The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Idea - adaptive_dual_scale_denoising

"Name": "adaptive_dual_scale_denoising", "Title": "Adaptive Dual-Scale Denoising for Dynamic Feature Balancing in Low-Dimensional Diffusion Models", "Experiment": "Modify MLPDenoiser to implement a dual-scale processing approach with two parallel branches: a global branch for the original input and a local branch for an upscaled input. Introduce a learnable, timestep- conditioned weighting factor to dynamically balance the contributions of global and local branches. Train models with both the original and new architecture on all datasets. Compare performance using KL divergence and visual inspection of generated samples. Analyze how the weighting factor evolves during the denoising process and its impact on capturing global structure vs. local details across different datasets and timesteps.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Review - adaptive_dual_scale_denoising

"Summary": "The paper introduces an adaptive dual-scale denoising approach for low-dimensional diffusion models, aiming to balance global structure and local details in generated samples. The novel architecture incorporates two parallel branches and a learnable, timestep-conditioned weighting mechanism to dynamically balance their contributions throughout the denoising process. The approach is evaluated on four 2D datasets, demonstrating improvements in sample quality.", "Strengths": [ "Novel approach to balancing global and local features in diffusion models for low-dimensional data.", "Comprehensive empirical evaluation on multiple 2D datasets.", "Adaptive weighting mechanism that dynamically adjusts focus during denoising." ], "Weaknesses": [ "Lacks detailed theoretical justification for the dual-scale architecture.", "Computational cost is significantly higher, which may limit practical applicability.", "Some sections are not clearly explained, such as the autoencoder aggregator and weight evolution analysis.", "Limited diversity in the datasets used for evaluation. More complex, real-world datasets could strengthen claims.", "Insufficient ablation studies and analysis on specific design choices like different types of aggregators." ], "Originality": 4, "Quality": 3, "Clarity": 3, "Significance": 3, "Questions": [ "Can you provide a more detailed theoretical justification for the dual-scale architecture?", "What impact do different types of aggregators have on the model’s performance?", "How does the model perform on more complex, real-world low-dimensional datasets?", "Can the computational cost be reduced without sacrificing performance?" ], "Limitations": [ "The paper should address the high computational cost and explore ways to optimize it.", "The limited diversity of datasets and lack of detailed theoretical backing for the proposed architecture are notable limitations." ], "Ethical Concerns": false, "Soundness": 3, "Presentation": 3, "Contribution": 3, "Overall": 5, "Confidence": 4, "Decision": "Reject"

Seed Idea - batch_size_grokking

"Name": "batch_size_grokking", "Title": "Batch Size Grokking: Assessing the impact of the training batchsize on the grokking phenomenon", "Experiment": "Modify the experiments to dynamically adjust the batch size during training, starting with a small batch size and gradually increasing it. This could potentially lead to faster generalization on the validation set.", "Interestingness": 6, "Feasibility": 4, "Novelty": 4, "novel": true

Idea 1/50 - model_size_grokking

"Name": "model_size_grokking", "Title": "Investigating the Impact of Model Size on the Grokking Phenomenon", "Experiment": "Modify the Transformer class to accept variable number of layers and dimension sizes. Test models with 1, 2, 4, and 8 layers, and dimension sizes of 64, 128, 256, and 512. For each dataset and model size, track the step at which grokking occurs (defined as validation accuracy exceeding 99% and final accuracy vs model size for each task.", "Interestingness": 8, "Feasibility": 7, "Novelty": 7, "novel": true

Idea 2/50 - optimizer_grokking

"Name": "optimizer_grokking", "Title": "Optimization Dynamics and Grokking: Comparing SGD and Adam with Different Learning Rate Schedules", "Experiment": "Modify the training loop to support two optimizers (SGD, Adam) and two learning rate schedules (constant, cosine annealing). For each combination, run multiple experiments with different random seeds. Track validation accuracy, training loss, and L2 norm of weight updates throughout training. Compare the timing and extent of grokking across these optimization strategies for each dataset. Analyze how different optimization dynamics correlate with grokking behavior, including statistical analysis of the results.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 3/50 - biased_data_grokking

"Name": "biased_data_grokking", "Title": "Grokking Under Biased Data: The Effect of Input Range Bias on Neural Network Generalization", "Experiment": "Modify the fetch_train_example method in AbstractDataset to introduce a simple bias: favoring lower-valued inputs. For modular arithmetic operations, sample 70% of the input range. For permutations, favor permutations with more elements in their original positions. Keep the validation set unbiased. Run experiments comparing grokking behavior on biased vs. unbiased training sets. Track metrics such as steps to 99% validation accuracy, and training loss. Analyze how this bias affects grokking across different operations.", "Interestingness": 8, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 4/50 - adaptive_noise_grokking

"Name": "adaptive_noise_grokking", "Title": "Adaptive Noise in Grokking: Investigating Input Perturbations on Algorithmic Learning and Representations", "Experiment": "Modify the GroupDataset class to add operation-specific noise during training: (1) For modular arithmetic, add small integer perturbations. (2) For permutations, occasionally swap two elements. Implement three noise levels (low, medium, high) for each operation. Compare grokking behavior across noise levels and operations, tracking steps to 99% loss. Analyze learned representations by visualizing attention patterns and performing principal component analysis (PCA) on hidden states at different training stages. Compare these representations between noisy and non-noisy training to understand how noise affects the abstraction of concepts.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 5/50 - attention_evolution_grokking

"Name": "attention_evolution_grokking", "Title": "Attention Evolution in Grokking: Quantifying the Transition from Memorization to Generalization", "Experiment": "Modify the Transformer class to output attention weights. Extract and store attention weights at key checkpoints: start, mid- training, grokking point (99% visualization tools for attention heatmaps and create plots showing attention evolution over time. Calculate the Frobenius norm of the difference between attention matrices at consecutive checkpoints to quantify attention evolution. Compare attention patterns and evolution metrics across different operations (modular arithmetic vs. permutations). Analyze attention for specific, informative input sequences to enhance interpretability. Correlate attention evolution metrics with validation accuracy and generalization performance.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 6/50 - local_vs_global_attention_grokking

