TL;DR
This paper introduces DeepDAM, a novel deep learning framework that bridges the gap between synthetic and real-world neural data for more accurate neural circuit connectivity inference. By employing domain adaptation and self-training, DeepDAM effectively overcomes the "model mismatch" problem, achieving near-perfect performance on challenging in vivo datasets and demonstrating broad applicability across different neural systems.
What Problem Does This Paper Address?
Accurately mapping neural circuits is crucial for understanding brain function, but it's incredibly difficult to obtain ground-truth connectivity data from living brains. Current deep learning methods often rely on synthetic data, which can lead to a "model mismatch" problem: the simulated neural activity doesn't perfectly reflect real-world dynamics. This mismatch significantly degrades performance when these models are applied to actual experimental recordings. The authors aim to solve this by developing a framework that can effectively learn from both synthetic and unlabeled real-world neural data, bridging the domain gap.
Methods
The core of DeepDAM is a five-step framework that integrates two key machine learning strategies:
- Model Formulation: A computational model (MAT network) is used to generate synthetic neural data that mimics real-world recordings.
- Training Dataset Synthetization: This model generates synthetic data with ground-truth labels for connectivity.
- Feature Alignment (Domain Adaptation): A deep neural network (DNN) is trained to align the feature representations of synthetic and real (unlabeled) experimental data. This is achieved by minimizing a domain adaptation loss (LDA) that bridges the distributional gap in the feature space.
- Classifier Construction: The DNN's classifier is initially trained on the labeled synthetic data.
- Classifier Refinement (Self-Training): To prevent overfitting to synthetic data and improve performance on real data, the classifier is further refined using unlabeled experimental data. This involves assigning "pseudo-labels" to the experimental data and training the classifier on these pseudo-labeled samples, guided by a self-training loss (LST).
A crucial aspect is the use of a gating mechanism that filters out unreliable experimental data during training, ensuring that only the most relevant data contribute to model improvement.
Datasets
- Synthetic Data: Generated by the authors' MAT neural network, mimicking rodent hippocampal CA1 circuits.
- Real In Vivo Data:
- A subset of extracellular recordings from the CA1 region of freely-running mice with ground-truth labels (from English et al., 2015). This dataset was used for validation and is noted as one of the few available in vivo datasets with such labels.
- A larger subset of in vivo spike data from the same source, characterized by lower signal-to-noise ratios and inconsistent ground-truth labels, to test robustness.
- Ex Vivo and In Vitro Data:
- Intracellular recordings from individual mouse neurons (Allen Cell Types Dataset).
- Data from a stomatogastric ganglion microcircuit of Cancer borealis.
Key Insights and Ideas
- Model Mismatch as an OOD Problem: The authors reframe the "model mismatch" problem in neuroscience as an out-of-distribution (OOD) challenge in machine learning, a perspective that has been successfully addressed by domain adaptation techniques.
- Domain Adaptation and Self-Training Synergy: The combination of domain adaptation (to align feature spaces) and self-training (to refine the classifier on real data) is key to DeepDAM's success. This dual approach effectively bridges the gap between synthetic and real data.
- Importance of Unlabeled Real Data: DeepDAM's ability to leverage unlabeled in vivo data is critical, as obtaining labeled in vivo data is extremely challenging.
- Robustness Across Conditions: The framework demonstrates remarkable robustness to variations in recording length, network size, and data ambiguity, suggesting its applicability in real-world, less controlled experimental settings.
- Generalizability Beyond Connectivity: The core DeepDAM framework can be adapted for other model-fitting tasks, such as inferring biophysical properties of neurons and microcircuits, highlighting its versatility.
Key Results
- Near-Perfect Connectivity Inference: On the rodent hippocampal CA1 dataset, DeepDAM achieved a Matthews correlation coefficient (MCC) of 0.97-1.0, significantly outperforming classical methods (MCC ~0.6-0.7) and vanilla deep learning approaches (MCC ~0.61).
- Robustness: The aggregated DeepDAM method achieved an MCC of 0.97 (SD=0.04), showing high consistency across different random seeds and hyperparameters.
- Improved Performance with Recording Length: Performance consistently improved with longer recording durations, with DeepDAM surpassing other frameworks even on shorter recordings.
- Broad Applicability: DeepDAM successfully inferred biophysical properties of single neurons and microcircuits, outperforming existing methods in accuracy and efficiency. For example, it achieved a 2.54 times smaller inference error than the state-of-the-art SNPE-C method for biophysical patterns.
- Domain Randomization vs. DeepDAM: While domain randomization improved vanilla deep learning performance, DeepDAM still significantly outperformed it.
Limitations and Future Work
- Model Specificity: DeepDAM relies on a computational model that captures region-specific firing dynamics. For application to new brain regions, new biophysical models that faithfully reproduce those region's characteristics would need to be developed.
- Multiple Feasible Properties: For model-fitting tasks, DeepDAM currently outputs only one set of feasible neural properties. The authors suggest that replacing the synthetic loss function could potentially address the issue of multiple feasible properties.
- Need for Large-Scale Labeled Datasets: The authors emphasize the urgent need for large-scale datasets with ground-truth connectivity to further refine and compare inference methods.
- Ambiguous Data: While DeepDAM performs well, ambiguous data (e.g., due to covariation or insufficient spikes) still presents challenges. Future work could involve incorporating probabilistic methods or ensemble methods to model uncertainty in these cases.
- Extending to Other Brain Regions and Species: The authors envision extending DeepDAM to other brain regions and species, and to more complex neural systems.
Summary updated