As a scientist, I regularly deal with problems regarding signal-to-noise ratio. The signal that I’m interested in is always accompanied by noise, and depending on the signal of interest, the contribution of noise may vary from negligible to substantial. One colleague of mine would even claim that a great amount of experimental results can be explained by noise alone. I agree; however, this claim doesn’t always entail negativity.  While we use the term noise as an umbrella for all types of signals we are not interested in, noise can be studied on its own. I found this book while strolling across the aisles in my library; the title sounded interesting: “Neuronal Noise” by Alain Destexhe and Michelle Rudolph-Lilith. The book fully lived up to my expectations.

Neuronal activity and information transduction in neuronal networks are based on the movement of charged ions across the cell membrane. Ions flow in and out of the neuron via ion channels (essentially pores in the membrane). At rest, neurons constantly maintain an asymmetrical distribution of charges inside themselves in comparison with outside in the extracellular fluid. At rest, the membrane of a neuron is negatively charged. When active, neurons transmit information to other neurons via action potentials. An action potential is a very rapid redistribution of charges inside and outside the neuron and, as a consequence, a change in membrane potential. After “firing” an action potential, the neuron seeks to return to rest by reestablishing the distribution of charged ions.

A neuron, as a cell, can be divided into three compartments. 1. Soma: the cell body, which gathers and integrates signals. 2. Dendrites: branches from the soma that collect information. 3. Axon: a branch from the soma that sends information by “firing” an action potential. The length of axons varies a lot; the longest axons in humans, which run from the base of the spine down to the big toe, stretch 1.5 meters. Dendrites of one neuron branch substantially, especially in the cerebellum, where Purkinje cells’ dendrites are branching through roughly 20–30 hierarchical levels. Axons of one neuron approach the dendrites of another.

Communication between neurons occurs at places called synapses. In a synapse, the sending (or “firing”) neuron’s axon comes in close proximity to the receiving neuron’s dendrite. The sending neuron releases molecules into the extracellular space between neurons. These molecules either directly or indirectly affect the ion flow across the membrane of the receiving neuron. For example, molecules may bind to proteins on the receiving neuron membrane and, via a chain of events, cause the opening of ion channels in the membrane of the receiving neuron. If the dendrite accumulates a sufficient amount of inputs from outside, the signal propagates to the soma. If the soma accumulates a sufficient amount of input from dendrites, the neuron will “fire” an action potential via its axon. And the same communication round will be repeated somewhere else. Synaptic inputs are the primary means by which neurons communicate with one another.

Many events in the aforementioned processes can be noisy. This book is about synaptic noise. This type of noise is characterized by a nonstop and discordant “bombardment” of neurons by irregular synaptic inputs from other neurons. It is as if communication goes on but without a message. Synaptic noise is a chatter with no meaning.

Turns out synaptic noise is a necessary component of neuronal communication. Without synaptic noise, neurons are unable to operate in the same way. There are five unusual and brilliant ways that synaptic noise influences communication. 

1. Increased sensitivity. In the absence of background activity, neurons only react to sufficiently strong stimuli. The response probability in this case (deterministic neuron model) is a step-wise function: below a certain input amplitude, there is no reaction, and above, a 100% reaction. Moreover, without background activity, there is little to no variability: all neurons react when the synaptic input arrives without a delay between each other. When background activity is present, neuronal responses are more nuanced. The response probability function is more sigmoidal in shape, meaning that even small-amplitude inputs have a chance of being detected. Moreover, the timing of the response of the population of neurons is different; some neurons will fire with a delay. Thus, the presence of synaptic noise leads to enhanced sensitivity and an increased range of detected inputs.

2. Neuronal code. Two theories have been put forward about the regime in which neurons code information. These regimes are not mutually exclusive. One regime is that neurons work as coincidence detectors. When two or more inputs arrive at the same time to a neuron, that neuron will “fire” and excite other neurons downstream. Another regime is that neurons detect a temporal profile in the input and integrate it over time. In other words, in one regime, neurons process the precise timing of inputs and, in another, rates. Simulation studies demonstrated that synaptic noise modulates neuronal processing in both of these regimes. As responses of a network are more variable with a larger amount of synaptic noise (see 1 above), coincidence detection is directly influenced by the amount of noise. Also, in the absence of noise, neurons are reliable in detecting rates of input for smaller input strengths, while with noise, higher strengths are required to be integrated. At the same time, smaller inputs are integrated with better reliability in noisy conditions than in non-noisy conditions. Overall, these results suggest that noise is a part of information processing in neuronal networks. 

3. Location independence. If the axon contacts the dendrite in the proximity of the soma, seemingly, the signal needs to travel a shorter distance than if the synapse were made at the distal part of a dendrite. However, in the presence of background noise, synapse location does not affect the efficacy of a synapse. It is still true that the chance that the dendritic input propagates all the way to the soma is higher for synapses that are closer to the soma. But in addition to that, synaptic noise increases the chance of dendritic response (more nuanced processing, see 1 above), and more so for distal locations. In distal parts of a dendrite, its diameter is smaller, and input resistance is higher (less surface, fewer ion channels, a larger effect if only a small influx of ions occurs). These two effects balance each other, and synaptic input effectiveness becomes independent of the location of the synapse on the dendrite.

4. Inhibitory nature. Conductance is a measure of the flow of ions through the membrane. Inhibitory conductance is the flow of ions that decreases the probability of a neuron “firing”. Excitatory conductance, in contrast, is the flow of ions that increases the probability of a neuron “firing”. In vivo, inhibitory conductance was measured to be 5-20 times larger than excitatory conductance, and fluctuations in inhibitory conductance are larger.  From this it follows that synaptic noise mostly depends on inhibitory conductance fluctuations. At the same time, the initiation of an action potential was correlated with an increase in excitatory conductance and with a decrease in inhibitory conductance. It means that synaptic noise modulations, to a certain extent, determine action potentials.

5. External inputs only modulate network activity. Experiments with awake animals demonstrated that intrinsic activity (not related to the stimulus) explains the bulk of variance in spatial and temporal characteristics of neuronal responses to the stimulus. To reiterate, not so much that properties of the stimulus explain stimulus-evoked activity, but neuronal network activity unrelated to the stimulus. Another finding is that most of the action potentials are initiated from the recurrent feedback intrinsic activity, not external stimuli. These observations, together with the omnipresence of synaptic noise in different animals, suggest that sensory (visual, auditory, tactile) inputs do not orchestrate neuronal activity but rather only modulate it.

The last claim is quite thought-provoking. In my research, I usually ask questions about how intrinsic activity affects perception or, equivalently, how brain state influences perceptual outcomes. Asking how stimuli modulate brain dynamics is turning everything upside down. It is a new way of thinking for me, but I will definitely think about it more.

The book is interesting but becomes technical quickly. Many studies are simulation studies, so the book covers details of simulations and the math behind them. The book will be easy to grasp for someone with experience in computational modeling or simulations, but it is also accessible for a general reader since the main arguments do not require a deep understanding of simulations. There are also a lot of references for further reading. I liked the book and wholeheartedly recommend it.

Favorite quote:

“Noisy networks can sustain a faithful propagation of firing rates across successive layers. These results are particularly interesting because noise allows populations of neurons to relay a signal across successive layers without attenuation or prevents a catastrophic invasion of synchronous activity“

July, 2026