"Name": "local_vs_global_attention_grokking", "Title": "Local vs Global Attention: Investigating the Impact of Attention Scope on Grokking in Algorithmic Learning", "Experiment": "Modify the DecoderBlock class to support two attention mechanisms: full (global) attention and local attention with a fixed window size. Implement these variants and run experiments across all datasets. Track metrics including time to grokking (99% validation accuracy, and training loss. Calculate and compare ’attention entropy’ for both mechanisms across tasks to quantify attention focus. Analyze how the scope of attention (local vs global) affects grokking behavior and final performance for different algorithmic tasks.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 7/50 - input_encoding_grokking

"Name": "input_encoding_grokking", "Title": "Binary vs One-Hot Encoding: Impact on Grokking in Algorithmic Learning Tasks", "Experiment": "Modify the AbstractDataset class to support two encoding schemes: one-hot (current) and binary. Implement binary encoding for modular arithmetic operations using log2(p) bits, and for permutations using ceil(log2(k!)) bits to represent each permutation uniquely. Adjust the Transformer class to accommodate different input sizes. Run experiments for each encoding scheme across all datasets, tracking metrics such as time to grokking (99% loss, and model memory usage. Analyze how different encoding schemes affect grokking behavior, convergence speed, and final performance for various algorithmic tasks. Compare the impact of input representations on the model’s ability to learn and generalize across different operations. Discuss how findings could inform input representation choices in other machine learning tasks beyond algorithmic learning.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 8/50 - curriculum_learning_grokking

"Name": "curriculum_learning_grokking", "Title": "Curriculum Learning in Grokking: The Effect of Structured Example Progression on Algorithmic Learning", "Experiment": "Modify the AbstractDataset class to implement a simple curriculum learning strategy. For modular arithmetic operations, start with operations involving numbers in the lower half of the range and gradually introduce larger numbers. For permutations, begin with permutations that differ from the identity by one swap and progressively increase the number of swaps. Implement a curriculum scheduler that increases difficulty every 500 steps. Run experiments comparing standard random sampling vs. curriculum learning across all datasets. Track metrics including time to grokking (99% loss. Plot learning trajectories (validation accuracy over time) for both approaches. Compare attention patterns between curriculum and random approaches at different stages of training. Analyze how the curriculum affects the grokking phenomenon across different operations and discuss implications for training neural networks on algorithmic tasks.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 9/50 - weight_init_grokking

"Name": "weight_init_grokking", "Title": "Weight Initialization Strategies and Their Impact on Grokking in Algorithmic Learning", "Experiment": "Modify the Transformer class to support three weight initialization strategies: Xavier/Glorot, Kaiming/He, and random normal (as baseline). Implement these initialization methods for linear layers and embeddings. Run experiments across all datasets for each initialization strategy. Track metrics including time to grokking (99% accuracy), final validation accuracy, training loss, and gradient norm during training. Plot learning curves and compare the distribution of weight values at different stages of training. Analyze the loss landscape by computing gradient variance as a proxy for local geometry at key points during training. Compare how different initialization strategies affect the grokking phenomenon, convergence speed, and final performance across various algorithmic tasks. Investigate potential correlations between initial weight distributions, gradient variance characteristics, and the timing/nature of the grokking transition.", "Interestingness": 9, "Feasibility": 9, "Novelty": 8, "novel": true

Idea 10/50 - task_complexity_grokking

"Name": "task_complexity_grokking", "Title": "Grokking Across Task Complexity: Mapping Neural Network Learning Dynamics to Algorithmic Difficulty", "Experiment": "1. Modify the AbstractDataset class to include new operations of increasing complexity: modular addition, subtraction, multiplication, and exponentiation. 2. Implement these operations in new dataset classes. 3. Quantify task complexity using metrics like algebraic degree and average solution time for humans (estimated). 4. Run experiments for each operation, tracking metrics such as time to grokking (99% validation accuracy), final validation accuracy, training loss, and a new ’complexity-adjusted learning rate’ (validation accuracy improvement per epoch, normalized by task complexity). 5. Plot learning curves and complexity-adjusted learning rates for each operation. 6. Analyze attention patterns and hidden state representations at different stages of training for each operation. 7. Investigate correlations between quantified task complexity and grokking characteristics (e.g., time to grokking, steepness of accuracy improvement).", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 11/50 - regularization_grokking

"Name": "regularization_grokking", "Title": "The Role of Regularization in Grokking: How L2 and Label Smoothing Affect Algorithmic Learning", "Experiment": "1. Implement L2 regularization by adding weight decay to the optimizer. 2. Implement label smoothing in the loss function. 3. Modify the training function to support these regularization techniques with two strength levels each (low and high). 4. Run experiments for each regularization technique and strength across all datasets, including a baseline without regularization. 5. Track metrics: time to grokking (99% validation accuracy), final validation accuracy, training loss, and a new ’grokking speed’ metric (rate of validation accuracy improvement from 50% to 90% for different regularization settings. 7. Analyze how L2 regularization and label smoothing affect the timing, speed, and nature of grokking across various algorithmic tasks, comparing against the non-regularized baseline.", "Interestingness": 9, "Feasibility": 9, "Novelty": 8, "novel": true

Idea 12/50 - grokking_extrapolation

"Name": "grokking_extrapolation", "Title": "Grokking and Extrapolation: Investigating the Limits of Algorithmic Understanding", "Experiment": "1. Modify AbstractDataset to create a separate test set with out-of-distribution examples (e.g., larger numbers for modular arithmetic, longer permutations). 2. Implement a new evaluation function for the test set. 3. During training, periodically evaluate on both validation and test sets. 4. Track metrics: time to grokking on validation set, final validation accuracy, test set accuracy at grokking point, final test set accuracy, and ’extrapolation gap’. 5. Implement visualization of test set performance and extrapolation gap over time, highlighting the grokking point. 6. Compare extrapolation capabilities across different operations and model sizes. 7. Analyze attention patterns on test set inputs before and after grokking. 8. Implement a simple MLP baseline for comparison.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 13/50 - label_noise_grokking

