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 synaptic scaling and network homeostasis—mechanisms that allow neural circuits to maintain stable function despite continuous learning and environmental fluctuations. When you train an artificial neural network, you typically initialize the weights to small random values and then adjust them through gradient descent until the network performs well on your task. But there's no built-in mechanism to prevent weights from growing without bound or shrinking to zero if the training dynamics push in those directions. You have to add explicit regularization—weight decay, dropout, normalization layers—to keep the network stable. Biological networks face an analogous problem. Hebbian plasticity mechanisms like long-term potentiation strengthen synapses based on correlated activity, but unchecked strengthening would drive neurons to saturation. How do biological circuits avoid runaway excitation while still learning?

Jennifer Brooks This is where homeostatic plasticity comes in. Unlike Hebbian mechanisms that depend on input correlations, homeostatic plasticity responds to the overall activity level of the postsynaptic neuron. If a neuron is firing too much, homeostatic mechanisms downscale its synapses globally. If it's firing too little, they upscale. This keeps the firing rate in a functional range regardless of what Hebbian learning is doing to individual synapses. The question is how this works mechanistically. What signals trigger scaling? How does the neuron sense its own activity level over the relevant timescales? How does it modulate synaptic strength uniformly across thousands of synapses? And does this mechanism actually operate during natural learning, or is it primarily a response to extreme perturbations?

Adam Ramirez To explore these questions, we're joined by Dr. Gina Turrigiano, a neuroscientist at Brandeis University whose research discovered and characterized synaptic scaling. Her work combines electrophysiology, imaging, and molecular biology to understand how neurons regulate their excitability and maintain stable network function. Dr. Turrigiano, welcome.

Dr. Gina Turrigiano Thank you. It's good to be here.

Jennifer Brooks Let's start with the basic phenomenology. What is synaptic scaling, and how did you discover it?

Dr. Gina Turrigiano Synaptic scaling is a form of homeostatic plasticity where neurons adjust the strength of all their excitatory synapses up or down to compensate for changes in overall network activity. We discovered it by blocking activity in cultured neurons for several days and then measuring synaptic strength. We found that when you silence a network with tetrodotoxin, which blocks action potentials, the neurons respond by upscaling their excitatory synapses. The amplitude of miniature excitatory postsynaptic currents—these are responses to spontaneous neurotransmitter release—increases by about a factor of two. Conversely, if you increase network activity with drugs that enhance excitation, synapses downscale. The effect is multiplicative and global—all excitatory synapses on a neuron scale by roughly the same factor.

Adam Ramirez Multiplicative scaling is computationally significant. If you scale all synapses by the same factor, you preserve their relative weights. Strong synapses stay strong relative to weak ones, just shifted up or down uniformly. This means scaling doesn't erase information encoded in the pattern of synaptic weights—it just adjusts the overall gain. Is that the right interpretation?

Dr. Gina Turrigiano That's correct. Multiplicative scaling preserves the rank order of synaptic strengths and their ratios. This is important because information is thought to be stored in the relative strengths of synapses, not their absolute values. If homeostasis worked by uniformly adding or subtracting the same amount to every synapse, it would compress the dynamic range and eventually erase weak synapses. Multiplicative scaling avoids this problem. It's like adjusting the volume on an audio signal—you change the amplitude but preserve the waveform.

Jennifer Brooks What's the timescale? Hebbian plasticity can happen in seconds to minutes. How long does it take for scaling to detect a change in activity and respond?

Dr. Gina Turrigiano Scaling operates on slower timescales—hours to days. In our initial experiments, we saw clear scaling after twenty-four to forty-eight hours of activity manipulation. This makes sense functionally. Scaling is meant to compensate for persistent changes in network activity, not transient fluctuations. If scaling responded too quickly, it would interfere with rapid learning. The slow timescale allows Hebbian plasticity to do its job of adjusting specific synapses based on correlation structure, while scaling operates in the background to keep overall excitability from drifting.

Adam Ramirez How does a neuron sense its own firing rate over these long timescales? You need some kind of integrator that accumulates activity and triggers scaling when the integrated signal crosses a threshold.

