Add Sketch RNN template. (SakanaAI#143)

* Add sketch RNN template.

* Add general description of experiment.py

* Add relevant citations to template.tex

* Update README.md

---------

Co-authored-by: Cong Lu <[email protected]>

README.md

@@ -280,6 +280,7 @@ We welcome community contributions in the form of new templates. While these are
 
 - Infectious Disease Modeling (`seir`) - [PR #137](https://github.com/SakanaAI/AI-Scientist/pull/137)
 - Image Classification with MobileNetV3 (`mobilenetV3`) - [PR #141](https://github.com/SakanaAI/AI-Scientist/pull/141)
+- Sketch RNN (`sketch_rnn`) - [PR #143](https://github.com/SakanaAI/AI-Scientist/pull/143)
 
 *This section is reserved for community contributions. Please submit a pull request to add your template to the list! Please describe the template in the PR description, and also show examples of the generated papers.*
 

templates/sketch_rnn/ideas.json

@@ -0,0 +1,38 @@
+[
+    {
+        "Name": "latent_space_decorrelation",
+        "Title": "Enhancing Sketch Diversity through Latent Space Decorrelation",
+        "Experiment": "Introduce a covariance penalty term in the loss function that encourages the covariance matrix of the latent vectors to be close to the identity matrix. This can be implemented by adding a term to the loss that penalizes the off-diagonal elements of the covariance matrix of the latent vectors. Modify the `train` method to include this additional regularization term in the loss computation.",
+        "Interestingness": 8,
+        "Feasibility": 9,
+        "Novelty": 7,
+        "novel": true
+    },
+    {
+        "Name": "temporal_smoothing_loss",
+        "Title": "Enhancing Sketch Generation with Temporal Smoothing Loss",
+        "Experiment": "Introduce a Temporal Smoothing Loss (TSL) to the existing model to encourage smooth transitions in pen movements. Define TSL as the L2 norm of the difference between consecutive pen movements. Modify the `train` method to include TSL in the loss computation. Implement an adaptive weighting mechanism that adjusts the influence of TSL based on the variance of pen movements within a sequence. Evaluate the quality and diversity of the generated sketches with and without TSL.",
+        "Interestingness": 7,
+        "Feasibility": 9,
+        "Novelty": 7,
+        "novel": true
+    },
+    {
+        "Name": "attention_mechanism",
+        "Title": "Enhancing Sketch Generation with Attention Mechanism",
+        "Experiment": "In this experiment, we introduce an attention mechanism to the decoder RNN. This involves adding an attention layer that computes a context vector at each decoding step by taking a weighted sum of the latent representations. The context vector is concatenated with the decoder's hidden state and the current input to generate the next output. Modify the `DecoderRNN` class to include the attention mechanism and adjust the `forward` method accordingly. Evaluate the quality and diversity of the generated sketches with and without the attention mechanism.",
+        "Interestingness": 9,
+        "Feasibility": 8,
+        "Novelty": 8,
+        "novel": true
+    },
+    {
+        "Name": "adaptive_dropout",
+        "Title": "Improving Sketch Generation with Adaptive Dropout",
+        "Experiment": "Introduce an adaptive dropout mechanism that adjusts the dropout rate dynamically based on the training step. Implement a `dropout_scheduler` function that decreases the dropout rate from 0.5 to 0.1 linearly over the training steps. Modify the `EncoderRNN` and `DecoderRNN` classes to accept a dynamic dropout rate parameter. Update the dropout rate in each forward pass of the encoder and decoder based on the current training step. Integrate the `dropout_scheduler` into the training loop, updating the dropout rates in each step. Evaluate the quality and diversity of the generated sketches with and without adaptive dropout.",
+        "Interestingness": 8,
+        "Feasibility": 8,
+        "Novelty": 7,
+        "novel": true
+    }
+]