"Name": "label_noise_grokking", "Title": "Grokking Under Noise: The Impact of Systematic and Random Label Errors on Algorithmic Learning", "Experiment": "1. Modify the AbstractDataset class to introduce two types of label noise: random (labels changed randomly) and systematic (specific labels consistently flipped). Add a ’noise_type’ parameter (random/systematic) and ’noise_level’ parameter (low: 5% high: 20% training set while keeping the validation set clean. 3. Run experiments for each noise type and level across all datasets. 4. Track metrics: time to grokking (99% loss, and model confidence (mean softmax probability of correct class). 5. Plot learning curves and model confidence for different noise types and levels, highlighting the grokking point for each. 6. Analyze how different types and levels of label noise affect the timing, speed, and extent of grokking across different operations. 7. Compare attention patterns between noisy and clean training at different stages to understand how the model adapts to noise.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 14/50 - compositional_grokking

"Name": "compositional_grokking", "Title": "Compositional Grokking: Investigating the Relationship Between Grokking and Compositional Learning in Modular Arithmetic", "Experiment": "1. Modify ModSumDataset and ModSubtractDataset to include composite operations: (a + b) - c mod p and (a - b) + c mod p. 2. Implement new dataset class CompositeModDataset for these operations. 3. Run experiments comparing learning curves for basic (a + b, a - b) and composite operations. 4. Track metrics: time to grokking for basic vs. composite operations, correlation between grokking times, final accuracies. 5. Analyze attention patterns to see if the model learns to attend to intermediate results in composite operations. 6. Implement a ’compositional generalization’ test by training on basic operations and testing on their compositions. 7. Compare internal representations (e.g., using PCA on hidden states) for basic vs. composite operations at different stages of training.", "Interestingness": 9, "Feasibility": 6, "Novelty": 9, "novel": true

Idea 15/50 - mutual_information_grokking

"Name": "mutual_information_grokking", "Title": "Information Dynamics in Grokking: Analyzing Mutual Information Evolution During Algorithmic Learning", "Experiment": "1. Implement a function to estimate mutual information using a binning approach for efficiency. 2. Modify the Transformer class to output hidden states from the final layer. 3. Update the training loop to calculate and store mutual information between (a) inputs and outputs, and (b) final hidden states and outputs, at regular intervals. 4. Run experiments across all datasets, tracking these mutual information metrics alongside validation accuracy and training loss. 5. Create plots showing the evolution of both mutual information metrics over training time, highlighting the grokking point. 6. Analyze how mutual information trends relate to grokking by testing specific hypotheses: (a) Rapid increase in hidden state-output mutual information coincides with grokking, (b) Input- output mutual information stabilizes post-grokking. 7. Compare mutual information dynamics between different operations and model sizes to identify common patterns in successful grokking.", "Interestingness": 9, "Feasibility": 6, "Novelty": 9, "novel": true

Idea 16/50 - sparse_subnetworks_grokking

"Name": "sparse_subnetworks_grokking", "Title": "Sparse Subnetworks in Grokking: Investigating the Emergence of Critical Structures During Algorithmic Learning", "Experiment": "1. Implement a simple magnitude-based pruning function for the Transformer model. 2. Modify the training loop to perform pruning at key points: before training, just before grokking (based on validation accuracy), and after grokking. 3. For each pruning point, create sparse networks at different sparsity levels (e.g., 50% these sparse networks from the original initialization for a fixed number of steps. 5. Track metrics: validation accuracy of sparse networks, sparsity level, and grokking speed (if it occurs). 6. Plot the performance of sparse networks at different sparsity levels and pruning points. 7. Compare the structure and performance of sparse networks found before and after grokking across different operations.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 17/50 - positional_encoding_grokking

"Name": "positional_encoding_grokking", "Title": "Inductive Biases in Grokking: The Impact of Positional Encoding Schemes on Algorithmic Learning", "Experiment": "1. Modify the Transformer class to support three positional encoding schemes: sinusoidal (current), learned embeddings, and a simple binary encoding (e.g., [0,1,0,1,0] for ’a o b = c’). 2. Implement these encoding schemes, ensuring they work with the existing sequence length. 3. Run experiments for each encoding scheme across all datasets, tracking: time to grokking (99% training loss, and attention entropy. 4. Analyze how different encoding schemes affect attention patterns and grokking behavior for each operation type. 5. Compare generalization capabilities on sequences with shuffled operands (e.g., ’b o a = c’). 6. Correlate encoding scheme performance with operation complexity to identify potential interactions between input representation and task structure. 7. Discuss implications for designing transformers for specific algorithmic tasks based on findings.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 18/50 - adversarial_robustness_grokking

"Name": "adversarial_robustness_grokking", "Title": "Adversarial Robustness During Grokking: Tracking Vulnerability Evolution in Algorithmic Learning", "Experiment": "1. Implement a simple perturbation method: randomly flip 1-2 bits in the input representation for modular arithmetic, and swap 1-2 elements for permutations. 2. Modify the training loop to generate perturbed inputs and evaluate model performance on them every 500 steps. 3. Track metrics: normal validation accuracy, accuracy on perturbed inputs, and ’robustness gap’ (difference between normal and perturbed accuracy). 4. Plot the evolution of robustness to perturbations alongside the grokking curve. 5. Compare robustness before, during, and after grokking across different operations. 6. Analyze examples of successful perturbations at different stages of training. 7. Investigate potential correlations between the timing of grokking and changes in robustness to perturbations.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 19/50 - critical_periods_grokking

"Name": "critical_periods_grokking", "Title": "Critical Periods in Grokking: The Impact of Timed Learning Rate Spikes on Algorithmic Learning", "Experiment": "1. Modify the training loop to support learning rate spikes at specific points (25% to apply these spikes, increasing the learning rate by 10x for 100 steps. 3. Run experiments for each spike timing across all datasets (modular arithmetic and permutations), including a control group with no spikes. 4. Track metrics: time to grokking, final validation accuracy, and ’spike impact’ (change in validation accuracy slope in 500 steps post-spike). 5. Plot learning curves highlighting spike points and their impacts. 6. Analyze how spike timing affects grokking across different operations, comparing modular arithmetic tasks with permutations. 7. Compare attention patterns immediately before and after impactful spikes. 8. Correlate spike impact with the stage of learning (pre-grokking, during grokking, post- grokking) to identify potential critical periods, assessing whether these periods are task-specific or general across operations.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 20/50 - lottery_tickets_grokking

