This paper introduces a rigorous framework for evaluating how structural changes to a neural network—such as pruning, quantization, or even hypothetical "mind-uploading" steps—affect its actual behavior. By moving beyond simple structural metrics, the authors provide a sobering look at the feasibility of speculative neurotechnology.
TL;DR
The authors propose the Functional Preservation Score (FPS), a metric that quantifies how much of a neural network's input-output behavior survives structural operations. Using this tool, they demonstrate that speculative "mind-migration" pipelines fail at every stage in simulation, primarily because biological identity is not a separable module and target wiring constraints are extremely restrictive.
What Problem Does This Paper Address?
In fields ranging from model compression to connectomics, we often perform "structural operations" on networks: we prune synapses, quantize weights, or attempt to align one brain's architecture with another. However, we lack a common currency to measure how much function is lost during these processes.
This is particularly relevant for the speculative concept of "mind-migration"—transferring a mind from a biological substrate to a digital or different biological one. Such proposals rely on a "partition–align–write" pipeline. This paper asks: if we try this in a controlled simulation using real neural data, does the function actually survive?
Methods
The core contribution is the Functional Preservation Score (FPS). Unlike standard Mean Squared Error (MSE), FPS is "null-normalized." It compares the transformed system's output to a reference system, but scales the result against a "trivial baseline" (like predicting the mean or a shuffled mapping).
$$FPS = \text{clip}\left(\frac{\text{raw}(Y) - \text{raw}(Y_{\text{null}})}{1 - \text{raw}(Y_{\text{null}})}, 0, 1\right)$$
An FPS of 1 means perfect functional preservation; 0 means the operation performed no better than a random guess.
The authors applied FPS to three main areas:
- Connectome Compression: Testing how low-bit quantization (like int4 or ternary) affects the C. elegans motor output.
- Replay Compression: Using autoencoders to compress hippocampal activity sequences.
- Migration Pipeline: Simulating the three stages of mind-transfer:
- Partition: Identifying a functional "identity payload."
- Transport: Mapping function across individuals using Gromov–Wasserstein alignment.
- Writing: Installing weights using biologically plausible local plasticity rules.
Datasets
The study utilizes several high-quality public datasets that are small enough to run on a standard laptop:
- Real C. elegans somatic connectome: 306 neurons (sensory, interneuron, and motor).
- Hippocampal Replay Data (DANDI 000044): Recordings of 48 pyramidal units from rats.
- Whole-brain calcium activity: Used for functional characterization.
Key Insights and Ideas
- The Sparsity Floor: You cannot quantize a real connectome indefinitely. Because synaptic weights are heavy-tailed, uniform quantization (like int4) forces a massive amount of "effective sparsity" (73%) because most weights simply round to zero. This creates a functional "floor" below which the network breaks.
- Identity is Not a Module: The "partition" step of mind-migration fails because there is no evidence of a separable "identity payload." Ablating the supposed identity-carrying neurons is no more disruptive than ablating a random subset of the same size.
- Wiring is the Bottleneck: Even if you have a perfect "map" of a mind, you cannot easily "write" it into a new substrate. If the target substrate's physical wiring (topology) doesn't match the original, functional transport collapses.
Key Results
- Quantization Limits: Engram-aware mixed precision can lower the sparsity floor compared to uniform quantization, but the physical distribution of weights remains a hard constraint.
- Replay Identifiability: Hippocampal sequences are remarkably resilient to compression. A 48x25 sequence can be compressed by $10^3\times$ and still remain identifiable well above chance.
- The Migration Gap: In the "Writing" stage, there is a significant "plausibility gap." While a global optimizer can theoretically install a function onto a target topology, biologically plausible local plasticity rules achieve only about 55% fidelity (FPS 0.55).
Limitations and Future Work
The authors are careful to note that their results concern coherence in simulation, not the biological feasibility of reading or writing to a living brain.
- Scale: The study is limited to small-scale networks (nematodes and ~40-neuron modules).
- Dynamics: It uses relatively simple leaky-rate models rather than complex spiking dynamics.
- Subjectivity: FPS measures functional preservation (behavior), not subjective experience or "qualia."
Future work could involve scaling these audits to larger connectomes (like the fruit fly) or exploring more complex plasticity rules that might bridge the "writing gap."
Summary updated