Opinion
Deep Neural Networks as Scientific Models

https://doi.org/10.1016/j.tics.2019.01.009Get rights and content

Highlights

Neurally inspired deep neural networks (DNNs) have recently emerged as powerful computer algorithms tackling real-world tasks on which humans excel, such as object recognition, speech processing, and cognitive planning.

In the absence of scientific explanations regarding how humans solve such tasks, some cognitive scientists have turned to DNNs as models of human brain responses and behaviour.

In visual and auditory processing, DNNs were found to predict human brain responses and behaviour better than other models.

The use of DNNs as models in cognitive science has created a heated debate about their scientific value: in particular, are DNNs only valuable as predictive tools or do they also offer useful explanations of the phenomena investigated?

Artificial deep neural networks (DNNs) initially inspired by the brain enable computers to solve cognitive tasks at which humans excel. In the absence of explanations for such cognitive phenomena, in turn cognitive scientists have started using DNNs as models to investigate biological cognition and its neural basis, creating heated debate. Here, we reflect on the case from the perspective of philosophy of science. After putting DNNs as scientific models into context, we discuss how DNNs can fruitfully contribute to cognitive science. We claim that beyond their power to provide predictions and explanations of cognitive phenomena, DNNs have the potential to contribute to an often overlooked but ubiquitous and fundamental use of scientific models: exploration.

Section snippets

The Contested Value of Deep Neural Networks in Cognitive Science

In recent years, neurally inspired 1, 2 artificial deep neural networks (DNNs; see Glossary) have revolutionised first computer vision [3] and subsequently other domains such as natural language processing [4], control and planning (such as playing games, e.g., Atari and Go 5, 6), and navigational tasks (such as finding the shortest path on a subway map [7]).

DNNs are computational models consisting of many simple processing units (akin to neurons) that work in parallel and are arranged in

The Nature of Good Scientific Models

To evaluate whether DNNs are good scientific models, we need to agree on what makes a good model in the first place. Can we formulate a list of properties good models fulfil and check DNNs against these properties? Or, can we select a particular model that we believe to be an excellent scientific model as a standard and compare DNNs against it? Reflecting on the use of models in a broader scientific context, we argue that the case is not as simple as that.

The Predictive Power of DNNs

In technology and engineering the primary goal is to create artefacts that do things (i.e., correctly predict a particular outcome), while explanation often takes second place [33]. Are DNNs useful in cognitive science in this way? While DNN critics concede the predictive power of DNNs, they often dismiss it as less valuable for science than explanation. Here, we affirm the value of prediction on two grounds.

A pragmatic reason is that due to its predictive power a DNN could be used akin to a

The Explanatory Power of DNNs

How can a model do explanatory work in cognitive science (Box 2)? The blueprint notion many researchers have in mind is so-called mathematical–theoretical modelling 43, 44. There, a few relevant variables for describing a phenomenon are identified, it is hypothesised how they interact, and the variables and their interaction are modelled mathematically. Each variable is a priori linked to a part of the phenomenon modelled in the world, such that changes in the model variable are directly

The Exploratory Power of DNNs

An idealised view of natural science is that it proceeds by deriving hypotheses from a theory and testing them in experiments. But what to do if a fully-fledged and convincing theory is missing? Then we need to explore 64, 65 to create starting points for new theories, rather than predict or explain. This means a shift of perspective from models as tools for predictions, or akin to theories for explanation, to exploration [21].

Observing the scientific practise indicates that exploration is an

Concluding Remarks and Future Directions

Taking a bird’s-eye view from the stance of philosophy of science, we took a fresh look on what is at stake in the debate on DNNs as models for behaviour and neural activity. We emphasise four take-home messages for future research (Figure 2). First, given the current level of theory development and the need to trade-off model desiderata, we should embrace DNNs as one of many diverse kinds of useful models. Second, through their predictive power DNNs have rich potential as tools for scientific

Acknowledgments

We thank Aude Oliva and Philippe Schyns for thoughtful comments and feedback. This work was supported by Deutsche Forschungsgemeinschaft (DFG) grants awarded to R.M.C. (CI241/1-1, CI241/3-1) and D.K. (KA4683/2-1).

