Bayesian Neural Network is one of the major approaches for obtaining uncertainty estimates for deep learning models. Key to the success is the selection of the approximate inference algorithms used to compute the approximate posterior, with mean-field variational inference (MFVI) and MC-dropout being the most popular variants. But is the good downstream uncertainty estimation performance of BNNs attributed to good approximate inference? In this talk I will discuss some of our recent results towards answer this question. I will also discuss briefly the computational reasons of the preference of MFVI/MC-dropout and describe our latest work to make BNNs more memory efficient.

Passcode: 667130

Abstract: Neural Machine Translation (NMT) has seen a tremendous spurt of growth in less than ten years, and has already entered a mature phase. While considered as the most widely used solution for Machine Translation, its performance on low-resource language pairs still remains sub-optimal compared to the high-resource counterparts, due to the unavailability of large parallel corpora. Therefore, the implementation of NMT techniques for low-resource language pairs has been receiving the spotlight in the recent NMT research arena, thus leading to a substantial amount of research reported on this topic. This paper presents a detailed survey of research advancements in low-resource language NMT (LRL-NMT), along with a quantitative analysis aimed at identifying the most popular solutions. Based on our findings from reviewing previous work, this survey paper provides a set of guidelines to select the possible NMT technique for a given LRL data setting. It also presents a holistic view of the LRL-NMT research landscape and provides a list of recommendations to further enhance the research efforts on LRL-NMT.

In this session, we will introduce some of the most relevant computer vision problems, and explore them through the lens of unsupervised learning to highlight when and how it can be beneficial.

Precision medicine in oncology aims to find a targeted treatment for each cancer patient sub-type or individual patient. This session will discuss how leveraging data-driven machine learning methods for integrating high-dimensional multi-omics cancer datasets and modelling their interplay can lead to improved risk assessment, early detection, and tailored cancer therapies for cancer patients.

Despite the recent attention to DeepFakes and other forms of image manipulations, one of the most prevalent ways to mislead audiences on social media is the use of unaltered images in a new but false context, commonly known as out-of-context image use. The danger of out-of-context images is that little technical expertise is required, as one can simply take an image from a different event and create a highly convincing but potentially misleading message. At the same time, it is extremely challenging to detect misinformation based on out-of-context images given that the visual content by itself is not manipulated; only the image-text combination creates misleading or false information. In order to detect these out-of-context images, several online fact-checking initiatives have been launched by newsrooms. However, they all heavily rely on manual human efforts to verify each post factually and to determine if a fact-checking claim should be labeled as "out-of-context". In this talk, I will discuss how can we build models that help determine the conflicting image-caption pairs and could be potential out-of-context misuse.

In this breakout session, we will discuss the motivation behind contrastive learning and various recent contrastive models. We will discuss the intuition behind when and why contrastive learning works and how to improve its performance. Finally, we will discuss some tricks for successful application of contrastive learning in vision

A discussion on information leakage in machine learning models and concerns in everyday life

Recent years have seen an overwhelming body of work on fairness and robustness in Machine Learning (ML) models. This is not unexpected, as it is an increasingly important concern as ML models are used to support decision-making in high-stakes applications such as mortgage lending, hiring, and diagnosis in healthcare. Currently, most ML models assume ideal conditions and rely on the assumption that test/clinical data comes from the same distribution of the training samples. However, this assumption is not satisfied in most real-world applications; in a clinical setting, we can find different hardware devices, diverse patient populations, or samples from unknown medical conditions. On the other hand, we need to assess potential disparities in outcomes that can be translated and deepen in our ML solutions. In this presentation, we will discuss how to evaluate skin-tone representation in ML solutions for dermatology and how we can enhance the existing models’ robustness by detecting out-out-distribution test samples (e.g., new clinical protocols or unknown disease types) over off-the-shelf ML models.

This un-workshop breakout session will discuss challenges and opportunities in ML for Health Care with a focus on interpretability in clinical decision making, the automation of clinical tasks, and fairness and bias concerns.

We will discuss the present tools, challenges & future of explainability research.

