AI Lunch Series Public Events

October 13, 2021

“Learning Causality with Graphs”

Dr. Jundong Li

Abstract

"The ability to learn causality is considered as a significant component of human-level intelligence and can serve as the foundation of AI. In causality learning, one fundamental problem is to understand the causal effects of a specific treatment (e.g., prescription of medicine) on an important outcome (e.g., cure of a disease), with significant implications in various high-impact domains such as health care, education, and e-commerce. One prevalent way to solve the problem is to directly use the observational data since the alternative randomized experiments could be expensive, time-consuming, and even unethical in many scenarios. However, existing data-driven methods are often limited since they: (1) assume that observational data is independent and identically distributed (i.i.d.); and (2) ignore the influence of hidden confounders (i.e., the unobserved variables that affect both the treatment and the outcome). Meanwhile, real-world data is often connected and can be abstracted as graphs (e.g., social networks, biological networks, and knowledge graphs). The ubiquitous of graph data across many influential areas also brings opportunities to control the influence of hidden confounders and build more effective models that yield unbiased causal effects estimation. In this talk, I will introduce our recent research efforts in causal effects learning with graphs. Specifically, we attempt to answer the following research questions: How to utilize graph information among observational data for causal effects learning? How to harness the power of historical information to tame the influence hidden confounders for causal effects learning when the graph is continuously evolving?"

Join with Bluejeans:  https://bluejeans.com/786906712/0315

June 30, 2021

"Extracting the Most from Collider Data with Deep Learning"

Dr. Ben Nachman (LBL)

Abstract

“Precise scientific analysis in collider-based particle physics is possible because of complex simulations that connect fundamental theories to observable quantities. These simulations have been paired with multivariate methods for many years in search of the smallest distance scales in nature.  Deep learning tools hold great promise to qualitatively change this paradigm by allowing for holistic analysis of data in its natural hyperdimensionality with thousands or millions of features instead of up to tens of features. These tools are not yet broadly used for all areas of data analysis because of the traditional dependence on simulations. In this talk, I will discuss how we can change this paradigm in order to exploit the new features of deep learning to explore nature at sub-nuclear distance scales. In particular, I will show how neural networks can be used to (1) overcome the challenge of intractable hypervariate probability density modeling and (2) learn directly from (unlabeled) data to perform hypothesis tests that go beyond any existing analysis methods. The example for (1) will be full phase space unfolding and the example for (2) will be anomaly detection. The talk will include a discussion of uncertainties associated with deep learning-based analyses.”

Join with Bluejeans:https://bluejeans.com/451415148/8427

June 2, 2021

“Graph Neural Networks and Metric Learning for the Future of Particle Physics”

Dr. Daniel Murnane (LBL)

Abstract

"The discovery of the Higgs boson at the Large Hadron Collider was a colossal effort that brought together many advanced experimental, theoretical and computational techniques. We are opening a door to a new generation of particle physics experiments that will try to answer questions about dark matter, dark energy and why the universe seems to be so perfectly tuned for complex structures, like life. Many Higgs-era techniques may not be powerful enough for these new energy/intensity scales, and so we turn to machine learning (ML) techniques to help answer these questions. It turns out that not only does ML boost our ability to discover new physics, it also challenges us to think about physics problems in new ways."

Join with Bluejeans: https://bluejeans.com/260264787

May 05, 2021

"Interpretable Machine Learning for Physics, using Symbolic Regression and Graph Neural Networks"

Miles Cranmer

Abstract

“In this talk I will argue two points. 1) Symbolic regression, a machine learning technique that fits data by iteratively searching the space of all possible analytic equations, should be a standard machine learning algorithm in physics. 2) Symbolic regression can be extended to high-dimensional spaces, such as to models for N-body simulations, using the method we have developed. To begin the talk, I will introduce symbolic regression (SR), which is a relatively old but underdeveloped technique. I will demonstrate our new high-performance open-source SR software "PySR". I will then discuss our research contribution: a technique for extending SR to high-dimensional spaces by the use of a neural network. I will focus on Graph Neural Networks (GNNs), which are a physically motivated neural network architecture, and very relevant to many physical and astrophysical problems. As a validation, I will demonstrate that we can use this technique of a GNN to SR to extract the correct force laws and Hamiltonians from simple particle simulations. I will then show work applying our method to the Quijote dark matter simulations – where it finds a simple analytic formula to estimate the overdensity of dark matter in a halo using learned metrics which summarize the nearby halos, giving a functional form that is more accurate than hand-derived formulas. Our approach simultaneously offers an interpretable ML technique for high-dimensional astrophysical data, and also a way of interpreting neural networks. (Demo code is available here).