"Name": "lottery_tickets_grokking", "Title": "Lottery Tickets in Grokking: Investigating Sparse Subnetworks Capable of Algorithmic Learning", "Experiment": "1. Implement an iterative magnitude pruning function for the Transformer model. 2. Modify the training loop to support multiple rounds of train-prune-reset cycles. 3. For each dataset, run experiments with pruning levels of 30% iteration, train the network to convergence, prune the specified percentage of smallest weights, then reset remaining weights to their initial values. 5. Track metrics for each sparse network: time to grokking (or maximum training time if grokking doesn’t occur), final validation accuracy, and training loss. 6. Introduce a ’grokking efficiency’ metric: the ratio of time to grokking for the sparse network vs. the dense network. For networks that don’t grok, use the maximum training time. 7. Plot learning curves for each pruning level, highlighting grokking points and grokking efficiency. 8. Compare the structure of sparse networks that achieve grokking across different operations, focusing on the distribution of preserved weights in different layers and attention heads. 9. Analyze the correlation between pruning level and grokking efficiency across different algorithmic tasks, including cases where grokking fails to occur.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": false

Idea 21/50 - algebraic_structure_grokking

"Name": "algebraic_structure_grokking", "Title": "Grokking and Algebraic Structure: How Group Properties Influence Neural Network Learning", "Experiment": "1. Implement new dataset classes for modular multiplication and division (modulo p, where p is prime, ensuring proper group structures). 2. For each operation (addition, multiplication, division), calculate and store two properties: group order and number of generators. 3. Run experiments for each operation type, tracking: time to grokking, final validation accuracy, and the two group properties. 4. Plot learning curves and grokking points for each operation, labeled with their group properties. 5. Analyze the correlation between group properties and grokking behavior (e.g., time to grokking, steepness of accuracy improvement). 6. Compare attention patterns across operations, focusing on how they reflect the underlying group structure (e.g., uniformity for commutative operations). 7. Test the model’s ability to generalize by evaluating on compositions of learned operations (e.g., a * b + c mod p) after training on individual operations.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 22/50 - mdl_grokking

"Name": "mdl_grokking", "Title": "Minimum Description Length and Grokking: Investigating the Relationship Between Model Compression and Algorithmic Learning", "Experiment": "1. Implement functions to calculate model complexity: (a) L2 norm of weights, (b) number of bits to store parameters at different precisions, (c) effective number of parameters using BIC. 2. Modify the training loop to track these complexity measures alongside existing metrics. 3. Run experiments across all datasets, recording complexity measures, validation accuracy, and training loss at regular intervals. 4. Plot the evolution of model complexity alongside the grokking curve. 5. Analyze the correlation between sudden decreases in model complexity and the onset of grokking, including statistical tests for significance. 6. Compare complexity dynamics across different operations and model sizes. 7. Visualize weight distributions at pre-grokking, during grokking, and post- grokking stages. 8. Implement and compare two early stopping mechanisms: one based on model complexity stabilization and another based on validation loss stabilization.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 23/50 - invariance_learning_grokking

"Name": "invariance_learning_grokking", "Title": "Learning Invariances in Grokking: Tracking Symmetry Awareness During Algorithmic Learning", "Experiment": "1. Modify AbstractDataset to generate transformed versions of inputs (cyclic shifts for modular arithmetic, relabelings for permutations). 2. Update the evaluation function to test model predictions on both original and transformed inputs. 3. Implement an ’invariance score’ metric: mean absolute difference between predictions on original and transformed inputs. 4. Modify the training loop to calculate and store the invariance score at regular intervals. 5. Run experiments across all datasets, tracking the invariance score alongside existing metrics. 6. Plot the evolution of the invariance score alongside the grokking curve. 7. Analyze how the invariance score changes before, during, and after grokking. 8. Compare invariance learning across different operations and model sizes. 9. Investigate correlation between invariance score and generalization performance.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 24/50 - grokking_double_descent

"Name": "grokking_double_descent", "Title": "Grokking and Double Descent: Exploring the Intersection of Two Deep Learning Phenomena", "Experiment": "1. Create a range of model sizes by varying num_layers (1 to 8) and dim_model (32 to 512). 2. For each dataset, train models of different sizes, tracking validation accuracy, training loss, and time to grokking (99% parameters to identify double descent behavior. 4. On the same plot, mark the point where grokking occurs for each model size. 5. Analyze the relationship between grokking timing and the different regimes of the double descent curve (under-parameterized, critical, over-parameterized). 6. Calculate the correlation between model size and time to grokking. 7. Compare double descent and grokking behavior across different operations (modular arithmetic vs. permutations). 8. Investigate whether grokking consistently occurs in a specific regime of the double descent curve.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": false

Idea 25/50 - ntk_alignment_grokking

"Name": "ntk_alignment_grokking", "Title": "NTK-Output Alignment in Grokking: Tracking Feature Learning Dynamics in Algorithmic Tasks", "Experiment": "1. Implement a function to compute the NTK-output alignment: the cosine similarity between the NTK’s top eigenvector and the output gradient. 2. Modify the training loop to compute and store this alignment metric every 100 steps. 3. Run experiments across all datasets, tracking NTK-output alignment alongside validation accuracy and training loss. 4. Plot the evolution of NTK-output alignment alongside the grokking curve. 5. Analyze how the alignment changes before, during, and after grokking, identifying any consistent patterns across different operations. 6. Investigate correlations between sudden changes in alignment and the onset of grokking. 7. Compare alignment dynamics for models that achieve grokking vs. those that don’t. 8. Experiment with using the alignment metric as an early stopping criterion or to adjust learning rates dynamically. 9. Discuss implications of findings for understanding feature learning and generalization in grokking.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 26/50 - loss_landscape_grokking

"Name": "loss_landscape_grokking", "Title": "Loss Landscape Evolution in Grokking: Geometric Insights into Algorithmic Learning", "Experiment": "1. Implement functions to compute and visualize 2D loss landscapes using filter-wise normalization. 2. Modify the training loop to save model checkpoints at key points: start of training, 25% just before grokking (based on validation accuracy), during grokking, and after grokking. 3. For each checkpoint, compute and store 2D loss landscape visualizations. 4. Define quantitative metrics for loss landscape characteristics: (a) local smoothness (average gradient magnitude), (b) global convexity (ratio of loss at edges to center), (c) barrier height (maximum loss along minimum loss path). 5. Run experiments across all datasets, generating loss landscapes and computing metrics at key points. 6. Create side-by-side comparisons of loss landscapes at different stages of training for each operation. 7. Analyze how loss landscape metrics change before, during, and after grokking. 8. Compare loss landscape evolution between operations that grok quickly vs. slowly. 9. Investigate correlations between changes in loss landscape metrics and the onset of grokking.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 27/50 - neural_collapse_grokking

