Neural system identification for large populations separating "what" and "where" (1711.02653v2)

Abstract: Neuroscientists classify neurons into different types that perform similar computations at different locations in the visual field. Traditional methods for neural system identification do not capitalize on this separation of 'what' and 'where'. Learning deep convolutional feature spaces that are shared among many neurons provides an exciting path forward, but the architectural design needs to account for data limitations: While new experimental techniques enable recordings from thousands of neurons, experimental time is limited so that one can sample only a small fraction of each neuron's response space. Here, we show that a major bottleneck for fitting convolutional neural networks (CNNs) to neural data is the estimation of the individual receptive field locations, a problem that has been scratched only at the surface thus far. We propose a CNN architecture with a sparse readout layer factorizing the spatial (where) and feature (what) dimensions. Our network scales well to thousands of neurons and short recordings and can be trained end-to-end. We evaluate this architecture on ground-truth data to explore the challenges and limitations of CNN-based system identification. Moreover, we show that our network model outperforms current state-of-the art system identification models of mouse primary visual cortex.

• The paper introduces a novel convolutional neural network (CNN) with a sparse readout layer for neural system identification, designed to efficiently separate spatial ('where') and feature ('what') dimensions in large neuron populations.
• Evaluations on both simulated ground truth data and empirical mouse visual cortex data demonstrate that this CNN architecture significantly outperforms conventional models in predictive performance and distinguishing neuron identity.
• This work's implications include facilitating a more nuanced understanding of functional cell types, informing experimental design, and offering a potential framework for understanding neural dynamics in various brain systems.

Neural System Identification: A Novel Convolutional Architecture

The paper "Neural system identification for large populations separating 'what' and 'where'" by David A. Klindt et al. presents an advanced approach to neural system identification by employing a convolutional neural network (CNN) architecture specifically designed to efficiently dissect the spatial ('where') and feature ('what') dimensions of neural data. This proposal signifies a step towards overcoming major limitations in precisely modeling neural responses at scale, particularly under constraints of limited data collection from individual neurons.

Summary of Methodology and Contributions

The paper addresses the challenge of building quantitative models for neurons which process stimuli nonlinearly and exhibit spatial invariance in feature extraction. Traditional generalized linear models (GLMs) simplify this complexity but fall short, especially in higher visual areas. The authors argue that the separation of spatial and feature dimensions can be better exploited using deep learning, specifically through a novel CNN design incorporating a sparse readout layer.

1. Sparse Readout Layer: The key innovation is a readout layer that decouples spatial location and feature extraction. It reduces the parameter space significantly compared to fully connected layers, allowing scalable modeling across large neuron populations with minimal data.
2. End-to-End Trainability: The proposed CNN architecture can be trained from end to end, distinguishing itself by directly learning from data without requiring heuristic preprocessing such as receptive field mapping.
3. Evaluation on Ground Truth Data: Utilizing simulated data from a simpler set of ground truth models, the approach demonstrates substantial enhancements in predictive performance over conventional GLMs. Moreover, the model shows proficiency in distinguishing individual neuron's location and functional identity within a larger population when other models were deficient.
4. Comparative Performance on Real Data: The proposed method surpasses existing state-of-the-art models when applied to empirical data from mouse primary visual cortex, affirming its practical relevance and broader applicability in neuroscience research.

Numerical Performance and Implications

Quantitative evaluation reveals that, with the introduction of the sparse readout mechanism, the network retains comparable performance even as the data volume increases and the neural population becomes more diverse. The results on complex ground truth models and empirical datasets underscore the architecture's potential in dealing with the intricacies of real-world neural systems.

Implications and Further Research

This work's implications extend beyond improving predictive accuracy. By facilitating a more nuanced understanding of functional cell types, the research opens avenues for more detailed neural taxonomies which could illuminate ingrained activity patterns across different brain areas. The insights regarding neurons' equivariance could substantially influence experimental design, guiding the strategic trade-off between neuron sample size and recording length to optimize data utility.

Moreover, the paper hints at potential expansions, suggesting that this framework could be adapted to other neural systems and cognitive domains where spatial and feature influences are similarly critical. Future developments might integrate additional equivariant transformations like orientation selectivity, thereby advancing our grasp of complex neural dynamics in visual and non-visual systems.

In conclusion, the introduction of a CNN architecture with a sparse readout system delineates a promising trajectory for neural system identification. Its ability to disentangle 'what' a neuron computes from 'where' it is located provides a robust platform for future explorations into neural computation, with the expectation that this methodology could be instrumental in refining brain-computer interfaces and other neurotechnological applications.