Join with Bluejeans: https://bluejeans.com/500985640

February 24, 2021

CTwin - Developing a Transportation/Mobility Digital Twin for Chattanooga

Dr. Jibo Sanyal (ORNL)

Abstract

"The Computational Urban Sciences Group at ORNL, in partnership with NREL and several external stakeholders, have stood up a real-time digital twin focused on mobility for Chattanooga. The system has brought in 500+ real-time data feeds from 5 systems across 3 institutions, with at least 40 other secondary data sets. This has created an unprecedented opportunity to observe, anticipate, and orchestrate cyber-physical controls towards a 20% energy savings objective for the region. This talk focuses on the experiences in obtaining real-time data from location enabled sensors from infrastructure and how it is heralding a fundamental shift in how we interact with our urban environment."

Join with Bluejeans: https://bluejeans.com/651785849

February 17, 2021

“Deeply Uncertain: Comparing Methods of Uncertainty Quantification in Deep Learning Algorithms”

Dr. João Caldeira (Google)

Abstract

"In recent years, many methods for quantifying uncertainty in the output of deep learning algorithms have been developed. In this talk we will introduce three of the most prominent methods - Bayesian Neural Networks (BNN), Concrete Dropout (CD), and Deep Ensembles (DE) - and compare them to the standard analytic error propagation in the context of a single pendulum. We will also relate the output of these methods to terms more familiar in the context of the physical sciences. Our results highlight some pitfalls that may occur when using these UQ methods. For example, when the variation of noise in the training set is small, all methods predicted the same relative uncertainty independently of the inputs. This issue is particularly hard to avoid in BNN. On the other hand, when the test set contains samples far from the training distribution, we found that no methods sufficiently increased the uncertainties associated to their predictions. This problem was particularly clear for CD. In light of these results, we make some recommendations for usage and interpretation of UQ methods."

Join with Bluejeans: https://bluejeans.com/950395297  

February 10, 2021

“Machine learning-based Design and Control of Particle Accelerators”

Dr. Adi Hanuka (SLAC)

Abstract

"Machine learning has been used in various ways to improve accelerator operation including advanced optimization of accelerator operating configurations, development of virtual diagnostics to ’measure’ beam parameters, and prognostics to detect anomalies and predict failures. In this talk I'll review machine learning techniques currently being investigated at particle accelerator facilities, with specific emphasis on physics-informed active-learning research efforts."

Join with Bluejeans: https://bluejeans.com/230468745

January 20, 2021

Deep Learning on Graphs: Methods and Applications”

Dr. Lingfei Wu (IBM Research)

Abstract

"Recent years have seen a significantly growing amount of interest in graph neural networks (GNNs), especially on efforts devoted to developing more effective GNNs for node classification, graph classification, and graph generation. However, there are relatively fewer studies on other important topics such as deep graph learning, graph-based encoder-decoder, and deep graph matching. In the first part of the talk, I will introduce the basics of deep learning on graphs. In the second part of the talk,  I will introduce an end-to-end graph learning framework, namely Iterative Deep Graph Learning (IDGL), for jointly learning graph structure and graph embeddings simultaneously. In the third part of the talk, I will introduce a Graph2Seq neural network framework and then talk about how to apply this model in various NLP tasks."

Join with Bluejeans: https://bluejeans.com/981665133 

November 18, 2020

“What Entropy and Impedance Mean in Data Science”

Dr. Pete Alonzi (UVA)

Abstract

"Today every grant proposal seems to demand "interdisciplinarity". This talk is about how to think interdisciplinarialy between Physics and Data Science. The fundamental pieces of Data Science thinking will be laid out starting from the 4+1 model developed by Alvarado. The universal pipeline theory will be presented along with how the concepts of Entropy and Impedance are used to understand the scope of any given pipeline. We will conclude with a discussion of how using these Data Science concepts overlapped with Physics investigation can produce enhanced results."

