April 17, 2020
Speaker: Andrew Ng
Stanford CS229: Machine Learning (Autumn 2018)
[video]
Independent Component Analysis (ICA) is a powerful method that is used in the process of some methods for denoising fMRI. However, ICA is a very general blind source separation method that is useful in many domains. The premise of ICA is to be able to take some observations and recover the independent source that generated the observations. The canonical example is two speakers speaking into microphones, where the microphones pick up the voice of both speakers. ICA can recover the voice of each speaker separately from the microphone recordings. In fMRI, one can think of neural activity to be the voices and voxels, equivalent to pixels in an image, to be the microphones. Similarly, one can recover the distinct neural activities (in theory) that may have been combined into voxels.
This lecture discusses the ICA method clearly and in detail so that ICA is no longer just a black box to the listener. The speaker is giving a lecture to students, so the structure of the presentation is designed for education, which makes it a very useful resource. The speaker uses Reinforcement Learning (RL) as an application but keeps this part distinct from the general ICA discussion, so the viewer may choose to stop watching when it gets to that part. It would be instructive for the viewer to think about problems in their research area that ICA may be able to help with. It could range from denoising or providing intuition for the sources of observations from experiments executed in non-full-controlled settings
]]>May 27, 2021
Speaker: Jonas Peters
Host: Broad Institute, MIT
[video]
This webinar is a lecture by Jonas Peters, who has contributed a lot of important work to causal discovery and inference. This presentation is slightly technical in that it provides mathematical definitions of problems in causality; however, it is an excellent introduction. The speaker explains many essential objects and terms such as interventions (do-operations), structural equation models, adjustment, and backdoor paths. In addition, the explanations of these concepts are very nicely tied to intuitive examples that the viewer can follow.
This talk is a good introduction for those who watched the first webinar I posted featuring a conversation between Judea Pearl and Lex Fridman and wanted to learn more. It provides a sound basis for understanding works in this field that might be useful for their own work. Given the generality of the tools and methods developed in this line of work of causal discovery and inference, I hope the user can use this as a starting point for rigorously surveying this field and use the resources to further their research. Perhaps they will be able to use instrumental variables to answer scientific questions that they weren’t able to before due to experimental limitations or use the backdoor criterion to design controls for experiments more carefully.
]]>September 14, 2021
Speaker: Dr. Rasmus Birn
Host: Open Minds @ Pitt
[video]
This webinar is a comprehensive overview of the influence of motion and physiological noise on fMRI. As the title says, Dr. Birn presents an end-to-end synopsis of the problem, from introducing the BOLD signal to current solutions to further challenges.
The BOLD signal (Blood Oxygenation Level Dependent), which aims to measure a proxy of neuronal activity, is tiny and therefore susceptible to being influenced by artifacts from motion and physiological noise. When interested in uncovering functional connectivity, connections between brain regions, these artifacts can lead to false conclusions. For instance, when looking at child vs. adult brain in fMRI, a significant confound is that children tend to move more than adults, leading to many more false positives in non-artifactual activations. Current techniques to mitigate this problem include measuring physiological confounds via external devices, such as a heart monitor, and regressing out the measurements from the scan, measuring frame to frame affine changes in the position of the skull, and regressing out the positional changes from the scan. Despite the availability of techniques used to remove artifacts, it remains a challenge to know if they worked. QC-FC vs. Distance is one attempt to address this challenge for motion. Distances between networks are correlated with the correlations between strengths of paired networks and estimated motion parameters. There is no correlation in the ideal case, which implies that the method applied for motion correction ‘worked.’ However, even this method has limitations, which Dr. Birn explains.
This webinar is impressive because of its accessibility. Viewers with no understanding of fMRI can follow along while developing experts can acquire new knowledge and understanding.
]]>December 11, 2019
Speaker: Judea Pearl
Host: Lex Fridman @ Lex Fridman Podcast
[video]
This webinar presents a discussion between Lex Fridman and Turing award winner Judea Pearl on the nuances of causal reasoning, counterfactuals, and the path to Artificial General Intelligence (AGI). At the core of science is uncovering truths about how the world works, and generally, understanding how the world works requires understanding interactions between the systems of the world – causality. Many hypotheses, however, are untestable in a direct fashion due to various challenges – whether it be a sheer impossibility or ethical limitations. How can one seek to uncover relationships under these constraints? This question is the focus of Judea Pearl’s work on causality. This discussion gives a high-level introduction to work that addresses how to answer questions such as i) I know there is a confounding variable, but I can’t measure it directly; ii) I don’t understand the system enough even to know which experiments I should design; iii) I have limited resources; what set of experiments should I run to gain the most information.
This webinar is beneficial because of its accessibility. Often, the conceptualization of a problem lends itself to a particular methodological approach. The speaker gives a broad overview of the questions this area of study can help address, and researchers can quickly connect their work and the abstract principles that the speaker explains. While there is not enough detail here to lead to any concrete contribution to the listener’s research, it provides enough background to decide if the body of work being described may be helpful and worth exploring in more detail.
]]>