If deep learning is the answer, what is the question?

Andrew Saxe, Stephanie Nelli, Christopher SummerfieldNature Reviews. Neuroscience 2020 · DOI

Abstract

Neuroscience research is undergoing a minor revolution. Recent advances in machine learning and artificial intelligence (AI) research have opened up new ways of thinking about neural computation. Many researchers are excited by the possibility that deep neural networks may offer theories of perception, cognition and action for biological brains. This perspective has the potential to radically reshape our approach to understanding neural systems, because the computations performed by deep networks are learned from experience, not endowed by the researcher. If so, how can neuroscientists use deep networks to model and understand biological brains? What is the outlook for neuroscientists who seek to characterise computations or neural codes, or who wish to understand perception, attention, memory, and executive functions? In this Perspective, our goal is to offer a roadmap for systems neuroscience research in the age of deep learning. We discuss the conceptual and methodological challenges of comparing behaviour, learning dynamics, and neural representation in artificial and biological systems. We highlight new research questions that have emerged for neuroscience as a direct consequence of recent advances in machine learning.

This perspective paper by Andrew Saxe, Stephanie Nelli, and Christopher Summerfield explores the burgeoning synergy between deep learning (DL) and systems neuroscience. Rather than just using DL as a tool for data analysis, the authors argue that it provides a "roadmap" for understanding how biological computation emerges from experience, while cautioning against treating these models as mere "black boxes."

The Shift Toward Emergent Computation

For decades, systems neuroscience relied on "hand-crafted" theories—researchers would propose specific wiring diagrams or tuning curves to explain how the brain processes information. However, as we record from thousands of neurons simultaneously, these intuitive models often fail to capture the "messy," high-dimensional, and multi-tasking nature of real neural populations.

The "Deep Learning Framework" addresses this by suggesting that we shouldn't build the circuits ourselves. Instead, we should specify the architecture (the layout), the cost function (the goal), and the learning rule (how to improve). The core challenge is scientific: if a deep network "solves" a task like a brain does, have we actually understood the brain, or have we just replaced one mystery with another?

Methodological Bridges

The paper synthesizes several key methodological approaches currently used to bridge AI and neuroscience:

  • Representational Similarity Analysis (RSA): Comparing the "distance" between internal representations of different objects in both artificial and biological networks.
  • Linear Mapping/Regression: Predicting actual neuronal firing rates using a linear combination of artificial "unit" activations.
  • State-Space Analysis: Examining the population-level "trajectories" of neural activity during tasks like decision-making or interval timing.
Comparing biological brain regions with artificial neural network models across representations, activity, and dynamics.

Key Insights and Ideas

  • The "End-to-End" Revolution: Unlike 1980s connectionism, modern DL learns from raw, naturalistic signals (pixels, audio). This allows for the de novo emergence of filters (like Gabor patches in V1) that weren't programmed in by the researcher.
  • Representational Equivalence is a Low Bar: Even untrained (randomly initialized) networks can explain a significant amount of brain activity variance. This suggests that architecture alone does a lot of the heavy lifting, and we need stricter tests for "shared coding."
  • Learning Dynamics Matter: We shouldn't just look at the final, trained state of a network. The path the network takes during training—such as learning broad categories before fine details—mirrors human developmental stages and provides deeper insight into biological constraints.
  • Model Idealization: Using "Deep Linear Networks" to derive exact mathematical solutions for learning trajectories allows researchers to see why certain behaviors (like stage-like learning) emerge. Paradoxically, "infinite-size" or "deep linear" networks are often more interpretable than small, complex ones because their learning trajectories are mathematically stable and predictable.
Theoretical and simulated learning dynamics in deep networks, showing training error, weight evolution, and the impact of initialization and width.

Key Results

  • Variance Explained: Top-performing CNNs can explain roughly 60% of the variance in primate IT cortex responses, though the gap between "good" and "bad" models is narrowing.
  • Robustness Gap: While DL models match humans on standard images, they fail significantly more on "out-of-distribution" tasks (e.g., recognizing an object made of different textures), highlighting a lack of biological-like abstraction.
  • Stage-like Transitions: Deep networks exhibit "quasi-discrete" stages in learning, explaining why children might suddenly "get" a concept after a period of stagnation.

Limitations and Future Work

The authors are candid about the "existential threat" DL poses to traditional neuroscience: if we just document how different AI architectures resemble the brain, are we still doing neuroscience, or just "AI-ology"? Key limitations include the Credit Assignment Problem (finding biologically plausible alternatives to backpropagation), Data Efficiency (humans learn from far fewer examples), and Higher Cognition (modeling reasoning and flexible abstraction).

The authors conclude that the goal shouldn't be to build the most complex simulation possible, but to use DL to find falsifiable predictions about how the brain learns and allocates its resources.

Summary updated

BibTeX
@article{saxe2020if,
  author = {Andrew Saxe and Stephanie Nelli and Christopher Summerfield},
  title = {{If deep learning is the answer, what is the question?}},
  journal = {Nature Reviews. Neuroscience},
  year = {2020},
  doi = {10.1038/s41583-020-00395-8},
}

Loading related content...

Figure