Join with Bluejeans: https://bluejeans.com/990782096 

October 28, 2020

"Discussion on Continual Learning and Learning Dynamics for Deep Networks"

Dr. Razvan Pascanu

Abstract

I will begin with a brief introduction to, and open problems in, understanding neural networks and discuss several approaches to this problem. Further I will discuss the importance of continual learning for the field. Continual Learning, sometimes called life-long learning, never-ending learning or open-ended learning, is a topic that has been growing in popularity in the last few years. I will provide an overview of the problem, contrasting sequential learning to multi-task learning and highlight why multi-task learning can be in practice sub-optimal. I will discuss the main family of approaches to continual learning and discuss recent research in the field. Finally I will provide a few ideas of where the field might go in the near future and open the floor for discussion.

Join with Bluejeans: https://bluejeans.com/395185997 

September 30, 2020

"Who is Building AI and Who is AI Being Built For?"

Dr. Emily Denton

Abstract

AI systems are currently being deployed in a range of socially consequential domains, including policing, healthcare, employment, and more. Yet, despite the narrative of progress that frequently surround the development and integration of AI technologies, the benefits and risks of these systems are highly unequally distributed. This talk will provide an overview of the ways AI systems are impacting society, and failing marginalized communities, and the steps practitioners can take to mitigate potential harms.

Join with Bluejeans: https://bluejeans.com/130442674 

September 2, 2020

"Situating AI on the Road from Data Sharing to Societal Impact"

Dr. Daniel Mietchen

Abstract

By its very nature, Artificial Intelligence depends on the availability of data at scale. In this presentation we will look at a range of factors that influence the nature and scale of data sharing, from open science to disasters, from research infrastructures to ethics, from cooperation to competition. We will then delve into how these factors affect the data life cycle and the research cycle and explore how data sharing (or the lack of it) translates into societal impact. On that journey, we will watch out for ways in which AI can and does contribute to or benefit from the sharing of data and associated resources (or not), which will then form the basis of our discussion.

Join with Bluejeans: https://bluejeans.com/215807327  

July 29, 2020

“The Convergence of HPC and AI: New Methods for Computer Modeling to Address Grand Challenges in Science”

Tom Gibbs

Abstract

"The talk introduces the challenge with conventional modeling to address Grand Challenge problems as Moore’s Law is coming to an end, and the opportunity that is being demonstrated with early examples using Converged HPC and AI Methods. Conventional ModSim methods introduce a constraint with cost/accuracy as model size and fidelity increase, and the end of Moore’s Law further imposes a hard limit on raw performance. GPU accelerators have been shown to extend the raw compute speed by an order of magnitude, but cost/accuracy constraint with conventional methods become a hard limit with Grand Challenge model sizes. A taxonomy is emerging for the application of HPC and AI methods to address science problems. These include methods to dramatically improve the scale, accuracy and time to solution with converged modeling and simulation. The converged methods are also being applied to experimental data processing where detection accuracy is improved dramatically, and indicate the potential to introduce real time control to future experiment runs. Examples are given for Grand Challenge problems from multiple domains including fusion, particle physics, weather and climate, biochemistry, astrophysics, and material science."

Join with Bluejeans: https://bluejeans.com/707697867

July 22, 2020

“Neural Networks Love to Cheat: Shortcut Learning in Deep Learning and Beyond”

Robert Geirhos

Abstract

"Deep learning has triggered the current rise of artificial intelligence and is the workhorse of today's machine intelligence. Numerous success stories have rapidly spread all over science, industry and society, but its limitations have only recently come into focus. In this talk I will seek to distill how many of deep learning's problem can be seen as different symptoms of the same underlying problem: shortcut learning. Shortcuts are decision rules that perform well on standard benchmarks but fail to transfer to more challenging testing conditions, such as real-world scenarios. Related issues are known in Comparative Psychology and Education, suggesting that shortcut learning may be a common characteristic of learning systems, biological and artificial alike. Based on these observations, I propose a set of recommendations for model interpretation and benchmarking."

