Deep Learning Theory & Optimization Dynamics
Extending Two-Layer Network Theory to Deep and Structured Architectures
Generated automatically from the limitations stated in 8 papers (NeurIPS, COLT, ICML), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
A substantial body of rigorous results—on training dynamics, predictor statistics, generalization, and under-fitting—is proven only for one- or two-layer fully-connected networks, often with fixed or special activations. Practitioners use deep residual, convolutional, and attention models, so none of these guarantees can be checked against the systems that matter. Until the theory carries over, each paper's conclusions remain unvalidated in practice, and the specific mechanisms they identify may be artifacts of the shallow, unstructured setting.
Why it matters
The existing shallow-network results become usable predictions about the deep architectures actually deployed, and the point-of-failure analysis itself becomes a research tool.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Layer-by-layer induction. Take an existing two-layer result whose proof factors through the feature kernel and attempt to propagate it through depth by showing the required statistics stabilize under repeated composition; measure how tightly the per-layer bounds compose (do they blow up exponentially, linearly, or hold?). A student can start within a month on depth-3–10 ReLU/softplus networks on synthetic data.
- 2
Fix-then-release the second layer. Several statements block on a fixed top layer; study the same models with the second layer trained, measuring whether the quantities the theory tracks (kernel spectra, predictor statistics) are stable under training, as a prerequisite for any deep guarantee.
- 3
Structured architectures as weight-sharing constraints. Treat convolution/attention as linear constraints on an otherwise fully-connected weight space and test whether the existing proofs survive under those constraints, measuring where the argument first breaks.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
The induction genuinely fails: the quantities tracked in two layers are simply not stable under composition, so each additional depth layer requires a new analysis and no unified result exists.
Sub-problems
Extending Optimization Dynamics and Representation Guarantees of Two-Layer ReLU Networks to Multivariate Inputs
Theoretical analyses of optimization dynamics and representation in shallow ReLU networks currently depend on the total ordering and simple geometry of one-dimensional inputs and fixed internal weights. As a consequence, existing proofs cannot predict whether training dynamics, underfitting phenomena, or convergence rates persist when inputs span multiple dimensions with non-trivial covariance structures. Extending these analyses beyond univariate settings is necessary to determine if current theoretical explanations capture general deep learning behavior or are merely artifacts of 1D scalar geometry.
Evidence
Each paper's own statement of the limitation, verbatim.
- Taking a Big Step: Large Learning Rates in Denoising Score Matching Prevent MemorizationCOLT 2025
Theory is restricted to one-dimensional data and two-layer ReLU networks with fixed ±1 inner weights; extension to multivariate settings is only suggested by experiments, not proven
- Solving Neural Min-Max Games: The Role of Architecture, Initialization & DynamicsNeurIPS 2025
Analysis is restricted to two-layer fully-connected networks with twice-differentiable activations (excluding ReLU, though extension to GeLU/softplus is suggested), whereas practice uses deep residual/convolutional architectures
- How Does Label Noise Gradient Descent Improve Generalization in the Low SNR Regime?NeurIPS 2025
Theory is restricted to squared ReLU activation and a two-layer CNN with fixed second layer; no guarantees for deeper or residual architectures
- Defining Neural Network Architecture through Polytope Structures of DatasetsICML 2024
Analysis restricted to fully connected ReLU networks; convolutional and other architectures not covered
- Critical feature learning in deep neural networksICML 2024
Only fully-connected feed-forward networks analyzed; extension to RNNs, CNNs, ResNets, and transformers is proposed but not demonstrated; predictor statistics not yet computed
- The Expressive Power of Tuning Only the Normalization LayersCOLT 2023
Results are for fully connected ReLU networks only; applicability to transformers/convolutional architectures with attention or other nonlinearities is not shown
- Regression as Classification: Influence of Task Formulation on Neural Network FeaturesAISTATS 2023
Results are proven only for one-dimensional data and two-layer ReLU networks; deeper models did not exhibit the under-fitting on their synthetic problem (Appendix G), so the explanation may not carry to practical deep networks
- Penalising the biases in norm regularisation enforces sparsityNeurIPS 2023
Results are restricted to univariate data and one-hidden-layer ReLU networks; multivariate extension is open and deemed very challenging.
