No Free Lunch from Deep Learning in Neuroscience: A Case Study through Models of the Entorhinal-Hippocampal Circuit

Rylan Schaeffer, Mikail Khona, Ila Fiete

In recent years, a wave of excitement has swept through the intersection of AI and Neuroscience. The narrative is compelling: if we train a deep neural network to perform a biological task—like navigating a room—and "grid cells" (the brain's internal GPS) naturally emerge within the network, we’ve discovered a fundamental truth about how the brain optimizes for that task.

However, a provocative paper from researchers at Stanford and MIT, titled "No Free Lunch from Deep Learning in Neuroscience," throws a bucket of cold water on this "emergence" narrative. Through a rigorous case study of the entorhinal-hippocampal circuit, the authors argue that these "breakthroughs" are often less about biological discovery and more about clever (and sometimes accidental) engineering.

The "Emergence" Narrative vs. Reality

The "Holy Grail" of computational neuroscience is to understand why the brain uses specific neural representations, like the hexagonal firing patterns of grid cells. Recent high-profile papers claimed that grid cells are the "optimal" solution for path integration (calculating position from movement velocity). They argued that if you train a simple recurrent neural network (RNN) to path-integrate, grid cells appear "for free."

Model setup and the contrast between the assumed 'free' emergence of grid cells versus the highly specific, nested constraints actually required.

The authors of this paper suspected a "post-hoc" fallacy. They wanted to know: Is grid cell emergence a robust, fundamental outcome of optimization, or is it a fragile result of specific coding choices made by researchers who already knew what grid cells looked like?

Stress-Testing the Models

To answer this, the researchers conducted a massive, "brute-force" investigation to stress-test the claims of previous models. They trained over 11,000 networks, varying architectures (RNNs, LSTMs, GRUs), activation functions, and optimizers.

The models were tasked with navigating a simulated 2.2m x 2.2m square arena. The networks received sequences of 2D velocity and an initial starting position, and were supervised to output the agent's current position using various encoding schemes (the "readout"). The researchers tested different ways to represent this position, including Cartesian coordinates, polar coordinates, and various "place cell" models (Gaussian, Difference-of-Gaussians, etc.). Finally, they used "gridness scores" to quantify how hexagonal the resulting neural firing patterns were.

Key Insights: The Illusion of "Free" Emergence

  • The Readout is Everything: Grid cells don't emerge from the task (path integration); they emerge from the readout. Specifically, only a very narrow set of "Difference-of-Softmaxes" (DoS) targets—which have a built-in inhibitory surround—consistently produced grids.
  • The "Baking In" Effect: The spatial scale (period) of the "emergent" grid cells was perfectly correlated with the hyperparameters of the target functions. In other words, the programmer chooses the grid size, not the optimization process.
  • Fragility over Robustness: Even with the "correct" settings, grid emergence was highly sensitive to random seeds and specific hyperparameter combinations. Most networks learned to navigate perfectly without ever developing grid cells.
  • Biological Realism Kills the Model: In the brain, place cells (which provide input/output for grid cells) vary in size and shape. When the authors added this realistic "heterogeneity" to their models, the grid cells vanished. This suggests current deep learning models are "over-fitted" to idealized, biologically implausible assumptions.

Key Results: Task Success $\neq$ Biological Accuracy

The most striking finding was the disconnect between solving the engineering problem and mirroring biology. Almost all 11,000+ networks successfully learned to path-integrate with an error of less than 10cm. However, only a tiny fraction produced grid-like units.

Comparison of task success (top) versus grid cell emergence (bottom) across different target encodings.

Furthermore, real brains have multiple "modules" of grid cells with different scales. The deep learning models only ever produced a single scale, entirely determined by the user-defined readout width. The authors also highlighted a "predictivity paradox": some models "explain" brain data well not because they are biologically accurate, but because they have high intrinsic dimensionality, allowing them to be easily regressed onto neural data through sheer mathematical flexibility.

Moving Forward: A Call for Transparency

This paper is primarily a critique, setting a higher bar for future research. It demonstrates an informal "No Free Lunch" theorem for neuroscience: we cannot expect deep learning models to produce accurate models of the brain without adding substantial, explicit inductive biases.

The authors urge the community to move away from "look, it emerged!" narratives and toward rigorous sweeps that show exactly which choices lead to which results. They propose that additional constraints—like non-negativity, error correction, and translation invariance—might be the missing ingredients for a truly robust model of grid cells. If we want to understand the brain using AI, we can't just throw data at a black box; we have to understand the math behind the box first.

Summary updated

Abstract

Research in Neuroscience, as in many scientific disciplines, is undergoing a renaissance based on deep learning. Unique to Neuroscience, deep learning models can be used not only as a tool but interpreted as models of the brain. The central claims of recent deep learning-based models of brain circuits are that they make novel predictions about neural phenomena or shed light on the fundamental functions being optimized. We show, through the case-study of grid cells in the entorhinal-hippocampal circuit, that one may get neither. We begin by reviewing the principles of grid cell mechanism and function obtained from first-principles modeling efforts, then rigorously examine the claims of deep learning models of grid cells. Using large-scale architectural and hyperparameter sweeps and theory-driven experimentation, we demonstrate that the results of such models may be more strongly driven by particular, non-fundamental, and post-hoc implementation choices than fundamental truths about neural circuits or the loss function(s) they might optimize. We discuss why these models cannot be expected to produce accurate models of the brain without the addition of substantial amounts of inductive bias, an informal No Free Lunch result for Neuroscience. Based on first principles work, we provide hypotheses for what additional loss functions will produce grid cells more robustly. In conclusion, circumspection and transparency, together with biological knowledge, are warranted in building and interpreting deep learning models in Neuroscience.
BibTeX
@inproceedings{schaeffer2022no,
  author = {Rylan Schaeffer and Mikail Khona and Ila Fiete},
  title = {{No Free Lunch from Deep Learning in Neuroscience: A Case Study through Models of the Entorhinal-Hippocampal Circuit}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
  url = {https://openreview.net/forum?id=o6Yq_mxY1oT},
}
Figure