Consequential decisions compel individuals to strategically react in response to the specifics of the decision rule. In aggregate, these individual-level responses can disrupt the statistical patterns that motivated the decision rule, leading to unforeseen consequences. In this breakout session, we will give an introduction to the area of strategic classification, with a focus on limitations of the standard models of agent behavior. We will then have a discussion about applications where strategic adaptation is observed, potential approaches for modeling strategic agents, and connections to other domains.

Natural language processing (NLP), or the task of turning human language into structured data for machines to understand, powers the chatbots, voice assistants, predictive text, and other pervasive language applications that make our lives easier. We feel open-source tools complement the rapidly evolving field of NLP by helping to drive innovation through collaboration. In this breakout session, participants will be invited to share their experiences and lessons learned using open-source tools for NLP and explore how open-source tools may be able to help with current challenges in NLP.

This session will provide an opportunity to explore tools and techniques to assist the next generation of Black women as they pursue STEM careers and STEM studies.

Even though it is rarely emphasized, having a supportive environment is crucial to one’s research success. While in a lot of cases we don’t really have 100% control over what lab environment, what peers, what advisors we end up finding ourselves in, there are still things we can do to cultivate an environment, a community, even a culture, that works for us.

To-date, a discussion around the relative merits of different compression methods has centered on the trade-off between level of compression and top-line metrics such as top-1 and top-5 accuracy. Along this dimension, compression techniques such as pruning and quantization are remarkably successful. It is possible to prune or heavily quantize with negligible decreases to test-set accuracy. However, top-line metrics obscure critical differences in generalization between compressed and non-compressed networks. In this talk, we will go beyond test-set accuracy and discuss some of my recent work measuring the trade-offs between compression, robustness and algorithmic bias. Characterizing these trade-offs provide insight into how capacity is used in deep neural networks — the majority of parameters are used to represent a small fraction of the training set. Formal auditing tools like Compression Identified Exemplars (CIE) also catalyze progress in training models that mitigate some of the trade-offs incurred by compression.

With the emergence of publicly shareable repositories on the Web (i.e. open sourced code in GitHub), there are now billions of lines of code easily to read and learn from. With the abundance of available code artifacts, researchers have begun to apply machine learning techniques to find statistical patterns in Big Code to guide developers to understand and write code better. This breakout session will cover the applications of Big Code, including program synthesis, code search, code autocomplete, code bug fixing.

Humans represent complex systems as compositions of entities and interactions among these entities, i.e., a graph. In this session, we particularly focus on the algorithmic and theoretical foundations of Graph Neural Networks (GNNs). GNNs are a family of methods that generalize and extend neural networks to operate on relational data, and provides a flexible interface for manipulating structured knowledge, learning structured representations, and relational reasoning. Despite the rapid growth in the past five years, there are several limitations to applying and generalizing current GNN techniques to model datasets in novel applications, e.g., computer vision, graphics, wireless communication, etc.”

Collaborative grounding is a fundamental aspect of human-human dialogue which allows people to negotiate meaning; in this talk, I argue that current deep learning approaches to dialogue systems don’t deal with it adequately. Making errors, and being able to recover from them collaboratively, is a key ingredient in grounding meaning, but current dialogue systems can’t do this. I will illustrate the pitfalls of being unable to ground collaboratively, discuss what can be learned from the language acquisition and dialog systems literature, and reflect on how to move forward. I will urge the community to proceed by addressing a research gap: how clarification mechanisms can be learned from data. Novel research methodologies which highlight the importance of the role of clarification mechanisms are needed for this. I will present an annotation methodology, based on a theoretical analysis of clarification requests, which unifies a number of previous accounts. Dialogue clarification mechanisms are an understudied research problem and a key missing piece in the giant jigsaw puzzle of natural language understanding. I will conclude this talk with an open call for collaborators that share the vision presented.

ML methods across the discipline make a variety of approximations in order to scale. We will discuss how to navigate these trade-offs appropriately, particularly in the context of high-impact ML applications.

The first half of the panel will be focused on discussions on “industry/academic research, how to choose your path” and “post-pandemic adjustment and tips”. The second half will be an "ask me anything" session, where the panelists will be answering questions from the participants.

Panelists - Sarah Aerni (Salesforce Einstein), Kalesha Bullard (Facebook), Sarah Dean (Cornell University), Sylvia Herbert (University of California, San Diego), and Amy Zhang (Facebook).