Tuesday, June 28th, 2022

9:30 am10:00 am
Speaker: David Danks (University of California, San Diego)

Modern cognitive neuroscience often requires us to identify causal "objects" (perhaps spatial aggregates, perhaps more complex dynamic objects) that can function in our neuroscientific theories. Moreover, we often hope or require that these "objects" be neuroscientifically understandable (or plausible). Of course, the brain does not come neatly segmented or packaged into appropriate aggregates or objects; rather, these objects are themselves the product of scientific work, and which objects we get depend on the goals that we have. I will argue that different goals map onto different learning criteria, which then map onto different extant methods in cognitive neuroscience. The philosophical and technical challenge is that these different methods can yield incompatible outputs, particularly if we require interpretability, and so we seem to be led towards a problematic pluralism. I will conclude by considering several ways to try to avoid problematic inconsistencies and conflicts between our theories.

10:30 am11:00 am
Speaker: Sina Fazelpour (Northeastern University)

In many applications of societal concern, algorithmic outputs are not decisions, but only predictive recommendations provided to humans who ultimately make the decision. However, current technical and ethical analyses of such algorithmic systems tend to treat them as autonomous entities. In this talk, I draw on works in group dynamics and decision-making to show how viewing algorithmic systems as part of human-AI teams can change our understanding of key characteristics of these systems—with a focus on accuracy and interpretability—and any potential trade-off between them. I will discuss how this change of perspective (i) can guide the development of functional and behavioral measures for evaluating the success of interpretability methods and (ii) challenge existing ethical and policy proposals about the relative value of interpretability.

11:00 am12:00 pm
Speaker: David Danks (University of California, San Diego), and Sina Fazelpour (Northeastern University)

TBD

3:00 pm3:30 pm
Speaker: Alex D'Amour (Google Brain)

There has been a strong intuition in the Machine Learning community that interpretability and causality ought to have a strong connection. However, the community has not arrived at consensus about how to formalize this connection. In this talk, I will raise questions about conceptual and technical ambiguities that I think make this connection hard to specify. The goal of the talk is to raise points for discussion, expressed in causal formalism, rather than to provide answers.

4:00 pm4:30 pm
Speaker: Joe Halpern (Cornell University)

I consider a definition of (causal) explanation that is a variant of one Judea Pearl and I gave. The definition is based on the notion of actual cause. Essentially, an explanation is a fact that is not known for certain but, if found to be true, would constitute an actual cause of the fact to be explained, regardless of the agent's initial uncertainty. I show that the definition handles well a number of problematic examples from the literature, and discuss various notions of partial explanation.

4:30 pm5:30 pm
Speaker: Alex D'Amour (Google Brain), and Joe Halpern (Cornell University)

TBD

Wednesday, June 29th, 2022

9:30 am10:00 am
Speaker: Shirley Ho (Flatiron Institute/ NYU)
TBD
10:00 am10:30 am
Speaker: Michele Ceriotti (Swiss Federal Institute of Technology in Lausanne) [REMOTE]

TBD

11:00 am11:30 am
Speaker: David Limmer (University of California, Berkeley)

In this talk, I will discuss the connections between physical nonequilibrium systems and common algorithms employed in machine learning. I will report how machine learning has been used to expand the scope of physical nonequilibrium systems that can be effectively studied computationally. The interpretation of the optimization procedure as a nonequilibrium dynamics will be also examined. Specific examples in reinforcement learning will be highlighted.

11:30 am12:30 pm
Speaker: Shirley Ho (Flatiron Institute/ NYU), Michele Ceriotti (Swiss Federal Institute of Technology in Lausanne) [REMOTE], and David Limmer (University of California, Berkeley)

TBD

2:30 pm3:00 pm
Speaker: Debbie Marks (Harvard University)

There’s now an amazing opportunity to accelerate discovery across important 21st century challenges by using computation tightly coupled to biological experiments and clinical medicine. I will describe some recent approaches from my lab for these challenges where we have developed new machine learning methods that can exploit the enormous natural sequence diversity and our ability to synthesize DNA at scale.  To demonstrate the power of these new approaches I will present recent work predicting the effects of human genetic variation on disease, anticipation of viral escape from the host immune system for vaccine design,  protein design for enzyme optimization and smart antibody libraries.

