Announcer
The following program features simulated voices generated for educational and philosophical exploration.
Adam Ramirez
Good evening. I'm Adam Ramirez.
Jennifer Brooks
And I'm Jennifer Brooks. Welcome to Simulectics Radio.
Adam Ramirez
Tonight we're examining brain-computer interfaces—systems that read neural activity to control external devices or, more speculatively, write information back into the brain. Current BCIs can decode motor intentions with enough fidelity to control robotic arms and computer cursors. The question is what fundamental limits constrain these systems, and whether the leap from reading motor commands to reading and writing memories is a matter of engineering or requires conceptual breakthroughs we don't yet have.
Jennifer Brooks
The engineering challenges are substantial. You need stable chronic recordings from many neurons, reliable decoding algorithms that generalize across time, and minimal tissue damage from electrode arrays. But beyond engineering, there's the question of what neural signals actually carry. Motor cortex encodes movement intentions in a relatively straightforward way. Memory, by contrast, is distributed, context-dependent, and involves coordinated activity across multiple brain regions. Reading a memory might require recording from thousands of precisely identified neurons across distant areas.
Adam Ramirez
To explore both the current state and future possibilities of BCIs, we're joined by two researchers at the forefront of this field. Dr. Krishna Shenoy is a neural engineer at Stanford whose work focuses on the computational principles underlying motor control and high-performance BCI systems. Dr. Leigh Hochberg is the principal investigator of BrainGate, the clinical trial that has demonstrated motor BCIs in paralyzed individuals. Welcome to both of you.
Dr. Krishna Shenoy
Thank you for having us.
Dr. Leigh Hochberg
Glad to be here.
Adam Ramirez
Dr. Hochberg, let's start with the current state of motor BCIs. BrainGate participants can control computer cursors and robotic arms using decoded neural activity from motor cortex. What are the practical limits right now? What determines whether someone can pick up a cup versus perform fine manipulation?
Dr. Leigh Hochberg
The primary limits are bandwidth and stability. We're recording from roughly one hundred electrodes, which gives us access to perhaps one hundred to two hundred neurons if we're doing well. That's enough for two or three degrees of freedom of smooth control—moving a cursor on screen, controlling arm position. But dexterous manipulation requires coordinating many more degrees of freedom—fingers, wrist rotation, grip force. We'd need recordings from thousands of neurons with stable signal quality over years. Current electrode arrays degrade over months to years as tissue encapsulation reduces signal quality.
Jennifer Brooks
Let's talk about that degradation. Implanted electrodes trigger an immune response—microglia activation, astrocyte scarring around the array. This forms an insulating sheath that progressively attenuates the neural signals. Are there approaches to minimize this reaction, or is it an unavoidable consequence of inserting foreign material into brain tissue?
Dr. Leigh Hochberg
It's a major challenge. We've tried coating electrodes with biocompatible materials, varying electrode geometry to reduce tissue strain, delivering anti-inflammatory drugs locally. These approaches help but don't eliminate the problem. One promising direction is to move away from rigid silicon probes toward flexible polymer electrodes that better match the mechanical properties of brain tissue, reducing chronic irritation. But even with perfect biocompatibility, you're still disrupting local circuits by inserting hundreds of electrode shanks.
Adam Ramirez
Dr. Shenoy, from a decoding perspective, what does the signal degradation do to your algorithms? Do you need to retrain decoders constantly as the neural signals change, or can you build decoders that adapt automatically?
Dr. Krishna Shenoy
We use adaptive decoders that recalibrate based on the user's intended movements. The system assumes that when someone is trying to move the cursor to a target, their neural activity should be decoded as movement toward that target. We use that assumption to update the decoder parameters continuously. This handles gradual signal drift reasonably well. But when electrodes fail completely—when a channel goes silent—you lose those degrees of freedom unless you can recruit new neurons to compensate. That requires the user to essentially relearn the mapping.
Jennifer Brooks
There's an interesting biological question here about neural population codes. Motor cortex doesn't encode movement through individual labeled neurons that each represent specific actions. Instead, populations of neurons with broad tuning collectively represent movement parameters. This population coding is redundant and distributed. Does that redundancy help BCI robustness? If you lose some neurons, can the remaining population still carry the information?