Join with Bluejeans: https://bluejeans.com/707697867

June 17, 2020

“Machine Learning Approaches for Enabling Robots to Handle Non-repetitive Manufacturing Tasks”

Dr. Krishna Kaipa

Abstract

“Many physical manipulation tasks—in diverse domains ranging from manufacturing and surgery to daily activities—rely on significant contact between the tool and the work surface (e.g., rust removal by scrubbing, finishing, cutting and manipulating a tissue during a surgical operation, etc). These physical tasks are highly non-repetitive, where optimal values for task parameters, change from one task instance to the next based on change in work-piece geometry, material, tool, etc. Mathematical models that map task parameters to performance are not tractable for such non-repetitive tasks, making it difficult to program and automate robots to perform them. Therefore, such tasks are currently only handled manually by humans. This talk will describe machine learning approaches for enabling robots to handle non-repetitive manufacturing tasks. First, an overview of an integrated decision making approach that combines perception, planning, control, and learning to realize adaptive robotic assistants will be provided. This will be followed by presenting an approach for robots to learn optimal task parameters by using self-directed experiments. Next, an extension of this approach to handle compliant manipulation tasks will be presented. Both humans and robots can make errors while performing manipulation tasks, hence creating contingency situations. A decision making approach for detecting and managing contingencies will be presented. Cleaning, pouring, and bin-picking tasks will be used as illustrative examples to show how robots can be used to handle non-repetitive manufacturing tasks.”

Join with Bluejeans: https://bluejeans.com/707697867

June 10, 2020

“Natural Language Processing for Accelerating Scientific Breakthroughs”

John Dagdelen

Abstract

“The majority of all materials data is currently scattered across the text, tables, and figures of millions of scientific publications. In my talk, I will discuss the work of our team at Lawrence Berkeley National Laboratory on the use of natural language processing (NLP) to extract and discover scientific knowledge through textual analysis of the abstracts of several million journal articles. With this data we are exploring new avenues for materials discovery and design such as how functional materials like thermoelectrics can be identified by using only unsupervised word embeddings for materials. I will also discuss how we have pivoted to COVID-19 research tools, such as models that can help identifying papers from before the COVID-19 pandemic that may contain highly relevant information on similar viruses. To date, we have used advanced techniques for named entity recognition to extract more than 100 million mentions of materials, structures, properties, applications, synthesis methods, and characterization techniques from our database of over 3 million materials science abstracts and we are in the process of performing similar knowledge extraction for the COVID-19 literature. Finally, I will also give an overview on how we are making all of this data freely available to the materials research community through our public-facing websites (https://www.matscholar.com/ and https://covidscholar.org/) and open-access APIs.”

Join with Bluejeans: https://bluejeans.com/707697867

May 27, 2020

"Machine Learning for Data Streams with Python”

Dr. Jacob Montiel

Abstract

“As traditional "batch" learning is no match for today’s data deluge, a new field emerges — data stream mining. In stream learning, data is considered infinite and models are trained and updated continuously, thereby adapting to changes in the data. This talk provides an overview of data stream learning and introduces scikit-multiflow, an open-source Python framework to implement algorithms and perform experiments in the field of ML on evolving data streams.”

Join with Bluejeans: https://bluejeans.com/274517153

May 20, 2020

“Introduction to Deep Learning”

Dr. Mustafa Mustafa

Abstract

I will introduce the modern incarnation of neural networks, specifically the construction process of deep models that are powering most recent advances in artificial intelligence. Assuming no prior knowledge of the field, the audience will be brought up to speed with how neural networks work, how they are built and optimized and common tricks and techniques to improve their performance. By the end of the talk the audience will be familiar with start-of-the-art convolutional architectures and the stages of developments that brought them about. If time permits, I will overview common science problem instances that can be tackled using deep learning.

Join with Bluejeans: https://bluejeans.com/707697867

May 13, 2020

“Deep Learning at the Frontier of Particle Physics”

Dr. Alex Radovic

Abstract

"Our knowledge of the fundamental particles of nature and their interactions is elegantly summarized by the standard model of particle physics. Advancing our understanding in this field has required experiments that operate at ever higher energies and intensities, which produce extremely large and information-rich data samples. The use of advanced machine-learning techniques is revolutionizing how we interpret these data samples, greatly increasing the discovery potential of present and future experiments. This talk will summarize the challenges and opportunities that come with the use of deep learning at the frontiers of particle physics, with a particular focus on examples from the thriving subfield of neutrino physics."