"Name": "neural_collapse_grokking", "Title": "Neural Collapse in Grokking: Investigating Feature Geometry During Algorithmic Learning", "Experiment": "1. Modify Transformer to output final layer features. 2. Implement functions to compute class means and covariances. 3. Calculate simplified neural collapse metrics: (a) average cosine similarity between class means, (b) ratio of within-class to between-class variances. 4. Track these metrics every 500 steps during training. 5. Run experiments on modular arithmetic and permutation datasets. 6. Plot neural collapse metrics alongside grokking curves. 7. Analyze changes in metrics before, during, and after grokking. 8. Compare neural collapse dynamics between operations that grok quickly vs. slowly. 9. Visualize class mean trajectories in 2D/3D using PCA. 10. Discuss implications for understanding both grokking and general neural network learning dynamics.", "Interestingness": 9, "Feasibility": 6, "Novelty": 9, "novel": true

Idea 28/50 - data_augmentation_grokking

"Name": "data_augmentation_grokking", "Title": "Data Augmentation in Grokking: The Impact of Input Transformations on Algorithmic Learning", "Experiment": "1. Implement task-specific augmentations: (a) For modular arithmetic: add random offsets (mod p) to inputs. (b) For permutations: apply random permutations to inputs and outputs. 2. Modify GroupDataset to apply augmentations with 0% for each augmentation level across all datasets. 4. Track metrics: time to grokking (99% ’augmentation generalization gap’ (difference between augmented and non- augmented validation accuracy). 5. Plot learning curves and generalization gaps for each augmentation level. 6. Analyze the correlation between augmentation level and grokking speed. 7. Compare attention patterns between augmentation levels to understand representation changes. 8. Discuss implications for designing data augmentation strategies in algorithmic learning tasks.", "Interestingness": 9, "Feasibility": 9, "Novelty": 8, "novel": true

Idea 29/50 - emergent_grokking

"Name": "emergent_grokking", "Title": "Emergent Abilities in Grokking: Investigating Scale-Dependent Algorithmic Learning", "Experiment": "1. Modify existing datasets to include ’simple’ and ’complex’ versions (e.g., mod sum with small vs. large primes). 2. Adjust Transformer class to scale from tiny (1 layer, 64 dim) to medium (4 layers, 512 dim). 3. For each operation, train models of increasing size, tracking grokking time and performance on both simple and complex versions. 4. Implement a generalization test for each operation (e.g., mod sum with even larger primes). 5. Plot learning curves for different model sizes, highlighting grokking points. 6. Create heatmaps of model size vs. operation complexity, showing grokking time and generalization test results. 7. Perform statistical analysis to identify significant jumps in performance across model sizes, using metrics such as accuracy increase rate and time to reach 99% behavior patterns across different operation types.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 30/50 - functional_modularity_grokking

"Name": "functional_modularity_grokking", "Title": "Functional Modularity in Grokking: Analyzing Emergent Specialization in Transformer Networks During Algorithmic Learning", "Experiment": "1. Implement functions to track weight update patterns and attention focus for each layer and head. 2. Modify the training loop to compute and store these metrics at regular intervals. 3. Define a ’functional modularity score’ based on the consistency of weight updates and attention patterns for specific input types. 4. Run experiments across all datasets, tracking the functional modularity score alongside existing metrics. 5. Plot the evolution of functional modularity alongside the grokking curve. 6. Analyze how functional modularity changes before, during, and after grokking. 7. Visualize the most consistent patterns at different stages of training and interpret their functions. 8. Compare functional modularity dynamics between different operations and model sizes. 9. Investigate correlations between functional modularity and grokking speed or generalization performance.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 31/50 - information_compression_grokking

"Name": "information_compression_grokking", "Title": "Information Compression in Grokking: Analyzing Representational Dynamics During Algorithmic Learning", "Experiment": "1. Modify Transformer class to include a bottleneck layer (smaller dimension linear layer) after the encoder. 2. Implement function to compute activation sparsity (% bottleneck layer. 3. Update training loop to compute and store activation sparsity and gradient magnitudes of the bottleneck layer at regular intervals. 4. Run experiments with different bottleneck sizes (e.g., 25% 50% to grokking, final validation accuracy, activation sparsity, and gradient magnitudes. 6. Plot activation sparsity and gradient magnitude evolution alongside grokking curves for each bottleneck size. 7. Analyze how these metrics change before, during, and after grokking. 8. Test generalization by evaluating models on slightly out-of-distribution examples (e.g., larger numbers in modular arithmetic). 9. Investigate correlation between optimal compression (measured by activation sparsity) and grokking speed, generalization performance.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 32/50 - critical_learning_periods_grokking

"Name": "critical_learning_periods_grokking", "Title": "Critical Learning Periods in Grokking: Temporal Dynamics of Algorithmic Understanding", "Experiment": "1. Modify the training loop to support ’intervention periods’ where learning rate is increased by 5x for 100 steps. 2. Implement a sliding window intervention strategy, with windows of 500 steps, starting every 250 steps. 3. Run experiments for each window across all datasets and three model sizes (small, medium, large), including a control group with no interventions. 4. Track metrics: time to grokking, final validation accuracy, and ’intervention impact’ (area under the validation accuracy curve for 500 steps post-intervention). 5. Plot learning curves highlighting intervention windows and their impacts. 6. Create heatmaps visualizing intervention impact across time windows and model sizes for each operation. 7. Analyze how intervention timing affects grokking across different operations and model sizes. 8. Compare attention patterns immediately before and after impactful interventions. 9. Investigate whether certain operations or model sizes have more pronounced critical periods than others. 10. Discuss implications for curriculum design in machine learning and potential applications in continual and transfer learning.", "Interestingness": 9, "Feasibility": 7, "Novelty": 9, "novel": true