Dr. Krishna Shenoy
Absolutely. Population coding provides graceful degradation. Losing ten percent of your neurons reduces decoding accuracy by much less than ten percent, because the information is redundantly encoded. But there's a limit. Below some critical population size, performance drops sharply. The challenge is that we don't get to choose which neurons we lose—it depends on which electrodes degrade. If the lost neurons happen to carry information about specific movement directions or speeds, you get asymmetric degradation of performance.
Adam Ramirez
Let's talk about bandwidth. You mentioned needing thousands of neurons for dexterous control. How does that scale with the complexity of the behavior? Is there a theoretical limit to how many bits per second you can extract from neural activity?
Dr. Krishna Shenoy
Neurons fire at maybe one to ten spikes per second on average, and their responses are noisy. Information theory gives us bounds on the mutual information between neural activity and behavior, which works out to maybe a few bits per second per neuron under optimistic assumptions. So to achieve communication rates comparable to speech—maybe fifty bits per second—you'd need dozens of neurons if they're perfectly informative. In practice, you need hundreds because individual neurons are noisy and partially redundant. For complex motor control that's richer than speech, you're looking at thousands of neurons.
Jennifer Brooks
I want to push back on the assumption that more neurons always means better performance. There's evidence that motor cortex contains both task-relevant signals and uncorrelated noise. If you indiscriminately add neurons to your decoder, you're also adding noise. How do you determine which neurons to include? Do you select based on task-related tuning, or is there an optimal population size beyond which additional neurons hurt performance?
Dr. Krishna Shenoy
You're right that naive inclusion of all neurons can hurt. We typically use dimensionality reduction techniques—PCA, factor analysis, neural network encoders—to extract the low-dimensional manifold that carries task-relevant information. Many neurons contribute to this manifold, but their contributions are weighted by how informative they are. In practice, we find that the first ten to twenty principal components capture most of the task-relevant variance, even when recording from hundreds of neurons. So the effective dimensionality is much lower than the number of neurons.
Adam Ramirez
That raises a question about the neural population you're recording from. Motor cortex contains diverse cell types—excitatory pyramidal neurons, various interneuron types, neuromodulatory inputs. Current electrode arrays mostly pick up large pyramidal neurons because they have the biggest spikes. Are we getting a biased sample of the population, and does that limit what we can decode?
Dr. Leigh Hochberg
We're definitely sampling non-uniformly. Extracellular electrodes detect spikes from neurons within a few hundred micrometers, and spike amplitude decreases with distance. Large pyramidal cells dominate the recordings. Interneurons have smaller cell bodies and lower amplitude spikes, so they're underrepresented. This might matter if interneurons carry distinct information—timing signals, coordination signals, gain modulation. We don't know what we're missing because we can't record it with current technology.
Jennifer Brooks
Let's transition to the more speculative topic of memory BCIs. The premise is that if you can decode memories from neural activity, you could read thoughts, and if you can encode memories by stimulating neurons, you could write information into the brain. What are the fundamental differences between motor decoding and memory decoding that make the latter far more difficult?
Dr. Krishna Shenoy
Motor control has a clear behavioral output—arm movement, cursor position—that you can measure and use to train decoders. Memory retrieval doesn't have an observable behavioral output in the same way. You can ask someone what they remember, but that's mediated by language and decision-making processes that add layers of complexity. You'd need to decode the memory representation itself, which is distributed across hippocampus, cortex, and other structures, and highly dependent on context and internal state.
Dr. Leigh Hochberg
There's also the problem of access. Motor cortex is on the surface of the brain, so you can reach it with cortical surface arrays. Hippocampus is buried deep in the medial temporal lobe. Recording from hippocampus requires depth electrodes, which are far more invasive and carry higher risks. And you don't just need hippocampus—you need simultaneous recordings from prefrontal cortex, parietal cortex, perhaps sensory cortices, depending on what kind of memory you're trying to decode. The spatial coverage requirement is orders of magnitude larger than for motor BCIs.
Adam Ramirez
Even if you solve the recording problem, there's the question of what neural activity patterns correspond to specific memories. In rodents, hippocampal place cells activate when the animal is in a specific location, and replaying those patterns can reactivate place memories. But human episodic memory seems far more abstract and associative. How would you even know you're looking at a memory trace versus ongoing thought, imagination, or sensory processing?
