Attention & Consciousness

Derivation 1: The Normalization Model of Attention

Reynolds and Heeger (2009) proposed that attention acts through an "attention field" that modulates neural responses before divisive normalization. The response of neuron$i$ is:

$$r_i = \frac{(\mathbf{a} \cdot \mathbf{s})_i \cdot E_i}{\sigma^2 + \sum_j (\mathbf{a} \cdot \mathbf{s})_j \cdot w_{ij}}$$

where $\mathbf{s}$ is the stimulus drive, $\mathbf{a}$ is the attention field (a spatial or feature-based gain), $E_i$ is the excitatory drive for neuron $i$,$w_{ij}$ are normalization weights, and $\sigma^2$ is a semi-saturation constant. The key insight is that the same mechanism produces different phenomenology depending on the relative sizes of the attention field and stimulus:

• Narrow attention field (smaller than stimulus): contrast gain, shifting the contrast-response function leftward
• Broad attention field (larger than stimulus): response gain, multiplicatively scaling the entire response
• Matched sizes: intermediate effects combining both gain changes

This explains why different experiments report different attentional effects — the mechanism is the same, but the relative geometry varies.

Derivation 2: Temporal Binding via Neural Synchrony

The temporal correlation hypothesis (von der Malsburg, 1981; Singer and Gray, 1995) proposes that neurons representing features of the same object synchronize their firing in the gamma band (30–80 Hz). The coherence between two spike trains $x(t)$ and$y(t)$ at frequency $f$ is:

$$C_{xy}(f) = \frac{|S_{xy}(f)|^2}{S_{xx}(f) \cdot S_{yy}(f)}$$

where $S_{xy}(f)$ is the cross-spectral density. To quantify binding, consider$N$ neurons organized into two groups (objects). Within-group synchrony is:

$$\Phi_{\text{within}} = \frac{1}{|G|^2}\sum_{i,j \in G} C_{ij}(f_\gamma)$$

For binding to work, within-object coherence must significantly exceed between-object coherence. The required precision of synchronization is approximately$\Delta t < 1/(2f_\gamma)$, or about 7–15 ms for gamma oscillations. This temporal window matches the STDP learning rule window, suggesting that synchrony-based binding may also drive the formation of object representations through Hebbian learning.

Derivation 3: Feature Integration Theory and the Saliency Map

Treisman's Feature Integration Theory (1980) proposes that features are first processed in parallel "feature maps," then bound through focal attention. The computational implementation uses a saliency map. For feature dimension $k$(color, orientation, etc.) and location $(x, y)$:

$$S_k(x, y) = \sum_{c=1}^{C} |F_k^c(x, y) - \bar{F}_k^c|$$

where $F_k^c$ is the feature map at scale $c$ and $\bar{F}_k^c$ is its mean. The combined saliency map integrates across feature dimensions:

$$S(x, y) = \sum_{k=1}^{K} w_k \cdot \hat{S}_k(x, y)$$

where $\hat{S}_k$ is the normalized conspicuity map and $w_k$ are learned or task-dependent weights. A winner-take-all network selects the most salient location for the next fixation. Top-down attention modulates the weights $w_k$, implementing feature-based attention (e.g., biasing toward color when searching for a red target).

Derivation 4: Global Workspace Theory (GWT)

Baars (1988) and Dehaene et al. (1998) proposed that consciousness arises when information is "broadcast" to a global workspace formed by long-range cortical connections. Mathematically, the workspace can be modeled as a dynamical system with an ignition threshold. The activity of workspace neuron $i$ is:

$$\tau \frac{dx_i}{dt} = -x_i + f\left(\sum_j W_{ij}^{\text{local}} x_j + \sum_k W_{ik}^{\text{global}} x_k + I_i^{\text{ext}}\right)$$

The system has two stable states: (1) a low-activity "subliminal" state where information remains in local modules, and (2) a high-activity "ignited" state where global broadcasting occurs. The transition happens when the input exceeds a threshold set by the global coupling strength:

$$I_{\text{crit}} = \frac{1 - W^{\text{local}} f'(0)}{W^{\text{global}} f'(0)}$$

This predicts the all-or-none nature of conscious access: stimuli either reach awareness (ignition) or remain subliminal, with no intermediate states. The model explains the P3b ERP component, the "late ignition" observed ~300 ms post-stimulus in conscious perception, and the role of prefrontal-parietal networks in broadcasting.

Derivation 5: Integrated Information Theory (IIT)

Tononi's IIT (2004, 2008) proposes that consciousness is identical to integrated information, $\Phi$. For a system of $N$ binary elements in state$\mathbf{x}$ with transition probability matrix $\mathbf{T}$:

$$\Phi = \min_{\text{partition } P} \left[ D_{\text{KL}}\left(p(\mathbf{x}_{t+1} | \mathbf{x}_t) \,\|\, \prod_k p(\mathbf{x}_{t+1}^k | \mathbf{x}_t^k)\right)\right]$$

where the minimum is taken over all possible bipartitions $P$ of the system, and $D_{\text{KL}}$ is the Kullback-Leibler divergence. Intuitively,$\Phi$ measures how much the whole system's behavior exceeds what its parts can account for independently. For a simple 2-element system with coupling matrix:

$$\Phi_{\text{2-element}} = I(\mathbf{x}_{t+1}; \mathbf{x}_t) - \max_P \sum_k I(\mathbf{x}_{t+1}^k; \mathbf{x}_t^k)$$

IIT makes specific predictions: the cerebellum, despite having more neurons than the cerebrum, contributes little to consciousness because its feedforward architecture generates low $\Phi$. The thalamocortical system, with dense recurrent connectivity, generates high $\Phi$. Computing $\Phi$ exactly is NP-hard, but approximations based on spectral analysis of the connectivity matrix are tractable.