Unsupervised Thinking

For this episode, we try to de-thorn one of the thorniest topics in neuroscience: consciousness! Starting with the (not so) simple task of defining the c-word, we first lay out our own definitions but end up relying on the work of philosopher David Chalmers (easy and hard problems of consciousness, philosophical zombies). Then, after establishing the myriad of reasons why science can't actually study consciousness, we get into its attempts to do so. Specifically, we talk about studies on anaesthesia and the role of optical illusions in probing conscious perception. After that we go through a laundry list of all kinds of theories/models of consciousness put forth by neuroscientists, psychologists, and (ugh) physicists. Throughout, Josh complains about people saying they have "heightened consciousness", Conor advocates a lot of drug use (for science), and we discuss the politics of urination. If you're confused by any of this, no worries, Conor is too.

We read:

Scholarpedia: Neuronal Correlates of Consciousness

Scholarpedia: Models of Consciousness

Neural Correlates of Consciousness (sensory paper by Rees)

Neural correlates of consciousness during general anesthesia using functional magnetic resonance imaging (fMRI)

And mentioned:

20 Years of ASSC: are we ready for its coming of age? (thanks to @tweetsatpreet for pointing us to this journal)

Sloppiness, stiffness, and stomatogastric ganglion! This episode on underdeterminacy in neural circuits will introduce you to all these topics, as well as to special guest Alex Williams! To start, we take you way back to algebra class with a refresher on what makes a system "underdetermined" (essentially, more unknowns than constraints). There are two ways this can be a problem in neuroscience: (1) neural circuit modelers don't have enough data to constrain their models, and (2) biology itself is underconstrained, leading to differences across individuals within a species. We talk about both of these issues separately, the ways in which they interact, and the practical effects they have for the study of the nervous system. The first topic spurs a broad discussion on the philosophy of modelling and the potential pitfalls that careful scientists need to avoid. To explore this in more detail, we discuss an excellent modelling paper on the oculor-motor system that demonstrates ways in which models should guide experiments. For the latter topic, we delve into Eve Marder's work on crustaceans, wherein she carefully documents the incredible variety across individuals. Having worked in Eve's lab himself, Alex provides expertise and anecdotes on this topic throughout!

We read:

Computational models in the age of large datasets

A modeling framework for deriving the structural and functional architecture of a short-term memory microcircuit

ROBUST CIRCUIT RHYTHMS IN SMALL CIRCUITS ARISE FROM VARIABLE CIRCUIT COMPONENTS AND MECHANISMS

And mentioned:

Our episode on "Does Neuroscience Need More Behavior?"

Why Are Computational Neuroscience and Systems Biology So Separate?

James Sethna's work on sloppiness

To much of the world, the face of neuroscience is an image of a brain with small colored blobs on it. Those images come from functional magnetic resonance imaging (fMRI), a technology that's made a big splash in its relatively short tenure. For this episode, we delve into fMRI and what scientists do with the data it produces. To start, we review the technology behind MRI and fMRI. We get into the thorny issue of relating the BOLD signal recorded from fMRI with actual neural activity, and what's been learned from animal studies that have looked at both simultaneously. After that we talk stats: particularly the trouble with traditional "voxel"-wise methods and putting all your eggs in one basket (or in separate, but similar, baskets?). Approaches to fMRI analysis are quickly evolving however, and so we discuss multi-voxel pattern analysis, comparing across individual brains, real time-analysis, mind reading, and lie detecting. Finally, we turn a little more philosophical and ask "What does it mean to measure information in the brain?". Is what an experimenter can see in these colored patterns even relevant to the brain itself?? We give examples of when it's not.

We read:

Interpreting the BOLD Signal

Computational Approaches to fMRI Analysis

Is Neuroimaging Measuring Information in the Brain?

Sometimes scientists decide to turn their tools of inquiry inward to understand their own fields and behaviors. For our 20th episode, we're diving into this meta-science by reading some papers about by papers written by scientists studying scientists. In particular, we start with a commentary discussing the increasing size of scientific teams, and what that means for credit assignment. Do we need to move to a more Hollywood approach by highlighting specific achievements in different roles? Also, when will we address the fact that most young researchers on these teams will not have a career in academic science? We then get into a modeling study that aims to show how incentivizing the publication of novel results can ultimate lead to a widespread decrease in scientific quality. This raises questions of whether individuals or the system is to blame for high rates of shoddy publications. We then touch on a small experiment that the conference NIPS (Neural Information Processing Systems) performed on their peer review system, showing that (spoiler alert!, or probably not if you've been subjected to peer review...) the process can appear somewhat random. Finally, we go over a report that tracked trends in neuroscience research over the past ten years. We find that a meta-study of a field can seem very different from the view inside of it. Finally, we mention how studies of science done by scientists differ from those done by the humanities, and how both may be of use.