Dr. Krishna Shenoy
That's the core problem. Memories aren't discrete, labeled patterns stored in specific neurons. They're emergent properties of coordinated activity across distributed networks. The same neurons that participate in encoding a memory also participate in many other memories and ongoing cognitive processes. Isolating the activity pattern corresponding to a specific memory from the background of general cognitive activity is extremely difficult. You'd need massive amounts of data—recording while the person forms memories, retrieves them, imagines similar events—to tease apart the components.
Jennifer Brooks
And then there's the writing side. If you wanted to implant a memory, you'd need to stimulate neurons with precise spatiotemporal patterns that replicate the natural activity during memory encoding. Current stimulation technologies—DBS, TMS, optogenetics in research—are far too coarse. DBS activates thousands of neurons indiscriminately. Optogenetics in humans doesn't exist yet due to gene therapy challenges. Even if you could stimulate specific neurons, you'd need to coordinate their activity in precisely the right temporal pattern to mimic natural encoding.
Dr. Leigh Hochberg
Right. The brain doesn't have a simple write interface where you can inject a pattern and have it stick. Memory consolidation involves synaptic plasticity mechanisms that unfold over hours to days. You can stimulate neurons to fire, but that doesn't mean you're triggering the plasticity mechanisms that create lasting memory. You might create spurious activations that are immediately forgotten, or worse, disrupt existing memories through interference.
Adam Ramirez
It sounds like the gap between motor BCIs and memory BCIs is not just quantitative—more neurons, better electrodes—but qualitative. We understand motor control well enough to build useful decoders. We don't understand memory representation well enough to even formulate the decoder architecture.
Dr. Krishna Shenoy
That's accurate. Motor BCIs benefit from decades of motor neuroscience that identified the relevant brain areas, characterized neural tuning properties, and established the coordinate transformations from neural activity to movement. Memory neuroscience has identified key structures, but we don't have comparable mechanistic understanding of how specific patterns of activity encode specific memories. Without that understanding, any memory BCI is essentially blind searching through neural activity space hoping to find meaningful patterns.
Jennifer Brooks
There have been some clinical attempts at memory enhancement through electrical stimulation in epilepsy patients with implanted electrodes. The results have been mixed—sometimes stimulation during encoding improves later recall, sometimes it impairs it, and the effects are highly variable across individuals and brain regions. What do those inconsistent results tell us?
Dr. Leigh Hochberg
They tell us that we don't understand what stimulation is doing at the circuit level. We're applying electrical fields that activate or suppress large populations of neurons, but we don't know which neurons, what their functional roles are, or how the stimulation interacts with ongoing activity. It's like trying to improve a computer program by randomly flipping bits. Occasionally you might get lucky, but usually you just corrupt the computation.
Adam Ramirez
Looking ahead, what technological advances would be necessary to make memory BCIs feasible? Is this a hundred-year problem or a ten-year problem?
Dr. Krishna Shenoy
We'd need several breakthroughs. First, recording technologies that scale to tens of thousands of neurons across multiple brain regions with stable signals over decades. Second, theoretical understanding of memory representation sufficient to build decoders. Third, stimulation technologies with single-neuron precision and temporal control. Fourth, validation that stimulation can actually induce plasticity that forms memories. I'd say we're at the very beginning of understanding these problems. Decades is more realistic than years.
Dr. Leigh Hochberg
I agree with that timeline for general memory read-write. But there might be narrower applications that come sooner. For example, detecting specific pathological brain states—early signs of seizures, sleep disorders, attentional lapses—and delivering targeted interventions. Those are still pattern recognition problems on neural data, but they don't require understanding memory representation. They might provide clinical value while we work on the harder problems.
Jennifer Brooks
Both of you, thank you for providing a realistic assessment of where BCI technology stands and what the actual roadblocks are.
Dr. Krishna Shenoy
Thank you.
Dr. Leigh Hochberg
Pleasure to be here.
Adam Ramirez
That's our program for this evening. Until tomorrow, stay critical.
Jennifer Brooks
And keep questioning. Good night.