Join with Bluejeans: https://bluejeans.com/707697867

October 13, 2021

“Learning Causality with Graphs”

Dr. Jundong Li

Abstract

"The ability to learn causality is considered as a significant component of human-level intelligence and can serve as the foundation of AI. In causality learning, one fundamental problem is to understand the causal effects of a specific treatment (e.g., prescription of medicine) on an important outcome (e.g., cure of a disease), with significant implications in various high-impact domains such as health care, education, and e-commerce. One prevalent way to solve the problem is to directly use the observational data since the alternative randomized experiments could be expensive, time-consuming, and even unethical in many scenarios. However, existing data-driven methods are often limited since they: (1) assume that observational data is independent and identically distributed (i.i.d.); and (2) ignore the influence of hidden confounders (i.e., the unobserved variables that affect both the treatment and the outcome). Meanwhile, real-world data is often connected and can be abstracted as graphs (e.g., social networks, biological networks, and knowledge graphs). The ubiquitous of graph data across many influential areas also brings opportunities to control the influence of hidden confounders and build more effective models that yield unbiased causal effects estimation. In this talk, I will introduce our recent research efforts in causal effects learning with graphs. Specifically, we attempt to answer the following research questions: How to utilize graph information among observational data for causal effects learning? How to harness the power of historical information to tame the influence hidden confounders for causal effects learning when the graph is continuously evolving?"

Join with Bluejeans:  https://bluejeans.com/786906712/0315

June 30, 2021

"Extracting the Most from Collider Data with Deep Learning"

Dr. Ben Nachman (LBL)

Abstract

“Precise scientific analysis in collider-based particle physics is possible because of complex simulations that connect fundamental theories to observable quantities. These simulations have been paired with multivariate methods for many years in search of the smallest distance scales in nature.  Deep learning tools hold great promise to qualitatively change this paradigm by allowing for holistic analysis of data in its natural hyperdimensionality with thousands or millions of features instead of up to tens of features. These tools are not yet broadly used for all areas of data analysis because of the traditional dependence on simulations. In this talk, I will discuss how we can change this paradigm in order to exploit the new features of deep learning to explore nature at sub-nuclear distance scales. In particular, I will show how neural networks can be used to (1) overcome the challenge of intractable hypervariate probability density modeling and (2) learn directly from (unlabeled) data to perform hypothesis tests that go beyond any existing analysis methods. The example for (1) will be full phase space unfolding and the example for (2) will be anomaly detection. The talk will include a discussion of uncertainties associated with deep learning-based analyses.”

Join with Bluejeans:https://bluejeans.com/451415148/8427

June 2, 2021

“Graph Neural Networks and Metric Learning for the Future of Particle Physics”

Dr. Daniel Murnane (LBL)

Abstract

"The discovery of the Higgs boson at the Large Hadron Collider was a colossal effort that brought together many advanced experimental, theoretical and computational techniques. We are opening a door to a new generation of particle physics experiments that will try to answer questions about dark matter, dark energy and why the universe seems to be so perfectly tuned for complex structures, like life. Many Higgs-era techniques may not be powerful enough for these new energy/intensity scales, and so we turn to machine learning (ML) techniques to help answer these questions. It turns out that not only does ML boost our ability to discover new physics, it also challenges us to think about physics problems in new ways."

Join with Bluejeans: https://bluejeans.com/260264787

May 05, 2021

"Interpretable Machine Learning for Physics, using Symbolic Regression and Graph Neural Networks"

Miles Cranmer

Abstract

“In this talk I will argue two points. 1) Symbolic regression, a machine learning technique that fits data by iteratively searching the space of all possible analytic equations, should be a standard machine learning algorithm in physics. 2) Symbolic regression can be extended to high-dimensional spaces, such as to models for N-body simulations, using the method we have developed. To begin the talk, I will introduce symbolic regression (SR), which is a relatively old but underdeveloped technique. I will demonstrate our new high-performance open-source SR software "PySR". I will then discuss our research contribution: a technique for extending SR to high-dimensional spaces by the use of a neural network. I will focus on Graph Neural Networks (GNNs), which are a physically motivated neural network architecture, and very relevant to many physical and astrophysical problems. As a validation, I will demonstrate that we can use this technique of a GNN to SR to extract the correct force laws and Hamiltonians from simple particle simulations. I will then show work applying our method to the Quijote dark matter simulations – where it finds a simple analytic formula to estimate the overdensity of dark matter in a halo using learned metrics which summarize the nearby halos, giving a functional form that is more accurate than hand-derived formulas. Our approach simultaneously offers an interpretable ML technique for high-dimensional astrophysical data, and also a way of interpreting neural networks. (Demo code is available here).