Idea 33/50 - simplicity_bias_grokking

"Name": "simplicity_bias_grokking", "Title": "Simplicity Bias in Grokking: Analyzing Weight Matrix Complexity During Algorithmic Learning", "Experiment": "1. Modify AbstractDataset to include two complexity levels for each operation (e.g., small vs. large prime for modular arithmetic, short vs. long permutations). 2. Implement a function to compute the effective rank of weight matrices using singular value decomposition. 3. Update the training loop to compute and store the effective rank for each layer every 500 steps. 4. Run experiments across all datasets and both complexity levels, tracking effective rank alongside existing metrics. 5. Plot the evolution of effective rank alongside grokking curves for each complexity level and operation. 6. Analyze how effective rank changes before, during, and after grokking, and how this relates to task complexity. 7. Investigate correlations between effective rank dynamics and grokking speed or generalization performance. 8. Compare effective rank patterns across different operations and model sizes. 9. Contrast effective rank dynamics between operations that grok quickly versus those that grok slowly or fail to grok. 10. Experiment with using effective rank as an indicator for the onset of grokking.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 34/50 - lucky_initializations_grokking

"Name": "lucky_initializations_grokking", "Title": "Lucky Initializations in Grokking: Identifying and Analyzing Favorable Starting Points for Algorithmic Learning", "Experiment": "1. Implement a function to generate and store 50 random initializations for the Transformer model. 2. Modify the training loop to support training from stored initializations and different learning rates. 3. For each dataset, train models from the 50 initializations with 3 learning rates, tracking ’grokking efficiency’ (ratio of validation accuracy to training steps at 99% initializations (top 20% characteristics of lucky initializations: weight distribution statistics, layerwise norms, and attention pattern initialization. 6. Implement a function to visualize the loss landscape around initial points using filter-wise normalization. 7. Compare lucky initializations across different operations to identify common patterns. 8. Develop a simple predictor for initialization ’luckiness’ based on identified characteristics. 9. Test transfer of lucky initializations across tasks and learning rates.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 35/50 - relative_attention_grokking

"Name": "relative_attention_grokking", "Title": "Relative Positional Attention and Its Impact on Grokking in Algorithmic Learning", "Experiment": "1. Modify the DecoderBlock class to support two attention types: standard (current) and relative positional. 2. Implement relative positional attention, ensuring it works with the existing sequence length. 3. Update the Transformer class to accept an attention_type parameter. 4. Run experiments for both attention types across all datasets, tracking: time to grokking (99% training loss, grokking transition sharpness (rate of validation accuracy increase), and post-grokking stability (variance in validation accuracy after reaching 99% highlighting grokking points and transition periods. 6. Visualize and compare attention patterns between the two mechanisms at key stages: pre- grokking, during grokking transition, and post-grokking. 7. Analyze how relative positional attention affects grokking behavior, transition sharpness, and stability for each operation type compared to standard attention. 8. Investigate correlations between attention type and grokking speed or post-grokking stability. 9. Discuss implications for designing transformers for specific algorithmic tasks based on findings.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 36/50 - grokking_task_interference

"Name": "grokking_task_interference", "Title": "Grokking and Task Interference: Exploring the Stability of Algorithmic Understanding", "Experiment": "1. Modify the training loop to support learning two modular arithmetic operations sequentially (e.g., addition then multiplication). 2. Implement a task scheduler that switches between tasks at regular intervals. 3. Create a ’dual-task evaluation’ function to assess performance on both tasks simultaneously. 4. Track metrics: time to grokking for each task, performance on the first task while learning the second, and a ’grokking stability’ score (maintenance of >95% task 1 while learning task 2). 5. Run experiments with different task switching frequencies. 6. Analyze how grokking on one task affects learning speed and grokking on the subsequent task. 7. Visualize attention patterns before and after introducing the second task to understand representation changes. 8. Investigate the correlation between grokking speed on the first task and stability of that understanding when learning the second task. 9. Compare results with a baseline of learning both tasks simultaneously to isolate the effects of sequential learning.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 37/50 - attention_inductive_bias_grokking

"Name": "attention_inductive_bias_grokking", "Title": "Inductive Biases in Attention Mechanisms: Their Impact on Grokking in Algorithmic Learning", "Experiment": "1. Modify DecoderBlock class to support two attention mechanisms: standard dot-product and additive (Bahdanau). 2. Implement these attention mechanisms, ensuring compatibility with existing architecture. 3. Update Transformer class to accept an attention_type parameter. 4. Select a subset of most illustrative datasets based on preliminary experiments. 5. Run experiments for each attention type on selected datasets, tracking: time to grokking (99% final validation accuracy, training loss, and ’grokking transition sharpness’ (defined as the maximum rate of validation accuracy increase over any 500-step window). 6. Implement a simple generalization test using slightly out-of-distribution examples (e.g., larger numbers for modular arithmetic). 7. Plot learning curves for each attention type, highlighting grokking points and transition periods. 8. Analyze how different attention mechanisms affect grokking behavior, transition sharpness, and generalization performance for each operation type. 9. Visualize attention patterns for each mechanism at key stages: pre-grokking, during grokking transition, and post-grokking. 10. Discuss implications for designing transformers with appropriate inductive biases for specific types of algorithmic learning tasks.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 38/50 - gradient_dynamics_grokking

"Name": "gradient_dynamics_grokking", "Title": "Gradient Dynamics in Grokking: Analyzing Information Flow Efficiency During Algorithmic Learning", "Experiment": "1. Modify the training loop to compute gradient statistics (sparsity and magnitude distribution) for each layer. 2. Implement functions to calculate gradient sparsity (% magnitude percentiles. 3. Update training process to store these metrics every 500 steps. 4. Run experiments across all datasets, tracking gradient metrics alongside existing performance metrics. 5. Plot the evolution of gradient sparsity and magnitude distributions alongside grokking curves. 6. Analyze how gradient dynamics change before, during, and after grokking. 7. Compare gradient patterns between operations that grok quickly vs. slowly. 8. Investigate correlations between changes in gradient dynamics and grokking speed or generalization performance. 9. Visualize gradient flow patterns at key stages: pre-grokking, during grokking transition, and post- grokking.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 39/50 - adaptive_curriculum_grokking