Dr. Gina Turrigiano The exact sensing mechanism is still being worked out, but calcium is a strong candidate. Calcium influx through voltage-gated channels during action potentials provides a direct readout of firing rate. Neurons have calcium-sensitive signaling pathways that could integrate this signal over hours to days and adjust gene expression or receptor trafficking accordingly. We've shown that blocking calcium signaling can prevent scaling, which supports this idea. But there may be multiple sensors operating in parallel—some sensitive to somatic firing, others to dendritic activity or synaptic input.

Jennifer Brooks Once the neuron detects a deviation from its target firing rate, what are the effector mechanisms? How does it actually change synaptic strength?

Dr. Gina Turrigiano Scaling involves changes in the number of postsynaptic receptors at excitatory synapses. When synapses scale up, more AMPA receptors are inserted into the postsynaptic membrane. When they scale down, receptors are removed. This is regulated through receptor trafficking—endocytosis and exocytosis. The neuron adjusts the expression of proteins involved in receptor insertion and anchoring, which changes the steady-state number of receptors at synapses. There are also changes in presynaptic release probability in some cases, and adjustments to intrinsic excitability—the neuron can change its membrane conductances to become more or less excitable overall.

Adam Ramirez This sounds complex. You've got a feedback control system with a slow integrator, a comparison to a set point, and multiple effector pathways. How robust is this system to parameter variations and perturbations?

Dr. Gina Turrigiano It's quite robust. Neurons can maintain stable firing rates across a wide range of input conditions and network states. There's also diversity in the mechanisms—different neuron types use different combinations of scaling and intrinsic plasticity to achieve homeostasis. This is reminiscent of the degeneracy we've discussed before. Multiple mechanisms can achieve the same functional outcome, which provides redundancy and flexibility. If one pathway is compromised, others can compensate.

Jennifer Brooks Does synaptic scaling actually happen during learning in intact animals, or is it primarily a response to artificial manipulations like blocking activity with drugs?

Dr. Gina Turrigiano This is a critical question. Most of the mechanistic work has been done in cultured neurons or reduced preparations where we can control activity precisely. Demonstrating that scaling operates during natural learning in vivo is harder. But there's accumulating evidence that it does. Sensory deprivation studies show homeostatic changes. If you close one eye during development, the deprived neurons initially lose responsiveness, but over days they scale up their synapses to compensate. Sleep may also involve homeostatic downscaling to offset synaptic strengthening that occurs during waking. The evidence is still being built, but it's consistent with scaling playing a role in maintaining stability during ongoing plasticity.

Adam Ramirez How does scaling interact with Hebbian plasticity? Are they completely independent, or do they influence each other?

Dr. Gina Turrigiano They interact. Hebbian plasticity changes the pattern of synaptic weights—strengthening some connections, weakening others. This can shift the overall distribution of synaptic strengths and alter the neuron's firing rate. Scaling then responds to this change in firing rate and adjusts the gain to bring it back to target. In principle, you could have a dynamic equilibrium where Hebbian mechanisms continuously adjust relative weights, and scaling continuously adjusts the gain to maintain stable firing. The two processes operate on different timescales and respond to different signals, but they're part of a coupled system.

Jennifer Brooks What happens if you disable scaling? Do networks become unstable?

Dr. Gina Turrigiano Yes. We've used genetic approaches to disrupt scaling mechanisms, and the result is often network instability. Neurons can become hyperexcitable, leading to epileptiform activity. Or they can become hypoexcitable and unresponsive. The network loses its ability to self-regulate. This suggests that homeostatic mechanisms are essential for stable network function, especially in the face of ongoing plasticity. Without homeostasis, Hebbian learning would drive the network into pathological states.

Adam Ramirez This raises the question of whether artificial neural networks need something like synaptic scaling. We use regularization techniques like batch normalization, layer normalization, and weight decay. Are these analogous to biological homeostasis, or are they solving different problems?

Dr. Gina Turrigiano There are similarities. Batch normalization normalizes activations to have stable statistics across layers, which prevents activations from growing or shrinking uncontrollably during training. This is functionally similar to homeostatic regulation of firing rates. Weight decay shrinks weights toward zero, which prevents them from growing without bound. But the motivation and implementation differ. In artificial networks, these techniques are primarily designed to improve optimization—making gradient descent converge faster and more reliably. In biological networks, homeostasis is about maintaining functional operating ranges in the face of continuous perturbations. The computational goals overlap, but the constraints are different.

