanding the strategies learned by these models. In this talk, we will describe advances in this direction.

I recently proposed the lottery ticket hypothesis: that the dense neural networks we typically train have much smaller subnetworks capable of reaching full accuracy from early in training. In this talk, I will discuss established results and the latest developments in my line of work.

Prescient Design at Genentech investigates, develops and deploys a machine learning driven platform for protein design. At the core of Prescient Design’s strategy is the deep manifold sampler which is a non-autoregressive sequence denoising autoencoder combined with a function predictor. In this talk, I will describe the foundations behind the deep manifold sampler. 

Supporting a safe and trustworthy online environment is challenging, as these environments are constantly threatened by abusive behaviors that cause real human harm.  I will present my work on characterizing threats, and empowering users with new techniques to combat these threats.

Jiwoong Im: Causal Effect Variational Autoencoder with Uniform Treatment

Aahlad Puli: Predictive Modeling in the Presence of Nuisance-Induced Spurious Correlations

While DNNs have achieved many breakthroughs, our understanding of their internal structure, optimization process, and generalization is poor, and we often treat them as black boxes. We attempt to resolve these issues by suggesting that DNNs learn to optimize the Information Bottleneck (IB) principle - the tradeoff between information compression and prediction quality.
 

Current AI systems have limited agency in the world, i.e., they lack the embodiment and sensory experience of the world. Hopefully, embodied AI systems will address these issues, such as autonomous vehicles mobilizing the visually impaired in our communities. The manifestation of AI in the physical world necessitates research into linking natural language to physical referents. The central theme of this talk is to understand and model this linking process, namely language grounding. 

Unlikelihood training and the consistency of distributions induced by language models and decoding algorithms.

1. Min Jae: On the Cryptographic Hardness of Learning Neural Networks.
2. Mahi Shaffiullah:  Surprising Effectiveness of Representation Learning for Visual Imitation.
3. Ben Evans: Context is Everything: Implicit Identification for Dynamics Adaptation.
4. Nan Wu: Joint speech and text pre-training for text normalization towards end-to-end TTS.

Tal Linzen, Sam Bowman, Lerrel Pinto, and Jason Weston (Meta) will present their recent research.