Join with Bluejeans: https://bluejeans.com/500985640

February 24, 2021

CTwin - Developing a Transportation/Mobility Digital Twin for Chattanooga

Dr. Jibo Sanyal (ORNL)

Abstract

"The Computational Urban Sciences Group at ORNL, in partnership with NREL and several external stakeholders, have stood up a real-time digital twin focused on mobility for Chattanooga. The system has brought in 500+ real-time data feeds from 5 systems across 3 institutions, with at least 40 other secondary data sets. This has created an unprecedented opportunity to observe, anticipate, and orchestrate cyber-physical controls towards a 20% energy savings objective for the region. This talk focuses on the experiences in obtaining real-time data from location enabled sensors from infrastructure and how it is heralding a fundamental shift in how we interact with our urban environment."

Join with Bluejeans: https://bluejeans.com/651785849

February 17, 2021

“Deeply Uncertain: Comparing Methods of Uncertainty Quantification in Deep Learning Algorithms”

Dr. João Caldeira (Google)

Abstract

"In recent years, many methods for quantifying uncertainty in the output of deep learning algorithms have been developed. In this talk we will introduce three of the most prominent methods - Bayesian Neural Networks (BNN), Concrete Dropout (CD), and Deep Ensembles (DE) - and compare them to the standard analytic error propagation in the context of a single pendulum. We will also relate the output of these methods to terms more familiar in the context of the physical sciences. Our results highlight some pitfalls that may occur when using these UQ methods. For example, when the variation of noise in the training set is small, all methods predicted the same relative uncertainty independently of the inputs. This issue is particularly hard to avoid in BNN. On the other hand, when the test set contains samples far from the training distribution, we found that no methods sufficiently increased the uncertainties associated to their predictions. This problem was particularly clear for CD. In light of these results, we make some recommendations for usage and interpretation of UQ methods."

Join with Bluejeans: https://bluejeans.com/950395297  

February 10, 2021

“Machine learning-based Design and Control of Particle Accelerators”

Dr. Adi Hanuka (SLAC)

Abstract

"Machine learning has been used in various ways to improve accelerator operation including advanced optimization of accelerator operating configurations, development of virtual diagnostics to ’measure’ beam parameters, and prognostics to detect anomalies and predict failures. In this talk I'll review machine learning techniques currently being investigated at particle accelerator facilities, with specific emphasis on physics-informed active-learning research efforts."

Join with Bluejeans: https://bluejeans.com/230468745

January 20, 2021

Deep Learning on Graphs: Methods and Applications”

Dr. Lingfei Wu (IBM Research)

Abstract

"Recent years have seen a significantly growing amount of interest in graph neural networks (GNNs), especially on efforts devoted to developing more effective GNNs for node classification, graph classification, and graph generation. However, there are relatively fewer studies on other important topics such as deep graph learning, graph-based encoder-decoder, and deep graph matching. In the first part of the talk, I will introduce the basics of deep learning on graphs. In the second part of the talk,  I will introduce an end-to-end graph learning framework, namely Iterative Deep Graph Learning (IDGL), for jointly learning graph structure and graph embeddings simultaneously. In the third part of the talk, I will introduce a Graph2Seq neural network framework and then talk about how to apply this model in various NLP tasks."

Join with Bluejeans: https://bluejeans.com/981665133 

November 18, 2020

“What Entropy and Impedance Mean in Data Science”

Dr. Pete Alonzi (UVA)

Abstract

"Today every grant proposal seems to demand "interdisciplinarity". This talk is about how to think interdisciplinarialy between Physics and Data Science. The fundamental pieces of Data Science thinking will be laid out starting from the 4+1 model developed by Alvarado. The universal pipeline theory will be presented along with how the concepts of Entropy and Impedance are used to understand the scope of any given pipeline. We will conclude with a discussion of how using these Data Science concepts overlapped with Physics investigation can produce enhanced results."

