Deep Neural Networks in Computational Neuroscience

Tim C. Kietzmann, Patrick McClure, Nikolaus KriegeskortebioRxiv 2017 · DOI

TL;DR

Deep neural networks (DNNs), despite their biological simplifications, are revolutionizing computational neuroscience by providing powerful models that can predict neural activity with unprecedented accuracy. By exploring DNNs through manipulating their architecture, input statistics, objective functions, and learning algorithms, researchers are gaining crucial insights into what biological details are essential for brain computation.

What Problem Does This Paper Address?

Computational neuroscience aims to understand how the brain processes information to produce cognition and behavior. Traditionally, this has involved building mathematical models that map sensory input to neural responses or behavior. While shallow models were easier to interpret, the brain's complexity demands more sophisticated models. Deep neural networks (DNNs), inspired by biological brains, have achieved remarkable success in artificial intelligence tasks. However, they often neglect many biological details, leading to questions about their explanatory power for neuroscience. The core challenge is to bridge the gap between the complexity of biological brains and the simplified yet powerful nature of DNNs, and to understand what biological features are truly necessary for brain computation.

Methods

The paper discusses several key approaches for using DNNs in computational neuroscience:

  • Encoding Models: These models use DNN features as a basis to predict neural responses, testing how well the model's representations align with brain activity.
  • Representational Similarity Analysis (RSA): This technique compares the representational geometry of DNNs and brains by analyzing the dissimilarity of their responses to various stimuli. RDMs are used to quantify these similarities.
  • Feature Visualization: Techniques like "in silico" electrophysiology and network dissection are used to probe the selectivity of individual units within DNNs, revealing what stimuli maximally activate them.
  • Feature Visualization by Optimization: This method involves modifying images to maximally excite or inhibit specific DNN units, providing insights into their learned features.
  • Exploring DNN Design Space: The paper emphasizes the importance of manipulating four key elements of DNNs: network architecture, input statistics, functional objective, and learning algorithm, to understand their impact on predictive power and biological plausibility.
  • Recurrent Neural Networks (RNNs): The paper highlights the growing importance of incorporating recurrence into DNNs to better model the brain's dynamic processing.

Datasets

Not explicitly detailed as distinct datasets for training or evaluation in the provided text. The paper refers to the use of "ImageNet competition" for visual object recognition and "category-labelled sets of training images" for vision tasks, implying standard large-scale image datasets are utilized.

Key Insights and Ideas

  • DNNs as Powerful Predictive Models: Despite their biological simplifications, DNNs are remarkably effective at predicting neural responses, often surpassing other model types. This suggests they capture essential computational principles.
  • The "Black Box" Argument is Nuanced: While DNNs can be complex, their internal workings are not entirely opaque. By manipulating their core components (architecture, input statistics, objective, learning algorithm), researchers can gain insights.
  • Biological Realism vs. Explanatory Power: There's a trade-off between biological detail and a model's ability to explain behavior. DNNs demonstrate that high performance can be achieved with simplified biological mechanisms.
  • Four Pillars of DNN Computation: Understanding DNNs involves considering their architecture, input statistics, functional objective, and learning algorithm. Manipulating these provides a framework for exploring brain computation.
  • Recurrence is Crucial: The brain's reliance on recurrent connections is a key feature that current feed-forward DNNs often miss, and incorporating recurrence is a promising avenue for more biologically plausible models.
  • Spiking Models as a Frontier: While current DNNs are typically deterministic, exploring stochasticity and spiking dynamics in neural networks is an exciting area for future research.

Key Results

  • DNNs can predict neural responses to novel sensory stimuli with accuracies "well beyond any other currently available model type."
  • DNNs achieve human-level classification performance in tasks like visual object recognition.
  • DNNs explain approximately 50% of the variance in spike counts in macaque IT cortex, comparable to established models in V1.
  • DNNs can predict representations across multiple levels of processing in both macaque and human brains, aligning with functional magnetic resonance imaging (fMRI) data.
  • The paper suggests that DNNs are currently the "best-performing model in explaining human eye-movements in free viewing paradigms."

Limitations and Future Work

  • Biological Simplifications: Current DNNs neglect many biological details, such as lateral and top-down connections, and often use continuous outputs instead of spikes.
  • "Black Box" Concern: While progress is being made, understanding the exact computations within DNNs remains a challenge.
  • Biological Plausibility of Learning: Backpropagation, the dominant learning algorithm for DNNs, is considered neurobiologically implausible, though research into biologically plausible alternatives is ongoing.
  • Stochasticity and Oscillations: DNNs are typically deterministic, unlike biological networks which exhibit stochasticity and oscillations, features whose roles are still being explored.
  • Future Work: The paper emphasizes the need for continued integration of biological details into DNNs, close collaboration between modelers and experimental neuroscientists, and exploring recurrent architectures and spiking neuron models. The authors suggest that DNNs will not replace shallow models but will enhance the investigative repertoire of computational neuroscientists.

Summary updated

Abstract

Summary The goal of computational neuroscience is to find mechanistic explanations of how the nervous system processes information to give rise to cognitive function and behaviour. At the heart of the field are its models, i.e. mathematical and computational descriptions of the system being studied, which map sensory stimuli to neural responses and/or neural to behavioural responses. These models range from simple to complex. Recently, deep neural networks (DNNs) have come to dominate several domains of artificial intelligence (AI). As the term “neural network” suggests, these models are inspired by biological brains. However, current DNNs neglect many details of biological neural networks. These simplifications contribute to their computational efficiency, enabling them to perform complex feats of intelligence, ranging from perceptual (e.g. visual object and auditory speech recognition) to cognitive tasks (e.g. machine translation), and on to motor control (e.g. playing computer games or controlling a robot arm). In addition to their ability to model complex intelligent behaviours, DNNs excel at predicting neural responses to novel sensory stimuli with accuracies well beyond any other currently available model type. DNNs can have millions of parameters, which are required to capture the domain knowledge needed for successful task performance. Contrary to the intuition that this renders them into impenetrable black boxes, the computational properties of the network units are the result of four directly manipulable elements: input statistics, network structure, functional objective , and learning algorithm . With full access to the activity and connectivity of all units, advanced visualization techniques, and analytic tools to map network representations to neural data, DNNs represent a powerful framework for building task-performing models and will drive substantial insights in computational neuroscience.
BibTeX
@misc{kietzmann2017deep,
  author = {Tim C. Kietzmann and Patrick McClure and Nikolaus Kriegeskorte},
  title = {{Deep Neural Networks in Computational Neuroscience}},
  year = {2017},
  doi = {10.1101/133504},
  howpublished = {bioRxiv},
}
Figure