We’re seeing a lot of excitement about what we loosely call “fundamental models of neural activity.” These neural-based models are essentially like artificial intelligence (AI) chatbots that have become an inescapable part of our lives. These models read large amounts of neural data and learn their underlying statistics through training. After training, these neural-based models can generalize what they have learned and make new predictions. neural activity pattern, motor output or in response to sensory stimulation.
Perhaps the most interesting aspect of neural basis models is that they can learn from activity across different animals, relevant brain regions, or similar tasks. This suggests something remarkable about the brain. Just as AI chatbots take advantage of the finite number of characters and grammatical rules of a language to learn the underlying statistics of an individual writer’s work in order to write sonnets, suggest recipes, or discuss the meaning of life, the generalizability of neural-based models suggests that neural activity must also consist of “letters” and “grammatical rules” that are shared across the brain.
This article attempts to explain why these neural basis models work by relating their performance to recent neuroscientific discoveries. I will also highlight what I believe are important opportunities for neuroscientists to use neural basis models to understand the brain and the challenges ahead.
○
Over the past decade, it has become clear that this is the best explanation for many aspects of brain function. group activities pattern It is distributed across many neurons rather than by specialized subpopulations of neurons. For example, when trying to understand how a monkey makes a decision based on the relative strength of two consecutive stimuli, there are no “what” or “when” neurons. Instead, all neurons in the prefrontal cortex form two sets of collective activity patterns that each define. What kind of stimulus was given and when?. Similarly, there appear to be no “motor preparation” or “motor execution” neurons. monkey or Mouse Motor cortex: All neurons participate in collective movement preparation and movement execution activity patterns. these findings and many others strongly suggest that the “language of the brain” is based on specific collective patterns of single neuron activity Implement key processes Used in cognition, sensory, and motor control.
Importantly for us, these collective patterns appear to be relatively consistent across neurons within the population from which we record. This means that it is possible to recover any neural process from any subset. recorded neurons. Furthermore, these collective activity patterns are conserved across animals. Same motor operation or navigate same environment. They also often similar across Related tasksHowever, neuroscientists have no way of knowing in advance when this will happen. I don’t Get used to it case. Given that brain function is well explained by collective activity patterns that can be revealed from completely distinct subsets of neurons within a population, and that these collective activity patterns are invariant across individuals and similar across relevant tasks, it follows that the powerful statistical machines that underpin neural basis models can learn these invariances and commonalities by aggregating data across individuals and tasks. In other words, recent neuroscience discoveries explain why the neural basis model works and why it works.
Neural basis models are powerful tools that not only combine recent observations but also allow us to learn new things about the brain in at least two main directions.
First, neural basis models are helpful. Continuing to quantify and, in some cases, “qualify” the organization of activity patterns that mediate behavior will deepen our understanding of the language of neural populations. For example, they can be used to characterize how different neurodevelopmental and neurological conditions change the activity of neural populations. Neural basis models can also help define how different ways of solving interesting tasks, such as playing the same game of Tetris, affect the collective activity of neural populations. Finally, these models also provide an opportunity to define relationships. across different speciesby quantifying their similarity in the activity of neural populations. This line of investigation is critical to establishing homology across the world. evolution tree It is based not only on neuroanatomy and behavioral repertoire, but also on neural function. Given that many neurotechnologies target this “system level”; reduce motor and mental illnessthe ability to establish similarities between the activity patterns of various model organisms and humans could accelerate translational research and have tremendous societal impact.
Second, neural basis models can help establish connections between different levels of brain function, from neurotransmitters to single neurons, circuits, and brain regions. The population patterns of neural activity that I have been discussing are undoubtedly implemented by neurons. In other words, in principle, neurons can be reduced to their constituent neurons. However, whether or not it can be “constructed” from neurons is a different matter. constituent neurons This is an important question, so there is no room for discussion here. But there are important pressing questions. Is there a tractable mapping between these aspects of collective activity patterns and well-defined subpopulations of neurons? Recent studies have shown that it is possible to predict with considerable accuracy which areas of the mouse visual system a neuron will originate from based solely on neural activity. is recordedand can even distinguish between as many as 11 types of excitable cells. These interesting results are based on an examination of the individual mapping of single neurons to ensemble patterns learned by neural-based models, and suggest that discoverable mappings between subclasses of single neurons and ensembles do indeed exist.
T
The opportunities outlined above are exciting. However, it remains to be seen whether neural basis models reveal neural explanations of behavior that we can understand. What I’m trying to say is that because there are so many parameters, it may be difficult to derive what kind of mental objects are easy to think about. scientific understanding. For example, I could distill Newton’s laws to explain to my nephew why he can spin faster on his chair if he holds his arms and legs closer to his body, but does such a compressed description exist to map the features of neural-based models to behavior, or does it remain in the realm of incredible predictive machines?
Another pressing question is whether whole-brain recordings across many behaviors can be integrated. Most results to date have focused on specific brain regions across fairly similar tasks, or on a set of related regions during the same task or behavior, but of course this concern applies to all subfields of neuroscience. One of the fundamental challenges to achieving this overarching goal, with models as data- and parameter-intensive as current neural-based models, is whether we will end up in the situation that Jorge Luis Borges predicted. perfect map Their map of the empire matched point-for-point with the empire itself, but no one cared about that. That’s because we couldn’t learn anything useful from the model, a map as detailed as what it represents.
Finally, the kinds of neural basis models I have been discussing do not have a body. In principle, one could argue that sensing and acting on the world through the body can be abstracted as a “transformation” that can be learned given enough neural and behavioral data. So how much data is enough data? My intuition is that this problem may actually become intractable. For example, it is not “just” learning complex transformations such as from the retina to the lateral geniculate nucleus and then to V1, or from motor cortical output to spinal circuits, motor neurons, and ultimately movement, but the fact that these transformations depend on many factors, such as: top-down modulation, Note, engagement, body conditionand perhaps even the organism’s developmental trajectory and lifetime experiences. This challenge gave me the motivation to work on: Basic model of brain-body-world interaction loopThis can potentially separate all these factors from what the neural population actually “does.”
Of course, neural basis models are useful even when we cannot understand the relationships between neurons, their collective activity, or the behaviors they reveal. For example, these relationships can be used to ‘know’ which neurons to target to shift population activity in the brain away from disease states, and to develop new neural techniques that can lead to new treatments. And perhaps, by using machine learning tricks, we might even be able to avoid falling into the same trap as Borges’ cartographers by using neural basis models as a new means of effectively graphing brain function. Ultimately, only by navigating this uncharted territory will we discover whether neural basis models can (no kidding) transform neuroscience and neurotechnology.