Join with Bluejeans: https://bluejeans.com/990782096 

October 28, 2020

"Discussion on Continual Learning and Learning Dynamics for Deep Networks"

Dr. Razvan Pascanu

Abstract

I will begin with a brief introduction to, and open problems in, understanding neural networks and discuss several approaches to this problem. Further I will discuss the importance of continual learning for the field. Continual Learning, sometimes called life-long learning, never-ending learning or open-ended learning, is a topic that has been growing in popularity in the last few years. I will provide an overview of the problem, contrasting sequential learning to multi-task learning and highlight why multi-task learning can be in practice sub-optimal. I will discuss the main family of approaches to continual learning and discuss recent research in the field. Finally I will provide a few ideas of where the field might go in the near future and open the floor for discussion.

Join with Bluejeans: https://bluejeans.com/395185997 

September 30, 2020

"Who is Building AI and Who is AI Being Built For?"

Dr. Emily Denton

Abstract

AI systems are currently being deployed in a range of socially consequential domains, including policing, healthcare, employment, and more. Yet, despite the narrative of progress that frequently surround the development and integration of AI technologies, the benefits and risks of these systems are highly unequally distributed. This talk will provide an overview of the ways AI systems are impacting society, and failing marginalized communities, and the steps practitioners can take to mitigate potential harms.

Join with Bluejeans: https://bluejeans.com/130442674 

September 2, 2020

"Situating AI on the Road from Data Sharing to Societal Impact"

Dr. Daniel Mietchen

Abstract

By its very nature, Artificial Intelligence depends on the availability of data at scale. In this presentation we will look at a range of factors that influence the nature and scale of data sharing, from open science to disasters, from research infrastructures to ethics, from cooperation to competition. We will then delve into how these factors affect the data life cycle and the research cycle and explore how data sharing (or the lack of it) translates into societal impact. On that journey, we will watch out for ways in which AI can and does contribute to or benefit from the sharing of data and associated resources (or not), which will then form the basis of our discussion.

Join with Bluejeans: https://bluejeans.com/215807327  

July 29, 2020

“The Convergence of HPC and AI: New Methods for Computer Modeling to Address Grand Challenges in Science”

Tom Gibbs

Abstract

"The talk introduces the challenge with conventional modeling to address Grand Challenge problems as Moore’s Law is coming to an end, and the opportunity that is being demonstrated with early examples using Converged HPC and AI Methods. Conventional ModSim methods introduce a constraint with cost/accuracy as model size and fidelity increase, and the end of Moore’s Law further imposes a hard limit on raw performance. GPU accelerators have been shown to extend the raw compute speed by an order of magnitude, but cost/accuracy constraint with conventional methods become a hard limit with Grand Challenge model sizes. A taxonomy is emerging for the application of HPC and AI methods to address science problems. These include methods to dramatically improve the scale, accuracy and time to solution with converged modeling and simulation. The converged methods are also being applied to experimental data processing where detection accuracy is improved dramatically, and indicate the potential to introduce real time control to future experiment runs. Examples are given for Grand Challenge problems from multiple domains including fusion, particle physics, weather and climate, biochemistry, astrophysics, and material science."

Join with Bluejeans: https://bluejeans.com/707697867

July 22, 2020

“Neural Networks Love to Cheat: Shortcut Learning in Deep Learning and Beyond”

Robert Geirhos

Abstract

"Deep learning has triggered the current rise of artificial intelligence and is the workhorse of today's machine intelligence. Numerous success stories have rapidly spread all over science, industry and society, but its limitations have only recently come into focus. In this talk I will seek to distill how many of deep learning's problem can be seen as different symptoms of the same underlying problem: shortcut learning. Shortcuts are decision rules that perform well on standard benchmarks but fail to transfer to more challenging testing conditions, such as real-world scenarios. Related issues are known in Comparative Psychology and Education, suggesting that shortcut learning may be a common characteristic of learning systems, biological and artificial alike. Based on these observations, I propose a set of recommendations for model interpretation and benchmarking."