3:00 pm3:30 pm
Speaker: Anshul Kundaje (Stanford University) [REMOTE]

The human genome sequence contains the fundamental code that defines the identity and function of all the cell types and tissues in the human body. Genes are functional sequence units that encode for proteins. But they account for just about 2% of the 3 billion long human genome sequence. What does the rest of the genome encode? How is gene activity controlled in each cell type? Where do the gene control units lie in the genome and what is their sequence code? How do variants and mutations in the genome sequence affect cellular function and disease? Regulatory instructions for controlling gene activity are encoded in the DNA sequence of millions of cell type specific regulatory DNA elements in the form of functional sequence syntax. This regulatory code has remained largely elusive despite exciting developments in experimental techniques to profile molecular properties of regulatory DNA. To address this challenge, we have developed high performance neural networks that can learn de-novo representations of regulatory DNA sequence to map genome-wide molecular profiles of protein DNA interactions and biochemical activity at single base resolution across 1000s of cellular contexts while accounting for experimental biases. We have developed methods to interpret DNA sequences through the lens of the models and extract local and global predictive syntactic patterns revealing many causal insights into the regulatory code. Our models also serve as in-silico oracles to predict the effects of natural and disease-associated genetic variation i.e. how differences in DNA sequence across healthy and diseased individuals are likely to affect molecular mechanisms associated with common and rare diseases. Our predictive models serve as an interpretable lens for genomic discovery.

4:00 pm4:30 pm
Speaker: Jack Gallant (University of California, Berkeley)

The mammalian brain is an extremely complicated, dynamical deep network. Systems, cognitive and computational neuroscientists seek to understand how information is represented throughout this network, and how these representations are modulated by attention and learning. Machine learning provides many tools useful for analyzing brain data recorded in neuroimaging, neurophysiology and optical imaging experiments. For example, deep neural networks trained to perform complex tasks can be used as a source of features for data analysis, or they can be trained directly to model complex data sets. Although artificial deep networks can produce complex models that accurately predict brain responses under complex conditions, the resulting models are notoriously difficult to interpret. This limits the utility of deep networks for neuroscience, where interpretation is often prized over absolute prediction accuracy. In this talk I will review two approaches that can be used to maximize interpretability of artificial deep networks and other machine learning tools when applied to brain data. The first approach is to use deep networks as a source of features for regression-based modeling. The second is to use deep learning infrastructure to construct sophisticated computational models of brain data. Both these approaches provide a means to produce high-dimensional quantitative models of brain data recorded under complex naturalistic conditions, while maximizing interpretability.

4:30 pm5:30 pm
Speaker: Debbie Marks (Harvard University), Anshul Kundaje (Stanford University) [REMOTE], and Jack Gallant (University of California, Berkeley)

TBD

Thursday, June 30th, 2022

9:30 am10:00 am
Speaker: Alice Xiang (Sony AI) [REMOTE]

Interpretability is a key component of many dimensions of building more trustworthy ML systems. In this talk, I will focus on a couple intersections between interpretability and algorithmic fairness. First, I will discuss some of the promises and challenges of causality for diagnosing sources of algorithmic bias. In particular, defining the nature and timing of interventions on immutable characteristics is highly important for appropriate causal inference but can create challenges in practice given data limitations. Second, I will discuss the strategy of collecting more diverse datasets for alleviating biases in computer vision models. Defining and measuring diversity of human appearance remains a significant challenge, especially given privacy concerns around sensitive attribute labels. To address this, I will present a method for learning interpretable dimensions of human diversity from unlabeled datasets.

10:30 am11:00 am
Speaker: Rebecca Wexler (Berkeley Center for Law & Technology) [REMOTE]

Making machine learning systems interpretable can require access to information and resources. Whether that means access to data, to models, to executable programs, to research licenses, to validation studies, or more, various legal doctrines can sometimes get in the way. This talk will explain how intellectual property laws, privacy laws, and contract laws can block the access needed to implement interpretable machine learning, and will suggest avenues for reform to minimize these legal barriers.

11:00 am11:30 am
Speaker: Aleksandra Korolova (University of Southern California) [REMOTE]

Relevance estimators are algorithms used by social media platforms to determine what content is shown to users and its presentation order. These algorithms aim to personalize the platform's experience for users, increasing engagement and, therefore, platform revenue. However, many have concerns that the relevance estimation and personalization algorithms are opaque and can produce outcomes that are harmful to individuals or society. Legislations have been proposed in both the U.S. and the E.U. that mandate auditing of social media algorithms by external researchers. But auditing at scale risks disclosure of users' private data and platforms' proprietary algorithms, and thus far there has been no concrete technical proposal that can provide such auditing. We propose a new method for platform-supported auditing that can meet the goals of the proposed legislations

11:30 am12:30 pm
Speaker: Alice Xiang (Sony AI), Aleksandra Korolova (University of Southern California), and Rebecca Wexler (Berkeley Center for Law & Technology) [REMOTE]

TBD