"Name": "adaptive_curriculum_grokking", "Title": "Adaptive Curriculum Learning in Grokking: Optimizing Example Difficulty for Efficient Algorithmic Understanding", "Experiment": "1. Modify AbstractDataset to include a difficulty scoring function (e.g., input magnitude for modular arithmetic, cycle length for permutations). 2. Implement adaptive sampling strategy: start with easiest 20% level exceeds 90% tracking difficulty of selected examples. 4. Run experiments comparing adaptive curriculum, random sampling, and static curriculum (increasing difficulty linearly) across all datasets. 5. Track metrics: time to grokking, final validation accuracy, learning trajectory smoothness, and example difficulty distribution over time. 6. Analyze relationship between difficulty progression and grokking onset. 7. Visualize learning curves and difficulty progression for each strategy. 8. Compare consistency and speed of grokking across different random seeds for each strategy. 9. Analyze computational efficiency by comparing total number of examples needed to achieve grokking for each strategy. 10. Compare attention patterns at key points (pre-grokking, during grokking, post-grokking) across strategies to understand how adaptive curriculum affects internal representations.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 40/50 - task_structure_grokking

"Name": "task_structure_grokking", "Title": "Task Structure and Grokking: Investigating the Relationship Between Algorithmic Complexity and Learning Dynamics", "Experiment": "1. Modify AbstractDataset to include a ’structural_complexity’ score based on: a) number of unique outputs, b) input-output correlation, c) algebraic degree for modular operations or cycle structure for permutations. 2. Extend existing dataset classes to include a wider range of operations (e.g., modular addition, multiplication, exponentiation; simple and complex permutations). 3. Run experiments across all operations, tracking time to grokking, final validation accuracy, and learning curve smoothness. 4. Plot grokking metrics against structural complexity scores, comparing trends between modular arithmetic and permutation tasks. 5. Analyze correlation between structural complexity and grokking behavior. 6. Compare attention patterns and gradient flows across tasks of different complexity. 7. Implement a generalization test where models trained on simpler structures are evaluated on more complex ones. 8. Discuss implications for neural network learning on structured vs. unstructured tasks in general machine learning contexts.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 41/50 - numerical_base_grokking

"Name": "numerical_base_grokking", "Title": "Numerical Base and Grokking: How Input Representation Affects Pattern Recognition in Algorithmic Learning", "Experiment": "1. Modify AbstractDataset and modular arithmetic dataset classes to support binary and decimal bases. 2. Implement functions to convert between bases and adjust the encode/decode methods. 3. Update the Transformer class to handle variable input lengths. 4. Run experiments for binary and decimal bases on modular addition and multiplication tasks. 5. Track metrics: time to grokking (99% accuracy, training loss, and ’cross-base generalization’ (accuracy when testing on the other base). 6. Plot learning curves for each base, highlighting grokking points. 7. Compare learning curves for binary (0-3) vs decimal (0-9) to isolate base effects from sequence length. 8. Analyze how different bases affect grokking speed and pattern recognition. 9. Compare attention patterns across bases at key stages: pre-grokking, during grokking, and post-grokking. 10. Discuss implications for choosing input representations in mathematical machine learning tasks.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 42/50 - activation_function_grokking

"Name": "activation_function_grokking", "Title": "Activation Functions and Grokking: Investigating the Role of Non- linearity in Algorithmic Learning and Generalization", "Experiment": "1. Modify the DecoderBlock class to support multiple activation functions (ReLU, GELU, Tanh). 2. Update the Transformer class to accept an activation_type parameter, allowing for both uniform and hybrid activation setups. 3. Run experiments comparing the baseline (GELU) with ReLU, Tanh, and a hybrid setup (ReLU in lower layers, Tanh in upper layers) across all datasets. 4. Track metrics: time to grokking (99% accuracy), final validation accuracy, training loss, ’grokking transition sharpness’, and gradient flow statistics. 5. Plot learning curves for each activation setup, highlighting grokking points and transition periods. 6. Visualize decision boundaries at different training stages for each activation setup. 7. Analyze how different activation functions affect grokking behavior, transition sharpness, and final performance for each operation type. 8. Compare hidden representations (using t-SNE) across activation setups at key stages: pre-grokking, during grokking transition, and post-grokking. 9. Investigate the relationship between activation function properties and the trade-off between memorization and generalization. 10. Discuss implications for choosing activation functions in tasks requiring pattern discovery and generalization beyond algorithmic learning.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 43/50 - phase_transition_grokking

"Name": "phase_transition_grokking", "Title": "Grokking as a Phase Transition: Characterizing Critical Behavior in Algorithmic Learning", "Experiment": "1. Implement functions to track key metrics: validation accuracy, training loss, gradient norm, and weight norm. 2. Modify training loop to compute and store these metrics every 100 steps. 3. Run experiments across all datasets, with finer-grained tracking (every 10 steps) around the suspected grokking point. 4. Implement analysis tools to detect sudden changes or discontinuities in metrics. 5. Plot all metrics on a single, multi-axis graph to visualize potential phase transitions. 6. Calculate susceptibility using fluctuations in validation accuracy near the grokking point. 7. Analyze scaling behavior of susceptibility to identify critical exponents, if any. 8. Compare phase transition characteristics across different operations and model sizes. 9. Investigate whether manipulating learning rate or gradient clipping can induce or prevent grokking phase transitions.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": false

Idea 44/50 - effective_dimension_grokking

"Name": "effective_dimension_grokking", "Title": "Effective Dimension Dynamics in Grokking: Analyzing Representational Complexity During Algorithmic Learning", "Experiment": "1. Implement functions to compute the rank and top-k singular values of weight matrices. 2. Modify the training loop to compute and store these metrics every 500 steps for each layer. 3. Run experiments across all datasets, tracking rank and singular value distributions alongside existing performance metrics. 4. Implement a simple MLP baseline that doesn’t exhibit grokking for comparison. 5. Plot the evolution of rank and singular value distributions alongside grokking curves for both Transformer and MLP models. 6. Analyze how these metrics change before, during, and after grokking in the Transformer, contrasting with the MLP. 7. Compare rank dynamics between operations that grok quickly vs. slowly. 8. Investigate correlations between changes in rank/singular values and grokking speed or generalization performance. 9. Visualize the relationship between these metrics and other performance indicators at different stages of training.", "Interestingness": 9, "Feasibility": 8, "Novelty": 9, "novel": true

