Dynamics of specialization in neural modules under resource constraints

TL;DR

This work interrogates whether structural modularity suffices for functional specialization in neural systems by using controlled two-module RNNs trained on a parity task. It introduces three functional specialization metrics and an adapted directed-Q measure to separate structure from function, then systematically varies environment structure and resource constraints to map where specialization emerges. The key finding is that modular structure alone does not guarantee specialization; specialization arises under meaningful environmental separability and under strong resource constraints, with the dynamics of information flow further shaping trajectories of specialization over time. The study highlights the importance of dynamic, context-dependent notions of modularity for neuroscience and brain-inspired engineering, and proposes stress-tested, simplified scenarios as a productive path toward robust definitions of functional modularity for complex systems.

Abstract

It has long been believed that the brain is highly modular both in terms of structure and function, although recent evidence has led some to question the extent of both types of modularity. We used artificial neural networks to test the hypothesis that structural modularity is sufficient to guarantee functional specialization, and find that in general, this doesn't necessarily hold. We then systematically tested which features of the environment and network do lead to the emergence of specialization. We used a simple toy environment, task and network, allowing us precise control, and show that in this setup, several distinct measures of specialization give qualitatively similar results. We further find that in this setup (1) specialization can only emerge in environments where features of that environment are meaningfully separable, (2) specialization preferentially emerges when the network is strongly resource-constrained, and (3) these findings are qualitatively similar across the different variations of network architectures that we tested, but that the quantitative relationships depend on the precise architecture. Finally, we show that functional specialization varies dynamically across time, and demonstrate that these dynamics depend on both the timing and bandwidth of information flow in the network. We conclude that a static notion of specialization, based on structural modularity, is likely too simple a framework for understanding intelligence in situations of real-world complexity, from biology to brain-inspired neuromorphic systems. We propose that thoroughly stress testing candidate definitions of functional modularity in simplified scenarios before extending to more complex data, network models and electrophysiological recordings is likely to be a fruitful approach.

Figures (7)

• Figure 1: Summary of methods. A. Schematic of input, model and task. Two digits $\mathcal{D}_1$ and $\mathcal{D}_2$ are fed to the network via either shared input weights (in which case each module sees each digit) or separate (in which case each module sees only a single digit). Two recurrent neural network modules of size $n$ then process the inputs, and communicate via a sparse interconnection containing a fraction $p$ of all $n^2$ possible connections. Outputs are computed either through shared readout (both modules feed into the same readout) or separate (in which case each module has its own readout, and the decision is based on a max function). A bottleneck consisting of an additional layer of 5 neurons can be placed between the modules and the final readout. B. All possible architectural choices. C. The three functional specialization metrics: Module Probing, where the trained network is frozen and a separate (probing) network is trained to extract digit identity from each module's hidden state; Module Ablation, where the trained network is frozen and a single network is trained to extract digit identity from both modules' hidden states, after which one is ablated and impact on performance is measured; and Correlation Analysis, where hidden states are correlated holding one digit fixed and varying the other.
• Figure 2: Levels of functional specialization of networks with (A) varying modularity $Q$, or equivalently (B) inter-module connectivity $p$, measured by three different metrics (columns). All metrics indicate a similar trend, sharply rising only at extreme levels of $Q$ modularity. Data are presented as mean values with a shaded standard error envelope.
• Figure 3: Functional specialization of networks with varying module size $n$, number of active synapses in the communication layer $p_s$, in an environment where input variables present covariance $c$. Each image shows functional specialization when varying two of these parameters (averaged across the third one), with color indicating the degree of functional specialization on a log scale, cut off at a mimal value of 0.01. Results are shown for three different architectures: (A.) separate pathways, (B.), fused pathways, and (C.) shared pathway.
• Figure 4: Specialization dynamics. A. Specialization of different parts of the network at different time-steps. The network unrolled through time is shown below, with inputs presented to modules at time $t=0$ at the left hand side, and subsequent time steps moving towards the right finishing in the readout layer at the final time step. In the plot above is shown the specialization of the corresponding module at a given time, where specialization close to 1 means full specialization on digit 0, close to -1 means full specialization on digit 1, and close to 0 means no specialization. The dark dashed lines in both the network diagram and the plot show the timing of inter-module communication. Each colored line represents networks with different inter-connection levels $p$. We vary the communication timings in the inputs, and find specialization collapses when communicating at high bandwidth ($p$ high). B. With dynamic noise in the inputs, specialization drops continually, especially in high noise setting. C. With stochasticity in the input dynamics (digits 0 and 1 turned on at $t_0$ and $t_1$ respectively, indicated by colored arrows), specialization dynamics closely follows the inputs' dynamics. Networks showed with 1 (dark blue), 10 (teal) and 100 (yellow) active inter-module connections. Data are presented as mean values with a shaded standard error envelope.
• Figure 5: Accuracy for networks composed of 25-neurons modules, with varying levels of inter-module communication $p$. Results are shown for 5 independent experiments per parameter-set (sparsity), with confidence intervals showed as shaded areas.