Join with Bluejeans: https://bluejeans.com/707697867

June 17, 2020

“Machine Learning Approaches for Enabling Robots to Handle Non-repetitive Manufacturing Tasks”

Dr. Krishna Kaipa

Abstract

“Many physical manipulation tasks—in diverse domains ranging from manufacturing and surgery to daily activities—rely on significant contact between the tool and the work surface (e.g., rust removal by scrubbing, finishing, cutting and manipulating a tissue during a surgical operation, etc). These physical tasks are highly non-repetitive, where optimal values for task parameters, change from one task instance to the next based on change in work-piece geometry, material, tool, etc. Mathematical models that map task parameters to performance are not tractable for such non-repetitive tasks, making it difficult to program and automate robots to perform them. Therefore, such tasks are currently only handled manually by humans. This talk will describe machine learning approaches for enabling robots to handle non-repetitive manufacturing tasks. First, an overview of an integrated decision making approach that combines perception, planning, control, and learning to realize adaptive robotic assistants will be provided. This will be followed by presenting an approach for robots to learn optimal task parameters by using self-directed experiments. Next, an extension of this approach to handle compliant manipulation tasks will be presented. Both humans and robots can make errors while performing manipulation tasks, hence creating contingency situations. A decision making approach for detecting and managing contingencies will be presented. Cleaning, pouring, and bin-picking tasks will be used as illustrative examples to show how robots can be used to handle non-repetitive manufacturing tasks.”

Join with Bluejeans: https://bluejeans.com/707697867

June 10, 2020

“Natural Language Processing for Accelerating Scientific Breakthroughs”

John Dagdelen

Abstract

“The majority of all materials data is currently scattered across the text, tables, and figures of millions of scientific publications. In my talk, I will discuss the work of our team at Lawrence Berkeley National Laboratory on the use of natural language processing (NLP) to extract and discover scientific knowledge through textual analysis of the abstracts of several million journal articles. With this data we are exploring new avenues for materials discovery and design such as how functional materials like thermoelectrics can be identified by using only unsupervised word embeddings for materials. I will also discuss how we have pivoted to COVID-19 research tools, such as models that can help identifying papers from before the COVID-19 pandemic that may contain highly relevant information on similar viruses. To date, we have used advanced techniques for named entity recognition to extract more than 100 million mentions of materials, structures, properties, applications, synthesis methods, and characterization techniques from our database of over 3 million materials science abstracts and we are in the process of performing similar knowledge extraction for the COVID-19 literature. Finally, I will also give an overview on how we are making all of this data freely available to the materials research community through our public-facing websites (https://www.matscholar.com/ and https://covidscholar.org/) and open-access APIs.”

Join with Bluejeans: https://bluejeans.com/707697867

May 27, 2020

"Machine Learning for Data Streams with Python”

Dr. Jacob Montiel

Abstract

“As traditional "batch" learning is no match for today’s data deluge, a new field emerges — data stream mining. In stream learning, data is considered infinite and models are trained and updated continuously, thereby adapting to changes in the data. This talk provides an overview of data stream learning and introduces scikit-multiflow, an open-source Python framework to implement algorithms and perform experiments in the field of ML on evolving data streams.”

Join with Bluejeans: https://bluejeans.com/274517153

May 20, 2020

“Introduction to Deep Learning”

Dr. Mustafa Mustafa

Abstract

I will introduce the modern incarnation of neural networks, specifically the construction process of deep models that are powering most recent advances in artificial intelligence. Assuming no prior knowledge of the field, the audience will be brought up to speed with how neural networks work, how they are built and optimized and common tricks and techniques to improve their performance. By the end of the talk the audience will be familiar with start-of-the-art convolutional architectures and the stages of developments that brought them about. If time permits, I will overview common science problem instances that can be tackled using deep learning.

Join with Bluejeans: https://bluejeans.com/707697867

May 13, 2020

“Deep Learning at the Frontier of Particle Physics”

Dr. Alex Radovic

Abstract

"Our knowledge of the fundamental particles of nature and their interactions is elegantly summarized by the standard model of particle physics. Advancing our understanding in this field has required experiments that operate at ever higher energies and intensities, which produce extremely large and information-rich data samples. The use of advanced machine-learning techniques is revolutionizing how we interpret these data samples, greatly increasing the discovery potential of present and future experiments. This talk will summarize the challenges and opportunities that come with the use of deep learning at the frontiers of particle physics, with a particular focus on examples from the thriving subfield of neutrino physics."

Join with Bluejeans: https://bluejeans.com/707697867