Jennifer Brooks Are there network-level homeostatic mechanisms beyond single-neuron scaling? Can networks regulate their overall excitability collectively?

Dr. Gina Turrigiano There's evidence for network-level regulation, though it's less well understood. Inhibitory neurons play a key role. Changes in excitatory drive onto inhibitory neurons can trigger homeostatic adjustments that stabilize network activity. There's also evidence that excitatory and inhibitory synapses scale in coordinated ways to maintain excitatory-inhibitory balance. If you perturb excitation, both excitatory and inhibitory synapses can scale together to preserve the ratio. This prevents runaway excitation or complete silencing. The network essentially has multiple feedback loops operating at different levels—individual neurons, local circuits, and broader networks.

Adam Ramirez How specific is scaling? You mentioned it's global—all excitatory synapses on a neuron scale together. But are there forms of homeostatic plasticity that are more synapse-specific?

Dr. Gina Turrigiano Yes. While classical synaptic scaling is global, there are other homeostatic mechanisms that can be more selective. For example, neurons can adjust the strength of inhibitory synapses independently of excitatory ones. There's also evidence for local homeostatic regulation within dendritic branches—a single dendrite can regulate its own excitability based on local activity. And some forms of metaplasticity—plasticity of plasticity—can adjust the threshold for Hebbian plasticity in a synapse-specific manner. So homeostasis isn't just a single mechanism. It's a family of regulatory processes operating at different spatial scales.

Jennifer Brooks What are the molecular targets for therapeutic intervention? If homeostatic mechanisms are disrupted in neurological disorders, could restoring them be beneficial?

Dr. Gina Turrigiano Disrupted homeostasis has been implicated in epilepsy, autism spectrum disorders, and neurodegenerative diseases. In epilepsy, for instance, there's evidence that homeostatic downscaling fails to compensate for increased excitability, allowing seizures to develop. In some forms of autism, there may be excessive or insufficient scaling that disrupts the balance between excitation and inhibition. Therapeutically, you could imagine targeting the sensors—manipulating calcium signaling to adjust the set point for homeostatic regulation. Or targeting the effectors—modulating receptor trafficking or intrinsic conductances. But this is speculative. We need a better understanding of how homeostasis is disrupted in specific disorders before we can design rational interventions.

Adam Ramirez Looking at computational models, have researchers implemented synaptic scaling in artificial networks, and does it improve performance or stability?

Dr. Gina Turrigiano There have been some attempts. Researchers have implemented homeostatic plasticity rules in spiking neural networks and shown that it can stabilize learning and prevent runaway dynamics. In some cases, it allows the network to learn continuously without catastrophic forgetting, because scaling helps maintain a functional operating range even as weights change. But these models are still relatively simple compared to biological networks. Most deep learning systems don't use explicit homeostatic mechanisms—they rely on normalization layers and regularization, which serve some of the same functions but aren't mechanistically analogous. There's potential for cross-pollination here. Biologically inspired homeostatic rules might offer new ways to stabilize training in recurrent networks or improve continual learning.

Jennifer Brooks What are the open questions? Where does the field need to go next?

Dr. Gina Turrigiano We need better tools for measuring homeostatic plasticity in vivo during natural behavior. Most of our mechanistic understanding comes from reduced preparations. We need to know whether scaling operates during learning, sleep, development, and aging in intact brains. We also need to understand the diversity of homeostatic mechanisms—which neuron types use which mechanisms, and why. And we need to connect homeostasis to circuit function. How does maintaining stable firing rates at the single-neuron level contribute to stable computations at the network level? Finally, there's the question of pathology. How do failures of homeostasis contribute to disease, and can we develop therapies that restore homeostatic balance?

Adam Ramirez Dr. Turrigiano, thank you for explaining how biological networks achieve stability without sacrificing plasticity.

Dr. Gina Turrigiano My pleasure. These mechanisms are fundamental to how the brain works.

Jennifer Brooks That's our program. Until tomorrow, stay critical.

Adam Ramirez And keep questioning. Good night.