Idea 45/50 - representation_entropy_grokking

"Name": "representation_entropy_grokking", "Title": "Representation Entropy in Grokking: Tracking the Simplification of Learned Concepts", "Experiment": "1. Implement a function to compute the entropy of the model’s internal representations. 2. Modify the Transformer class to output intermediate representations. 3. Update the training loop to compute and store the representation entropy every 500 steps. 4. Run experiments across all datasets, including configurations that lead to successful grokking and those that don’t (e.g., by varying model size or learning rate). 5. Track entropy alongside existing performance metrics. 6. Plot the evolution of representation entropy alongside grokking curves for both successful and unsuccessful cases. 7. Analyze how representation entropy changes before, during, and after grokking in successful cases, and compare with unsuccessful cases. 8. Investigate correlations between changes in representation entropy and grokking speed or generalization performance. 9. Visualize the relationship between entropy and other performance indicators at different stages of training. 10. Plot entropy distributions across different layers of the model to understand how different parts contribute to concept simplification.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 46/50 - mutual_information_grokking

"Name": "mutual_information_grokking", "Title": "Mutual Information Dynamics in Grokking: Tracing Information Flow During Algorithmic Learning", "Experiment": "1. Modify Transformer class to output representations from input embedding, middle layer, and final layer. 2. Implement MINE (Mutual Information Neural Estimation) for efficient mutual information approximation. 3. Update training loop to compute and store mutual information estimates between input-middle, input-output, and middle-output every 500 steps. 4. Run experiments across all datasets, tracking mutual information alongside existing performance metrics. 5. Plot the evolution of mutual information alongside grokking curves and generalization gap. 6. Analyze how mutual information changes before, during, and after grokking, particularly in relation to the generalization gap. 7. Compare mutual information dynamics between operations that grok quickly vs. slowly. 8. Investigate correlations between changes in mutual information and grokking speed or generalization performance.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 47/50 - lottery_tickets_grokking

"Name": "lottery_tickets_grokking", "Title": "Lottery Tickets in Grokking: Sparse Subnetworks and Sudden Generalization", "Experiment": "1. Implement iterative magnitude pruning for the Transformer model. 2. Modify training loop for train-prune-reset cycles. 3. For each dataset, run experiments with pruning levels of 50% iterations. 4. Track metrics: time to grokking, final validation accuracy, training loss, and ’grokking efficiency’ (ratio of time to grokking for sparse vs. dense network). 5. Plot learning curves for each pruning level, highlighting grokking points. 6. Compare sparse network structures that achieve grokking across operations. 7. Analyze correlation between pruning level and grokking efficiency. 8. Implement simple MLP baseline without grokking for comparison. 9. Visualize weight distributions of winning tickets pre- and post-grokking.", "Interestingness": 9, "Feasibility": 9, "Novelty": 8, "novel": false

Idea 48/50 - architecture_inductive_bias_grokking

"Name": "architecture_inductive_bias_grokking", "Title": "Architectural Inductive Biases and Grokking: Comparing Sudden Generalization Across Neural Network Types", "Experiment": "1. Implement simplified 1D CNN and LSTM model classes compatible with existing sequence-based datasets. 2. Modify training loop to support multiple model types. 3. Run experiments comparing Transformer, 1D CNN, and LSTM models across modular arithmetic datasets. 4. Track metrics: time to grokking, final validation accuracy, training loss, and architecture-specific indicators (attention patterns for Transformer, filter activations for CNN, forget gate activations for LSTM). 5. Plot learning curves for each architecture, highlighting grokking points. 6. Analyze how different architectures affect grokking behavior, speed, and final performance for each operation type. 7. Compare internal representations (using t-SNE) across architectures at key stages: pre- grokking, during grokking transition, and post-grokking. 8. Investigate the relationship between architectural inductive biases and the trade-off between memorization and generalization in modular arithmetic tasks.", "Interestingness": 9, "Feasibility": 8, "Novelty": 8, "novel": true

Idea 49/50 - shortcut_learning_grokking

"Name": "shortcut_learning_grokking", "Title": "Shortcut Learning and Grokking: The Interplay Between Surface Patterns and Deep Understanding in Algorithmic Learning", "Experiment": "1. Modify AbstractDataset to include operation-specific shortcuts: for modular arithmetic, make the result always even if the first operand is even; for permutations, always swap the first two elements. 2. Implement a function to gradually remove these shortcuts over training by reducing their frequency. 3. Update the training loop to apply the shortcut removal function. 4. Add a ’shortcut reliance’ metric: the accuracy difference between shortcut-following and shortcut-violating examples. 5. Run experiments with varying shortcut removal rates across datasets. 6. Track metrics: time to grokking, final validation accuracy, shortcut reliance over time, and performance on a shortcut-free test set. 7. Plot learning curves and shortcut reliance alongside grokking curves. 8. Analyze how shortcut presence and removal affect grokking timing and quality. 9. Compare attention patterns between models trained with and without shortcuts at key stages.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true

Idea 50/50 - grokking_forgetting_complexity

"Name": "grokking_forgetting_complexity", "Title": "Grokking and Forgetting: The Interplay of Task Complexity and Sudden Generalization in Algorithmic Learning", "Experiment": "1. Modify ModSumDataset to support multiple complexity levels (e.g., modular addition with increasing prime moduli). 2. Update the training loop to gradually introduce higher complexity levels while continuously evaluating on all levels. 3. Implement a ’multi-complexity evaluation’ function to assess performance across all complexity levels simultaneously. 4. Track metrics: time to grokking for each complexity level, performance on lower complexity levels when grokking occurs on a higher level, and a ’complexity forgetting score’ (decrease in accuracy on lower complexity levels). 5. Analyze the correlation between grokking events and performance changes on other complexity levels. 6. Compare internal representations (using cosine similarity of hidden states) across complexity levels before and after grokking events. 7. Investigate trends in grokking speed across increasing complexity levels. 8. Plot learning curves for all complexity levels simultaneously, highlighting grokking points and potential forgetting events. 9. Visualize the evolution of representation similarities over time using heatmaps.", "Interestingness": 9, "Feasibility": 9, "Novelty": 9, "novel": true