We read:

Together We Stand

The Natural Selection of Bad Science

The NIPS Experiment

The Changing Landscape of Neuroscience Research, 2006–2015: A Bibliometric Study

Way back in Episode 16 we paired up with Always Already to talk about a book on gender and the interaction of science and society. Unsurprisingly, that conversation spanned far beyond the scientific study of gender and so we never really got into the biological weeds. Our intent with this episode was to go back to gender, with a focus on the explaining the current state of the science. What we quickly learn, however, is that it's very difficult to talk about gender without talking about society. So we first work through this by airing our anxieties on the topic, and our personal motivations for finding this science interesting. Eventually though, we break into the biology of embryonic sexual differentiation and certain "natural experiments" that alter the course of this differentiation. People with abnormal differentiation offer a chance to see what happens when things like chromosomal sex (XX vs XY) and external genitalia are decoupled, which offers some insight into normal gender development. Next we cover some biological hypotheses on sex that didn't pan out (but are still being promoted...). Finally we turn to the better controlled world of animal experimentation and cover what factors impact gendered behavior in macaque monkeys.

As it turns a lot of findings on gender don't replicate, but here's one that does: whenever the Unsupervised Thinking crew has a conversation on gender it takes more than an hour.

For most people, the desire to study neuroscience comes from a desire to understand how, in some form, the brain leads to behavior. Generally, neuroscientists focus on the brain side of that relationship, but what obligation do they have to study behavior? Is it even possible to do proper neuroscience without a clear documenting of the behavior we seek to understand? We use a recent opinion article ("Neuroscience Needs Behavior: Correcting a Reductionist Bias") as a jumping off point to discuss these issues. In the paper, the authors argue that behavior is being neglected amongst neuroscientists and it must return to its status as "epistemologically prior." In particular, there are arguments for studying more natural behavior and quantifying behavior more precisely.

In this episode we explain our general sympathies with this argument, but question the extent to which change is required. Should all neuroscientists stop what they're doing and study behavior? Are modern technologies drawing scientists away from the "bigger questions"? No, probably not. But this article does bring up questions about how we, as individuals and as a field, choose what to study. Different implicit beliefs about what levels of explanation are satisfying lead to different research priorities. Progress in neuroscience would be best suited by neuroscientists who better understand these implicit beliefs in themselves and others.

There's a lot to be said about the technical progress being made with artificial intelligence, but what about the impact these rapid advances have on the society in which they unfold? In this episode, we tackle a broad range of such issues, from the possibility of removing human bias from algorithms to how likely we are to fall in love with an AI (Conor might). We speculate on how difficult the transition from humans to self-driving cars will be and our wild uncertainty about the future of jobs/the value of human labor. Throughout you will see a poorly-veiled concern about the current political state of the world and how wealth and power will be distributed in the future. What we learn though, is that in addition to the economic and technological impacts, the use of AI is having at least one major side effect: it's forcing us to explicitly define our goals and values, such that we can impart them to our digital offspring. Now if we could just agree on what those goals and values are...

On this "very special" episode of Unsupervised Thinking, we partner with our fellow podcasters over at Always Already, a critical theory podcast, to burst out of our respective academic bubbles and tackle issues of science and society. The fodder for our conversation is Brain Storm, a book by Rebecca Jordan-Young, that lays out the evidence that prenatal hormone exposures influence gender differences in behavior later in life. In the book, she claims that the sum total of the studies she covers only offers weak support for the hypothesis, and that scientists need to appropriately incorporate other factors into their models such as socialization and environment.

While we use this book as a common starting point, our conversation quickly moves beyond the particulars of these gender science studies. We start by questioning who is the intended audience of this book and what it's trying to say to different groups. This moves us into a discussion on critiques of science made by non-scientists and the role that those should/could play in shaping research agendas. We also spend some time dissecting the two-way street between science and society: particularly, how are common notions of gender shaped by scientific studies and how do society's stereotypes seep into the methods of science? An underlying disagreement about the nature of truth peppers the discussion, but we hold off on a full blown debate on that. Ultimately it is clear that the extent and cause of gender differences in behavior is far from settled science, and that is something on which we all can agree.

In the late 1700s, English minister Thomas Bayes discovered a simple mathematical rule for calculating probabilities based on different information sources. Since then Bayesian models for describing uncertain events have taken off in a wide variety of fields, not the least of which is psychology. This Bayesian framework has been used to understand far-reaching psychological processes, such as how humans combine noisy sensory information with their prior beliefs about the world in order to come to decisions on how to act.

But not everyone is riding the Bayesian train. In this episode, we discuss a published back and forth between scientists arguing over the use and merits of Bayesian modeling in neuroscience and psychology. First, though, we set the stage by describing Bayesian math, how it is used in psychology, and the significance of certain terms such as "optimal" (it may not mean what you think it does) and "utility". We then get into the arguments for and against Bayesian modeling, including its falsifiability and the extent to which Bayesian findings are overstated or outright confused. Ultimately, it seems the expansive power of Bayesian modeling to describe almost anything may in fact be its downfall. Do Bayesian models give us insight on animal brains and behaviors, or just a bunch of "just-so" stories?

Computational psychiatry is a field in its infancy, but it offers potential to change the way mental disorders are diagnosed, treated, and understood. In this episode, we cover the different components of computational psychiatry, compare and contrast it to computational neuroscience (and old-school psychiatry), and discuss which of its promises are actually likely to be fulfilled. In particular, we get into:

-online games that inform diagnoses,

-models of learning mechanisms that can explain disorder,

-machine learning techniques that advise treatment plans,

-neural circuit mechanisms that can('t) explain disease