Glossary

Analogy
a similarity between relations in two different domains.
Box-and-arrow model
a model of information processing in which boxes represent components of an information processing system, and arrows represent information flow between those components.
Deep neural network
a computer algorithm inspired by biological neural networks, consisting of units akin to neurons and defined in function by the connection between the units (Figure 1A). Units are often not connected to all other units but

References (93)

  • Graves, A. et al. (2013) Speech recognition with deep recurrent neural networks. IEEE International Conference on...
  • V. Mnih

    Human-level control through deep reinforcement learning

    Nature

    (2015)
  • D. Silver

    Mastering the game of Go with deep neural networks and tree search

    Nature

    (2016)
  • A. Graves

    Hybrid computing using a neural network with dynamic external memory

    Nature

    (2016)
  • Y. LeCun

    Deep learning

    Nature

    (2015)
  • J. Kubilius

    Deep neural networks as a computational model for human shape sensitivity

    PLoS Comput. Biol.

    (2016)
  • J.C. Peterson

    Adapting deep network features to capture psychological representations: an abridged report

  • S.-M. Khaligh-Razavi et al.

    Deep supervised, but not unsupervised, models may explain IT cortical representation

    PLoS Comput. Biol.

    (2014)
  • R.M. Cichy

    Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object recognition reveals hierarchical correspondence

    Sci. Rep.

    (2016)
  • T. Horikawa et al.

    Generic decoding of seen and imagined objects using hierarchical visual features

    Nat. Commun.

    (2017)
  • D.L.K. Yamins

    Performance-optimized hierarchical models predict neural responses in higher visual cortex

    Proc. Natl. Acad. Sci. U. S. A.

    (2014)
  • U. Güçlü et al.

    Deep neural networks reveal a gradient in the complexity of neural representations across the ventral stream

    J. Neurosci.

    (2015)
  • S.A. Cadena

    Deep convolutional models improve predictions of macaque V1 responses to natural images

    bioRxiv

    (2017)
  • A. Gelfert

    How to Do Science with Models: A Philosophical Primer

    (2016)
  • C.G. Hempel

    Explanation in science and in history

  • A. Gelfert

    Exploratory uses of scientific models

  • A. Gelfert

    Strategies and trade-offs in model-building

  • R. Levins

    The strategy of model building in population biology

    Am. Sci.

    (1966)
  • M. Morrison et al.

    Models as mediating instruments

  • E. Ising

    Beitrag zur theorie des ferromagnetismus

    Z. Phys.

    (1925)
  • A. Baddeley

    Working memory: theories, models, and controversies

    Annu. Rev. Psychol.

    (2011)
  • J.R. Anderson

    The Adaptive Character of Thought

    (1990)
  • T. van Gelder

    The dynamical hypothesis in cognitive science

    Behav. Brain Sci.

    (1998)
  • K. Kording

    Appreciating diversity of goals in computational neuroscience

    OSF Prepr.

    (2018)
  • A. Gelfert

    Between theory and phenomena: what are scientific models?

  • N. Cartwright

    Models and the limits of theory: quantum hamiltonians and the BCS model of superconductivity

  • M. Boon et al.

    Models as epistemic tools in engineering sciences: a pragmatic approach

  • R. Rajalingham

    Large-scale, high-resolution comparison of the core visual object recognition behavior of humans, monkeys, and state-of-the-art deep artificial neural networks

    J. Neurosci.

    (2018)
  • H. Hong

    Explicit information for category-orthogonal object properties increases along the ventral stream

    Nat. Neurosci.

    (2016)
  • D.L.K. Yamins et al.

    Using goal-driven deep learning models to understand sensory cortex

    Nat. Neurosci.

    (2016)
  • E.Y. Walker

    Inception in visual cortex: in vivo-silico loops reveal most exciting images

    bioRxiv

    (2018)
  • P. Bashivan

    Neural population control via deep image synthesis

    bioRxiv

    (2018)
  • M. Schrimpf

    Brain-Score: which artificial neural network for object recognition is most brain-like?

    bioRxiv

    (2018)
  • W.C. Wimsatt

    False models as means to truer theories

  • L. Breiman

    Statistical modeling: the two cultures

    Stat. Sci.

    (2001)
  • T. Yarkoni et al.

    Choosing prediction over explanation in psychology: lessons from machine learning

    Perspect. Psychol. Sci.

    (2017)

    • Cited by (0)

    View full text
    © 2019 Elsevier Ltd. All rights reserved.