Balanced amplification

An interactive walkthrough of Murphy & Miller (2009), Balanced Amplification: A New Mechanism of Selective Amplification of Neural Activity Patterns, Neuron 61, 635–648.

The puzzle

Cortical networks have strong recurrent excitation, balanced by similarly strong feedback inhibition. In cat V1, even spontaneous activity (no stimulus) contains structure that resembles evoked orientation maps — about 2× larger in those directions than random control patterns (Kenet et al., 2003). This implies the circuit selectively amplifies some patterns more than others.

The classical story is Hebbian amplification: a pattern excites itself through recurrent connections, slowing its decay and so accumulating to larger amplitude. But Hebbian amplification couples amplitude tightly to timescale — the more you amplify, the more you slow down. That is at odds with the ~80 ms timescale observed in V1.

Murphy & Miller introduce a different mechanism: balanced amplification. The key insight is that biological connectivity matrices are nonnormal (because individual neurons are either excitatory or inhibitory). Nonnormal matrices have non-orthogonal eigenvectors, and that geometry hides a feedforward connection between difference patterns (where E and I rates differ) and sum patterns (where they move together). Small imbalances in E/I activity drive large balanced responses — without slowing.

Fig 1: the two-population case

The simplest case is two populations, one excitatory and one inhibitory, each projecting to both. With a single recurrent strength w and an inhibition ratio k_I, the dynamics decompose neatly in the sum/difference basis. The difference mode decays passively; the sum mode receives feedforward drive from the difference mode with weight w_FF = w(1 + k_I). Try setting r_E(0) = 1 and r_I(0) = 0 — an "all-excitatory" perturbation — then watch how the sum (blue) transiently grows even as the difference (black) decays.

Figure 1

The two-population balanced circuit

Panel A — circuit

Green = excitatory ( $w$ )
Red = inhibitory ( $- w k_{I}$ )
Each cell projects to itself and the other.
Cell fill = live activity at t = 0.00τ
rE = 1.00, rI = 0.00

Panel B (top) — E & I activity at each marked time

t = 0.0τ

1.000.00

t = 0.4τ

1.220.55

t = 0.8τ

1.160.71

t = 1.5τ

0.840.61

t = 3.0τ

0.290.24

−1.9+1.9activity (rE, rI; deviation from baseline)

Preset: E pulse — Paper's example — imbalanced. Hover the plot to read values at any time.

Feedforward weight

w_{F F} = w (1 + k_{I}) = 4.490

Sum self-weight

w_{+} = w (k_{I} - 1) = 0.214

Non-zero eigenvalue

λ = - 0.214

What to look for: Switch between E pulse and Both equal presets — same total starting activity, but only the imbalanced one drives a transient overshoot in the sum (blue). That's the central message: the imbalance between E and I is what gets amplified, not their absolute activity. Then crank

w

up with

k_{I}

just above 1: the sum's peak grows, but the timescale barely moves. That is balanced amplification.

Fig 2: Hebbian vs. balanced

To see why balanced amplification matters, compare two networks tuned to the same steady-state amplification. The Hebbian network is a single excitatory population with self-strength w; its effective timescale is τ/(1 − w), which diverges as w approaches 1. The balanced network is the two-population circuit from Fig 1.

Use the preset selector to walk from 1× → 10× amplification. The Hebbian network slows down dramatically. The balanced network barely changes its timecourse — only its height.

Figure 2

Hebbian vs. balanced amplification

Hebbian (single E pop)

Balanced (E + I)

The Hebbian network (left) achieves amplification by slowing down: as

w \to 1

the decay timescale

τ / (1 - w)

diverges. The balanced network (right) achieves the same steady-state amplification with essentially the same dynamics as the unrecurrent case — only the height of the response changes, not its width.

Fig 3: spatially extended networks

In a real cortex you don't have two lumped populations — you have thousands of neurons with structured connectivity. In V1, intracortical connections depend on both distance and orientation preference. With that structure, balanced amplification picks out specific spatial patterns: those whose pattern of E/I difference best reproduces itself through W_E + W_I.

The top sum/difference mode pairs (the eigenvectors of W_E + W_I with the largest eigenvalues) are the most amplifiable patterns. In a network with V1-like connectivity, modes 2 and 3 happen to look exactly like evoked orientation maps — which is precisely what's observed in spontaneous V1 activity.

Figure 3

Spatial network: orientation map + sum/difference modes

Loading modes.json…

Fig 4: why the eigenvector picture deceives

All this is hidden if you only look at eigenvalues. The eigenvalues of the balanced matrix can all have negative real part — yet the matrix transiently amplifies. The trick is that the eigenvectors are non-orthogonal, so the "amplitude" of an eigenmode is not directly observable as a firing rate. The orthonormal Schur basis makes the hidden feedforward connectivity explicit: same eigenvalues on the diagonal, but with feedforward weights above.

Figure 4

Eigenvector basis vs. Schur basis

Eigenvector picture

Schur (orthonormal) basis

Limit: pure feedforward

Eigenvalues

[-0.030, -0.060, 0.090]

Schur off-diagonals (feedforward weights)

T₁₂ = 1.40, T₁₃ = 0.35, T₂₃ = 0.84

For a non-normal matrix, the eigenvector basis (left) hides feedforward connections between activity patterns. The Schur basis (middle) is orthonormal and exposes those feedforward weights as the upper-triangular entries. As the matrix becomes more "balanced" (eigenvalues → 0), the self-loops shrink and the dynamics become essentially feedforward (right).

Take-aways

Strong recurrent excitation + strong feedback inhibition → nonnormal connectivity → hidden feedforward structure.
Small E/I imbalances ("difference patterns") drive large balanced responses ("sum patterns") — selective amplification without slowing.
Which patterns are amplified is controlled by the eigenvectors of W_E + W_I, not eigenvalues.
This is likely ubiquitous in cortex and can quantitatively account for the spontaneous orientation-map structure observed in cat V1.

References

Murphy, B. K., & Miller, K. D. (2009). Balanced amplification: A new mechanism of selective amplification of neural activity patterns. Neuron, 61(4), 635–648.
Kenet, T., Bibitchkov, D., Tsodyks, M., Grinvald, A., & Arieli, A. (2003). Spontaneously emerging cortical representations of visual attributes. Nature, 425, 954–956.
Christodoulou, S., & Vogels, T. (2022). The eigenvalue value (in neuroscience). [Companion reading for 200C.]