Nearest existing work
- On the Dynamics of Gradient Descent for AutoencodersAISTATS 2019
- When Representations Align: Universality in Representation Learning DynamicsICML 2024
- Gradient Descent Finds Global Minima of Deep Neural NetworksICML 2019
- Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural NetworksICML 2019
- The Implicit Bias of Depth: From Neural Collapse to Softmax CodesICML 2026
- Layer-wise Conditioning Analysis in Exploring the Learning Dynamics of DNNsECCV 2020
- A Convergence Theory for Deep Learning via Over-ParameterizationICML 2019
- Linear CNNs Discover the Statistical Structure of the Dataset Using Only the Most Dominant FrequenciesICML 2023
- The Surprising Simplicity of the Early-Time Learning Dynamics of Neural NetworksNeurIPS 2020
- A Mean Field Analysis Of Deep ResNet And Beyond: Towards Provably Optimization Via Overparameterization From DepthICML 2020
- Learning Overparameterized Neural Networks via Stochastic Gradient Descent on Structured DataNeurIPS 2018
- Asymptotics of feature learning in two-layer networks after one gradient-stepICML 2024
- NAIS-Net: Stable Deep Networks from Non-Autonomous Differential EquationsNeurIPS 2018
- Batch normalization provably avoids ranks collapse for randomly initialised deep networksNeurIPS 2020
- Deeply-Supervised NetsAISTATS 2015
Related open problems
Deep Learning Theory & Optimization Dynamics
Bridging the Activation Divide in Neural Network Optimization Dynamics Theory
Current theoretical analyses of neural network training dynamics are split across incompatible mathematical preconditions: one major body of literature strictly requires high-order smoothness ($C^2$, $C^\infty$, bounded derivatives, or odd symmetry), which explicitly excludes standard piecewise-linear activations like ReLU, while another body relies strictly on piecewise linearity and positive homogeneity, failing to transfer to smooth activations like GELU, SiLU, or tanh. As a result, foundational theoretical predictions—such as edge-of-stability behavior, width-independent convergence rates, and representation learning bounds—cannot be generalized across activation families. Without analytical tools that either handle subgradient boundary crossings or bound the divergence between smooth approximations and exact piecewise-linear trajectories during training, optimization theory remains fragmented into mutually exclusive activation regimes.
Deep Learning Theory & Optimization Dynamics
Cross-Architecture Empirical Validation of Deep Learning Theory Predictors
Theoretical analysis and optimization dynamic guarantees in deep learning are overwhelmingly developed and proved on shallow fully connected ReLU models or two-layer networks. As a result, theoretical claims regarding convergence rates, generalization bounds, and optimization statistics are rarely verified on non-feedforward or deeper architectures such as CNNs, ResNets, and Transformers. Without systematic empirical evaluation across these structural settings, it remains unknown whether these theoretical quantities correlate with actual training dynamics or are merely artifacts of idealized shallow feedforward assumptions.
Deep Learning Theory & Optimization Dynamics
Empirical Benchmarking of Generalization Failure in Alternative Optimization Dynamics
Alternative optimization methods, such as exact Gauss-Newton (GN) and non-backpropagation dynamics (e.g., NMNC), exhibit severe generalization gaps and early loss saturation when scaled to deep networks and mini-batch settings. Standard regularizers (dropout, weight decay, data augmentation, pseudoinverse regularization) and initialization heuristics developed on shallow models consistently fail to close these train-test gaps. Without a controlled comparative evaluation across these distinct settings, it remains unknown whether these generalization failures share common optimization dynamics or require fundamentally different stabilization interventions.