[Paper Review] Context-sensitive neocortical neurons transform the effectiveness and efficiency of neural information processing

Khubaib Ahmed, Ahsan Adeel+2 more|arXiv (Cornell University)|Jul 15, 2022

Advanced Memory and Neural Computing4 citations

TL;DR

This paper proposes a biologically inspired deep learning architecture using context-sensitive two-point neocortical neurons that integrate feedforward inputs with contextual modulation from apical dendrites. By amplifying relevant signals and suppressing irrelevant ones based on contextual input, the model achieves superior efficiency and robustness in processing complex, noisy real-world data—demonstrated through audiovisual speech denoising with reduced neural activity and enhanced resilience to component failure.

ABSTRACT

Deep learning (DL) has big-data processing capabilities that are as good, or even better, than those of humans in many real-world domains, but at the cost of high energy requirements that may be unsustainable in some applications and of errors, that, though infrequent, can be large. We hypothesise that a fundamental weakness of DL lies in its intrinsic dependence on integrate-and-fire point neurons that maximise information transmission irrespective of whether it is relevant in the current context or not. This leads to unnecessary neural firing and to the feedforward transmission of conflicting messages, which makes learning difficult and processing energy inefficient. Here we show how to circumvent these limitations by mimicking the capabilities of context-sensitive neocortical neurons that receive input from diverse sources as a context to amplify and attenuate the transmission of relevant and irrelevant information, respectively. We demonstrate that a deep network composed of such local processors seeks to maximise agreement between the active neurons, thus restricting the transmission of conflicting information to higher levels and reducing the neural activity required to process large amounts of heterogeneous real-world data. As shown to be far more effective and efficient than current forms of DL, this two-point neuron study offers a possible step-change in transforming the cellular foundations of deep network architectures.

Motivation & Objective

• To address the high energy consumption and inefficiency of current deep learning models, which rely on integrate-and-fire point neurons that process all inputs uniformly regardless of context.
• To explore whether mimicking context-sensitive neocortical neurons—where apical dendrites modulate basal input transmission based on contextual relevance—can improve processing efficiency and effectiveness.
• To demonstrate that such neurons can reduce unnecessary neural firing and conflict in information transmission, leading to more efficient learning and lower energy use.
• To validate the approach in a real-world, complex task: audiovisual speech denoising under multi-talker, low SNR conditions.
• To show that the model maintains high performance even when components are damaged, indicating robustness through context-aware processing.

Proposed method

• The model uses a two-point neuron architecture where basal dendrites receive feedforward inputs (e.g., audiovisual features), and apical dendrites receive contextual inputs from neighboring processors, distant network regions, and memory.
• Contextual integration at the apical dendrites modulates the transmission of feedforward signals via a learned gating mechanism, amplifying coherent, relevant inputs and attenuating irrelevant ones.
• The modulation is implemented using a context-aware gating function that combines local, distant, and universal inputs to compute a dynamic gain factor applied to feedforward activation.
• The network is trained using a modified autoencoder setup with shared encoder-decoder structure, where the encoder uses context-sensitive processors and the decoder reconstructs clean audio from encoded features.
• A resilience test is performed by randomly deactivating processors (set to zero) at varying rates (P = 0 to 0.5) to evaluate robustness under component failure.
• Loss functions include standard reconstruction loss and ELBO loss, with ReLU activations, batch normalization, and Glorot uniform weight initialization.

Experimental results

Research questions

• RQ1 Can context-sensitive two-point neurons reduce neural activity and improve energy efficiency in deep learning by selectively transmitting only relevant information?
• RQ2 How does contextual modulation from apical dendrites affect the transmission of feedforward signals in a real-world, noisy audiovisual processing task?
• RQ3 To what extent does the proposed architecture outperform standard deep learning models in terms of robustness to component failure?
• RQ4 Does the integration of contextual inputs from multiple sources (local, distant, memory) enhance the model’s ability to process heterogeneous, complex real-world data?
• RQ5 Can this biologically inspired architecture achieve high performance with minimal neural activity, mimicking the efficiency of the neocortex?

Key findings

• The context-sensitive two-point neuron model achieved superior performance in audiovisual speech denoising under low SNR and multi-talker conditions, with significantly reduced neural activity compared to standard deep learning models.
• The model demonstrated high resilience to processor damage, maintaining performance even when up to 25% of processors were randomly deactivated, outperforming baseline models in robustness.
• Empirical results showed that reconstruction quality dropped sharply beyond a 1% error threshold, indicating a critical point in model stability.
• The inclusion of contextual inputs from neighboring and distant processors enhanced feature selection, reducing transmission of conflicting or irrelevant signals.
• The model achieved near-perfect reconstruction when provided with clean features, confirming that the encoder’s feature quality—driven by context-sensitive processing—was the primary determinant of final output quality.
• The use of contextual integration zones enabled dynamic, adaptive gating of feedforward signals, leading to more efficient and effective information processing.