Project in collaboration with Vicki Li and Pawan Jayakumar for a Program Synthesis course at UCSD (taught by Loris D’Antoni).
<aside> 💡
We attempted to implement a neurosymbolic architecture to solve the ARC-AGI challenge by applying a lightweight (yet expressive) domain specific language (DSL) on the ‘object-level’ within ARC problems.
Several elements were successful — object-level segmentation, learning semantically meaningful object embeddings, generating synthetic data using our DSL, applying a graph-bijection solver — but our ultimate ambition fell short due to a key neural network in our proposed approach failing to learn non-trivial mappings.
📄 Formal write-up here.
</aside>
Visualization of attention logits from the class token in our embedding model.
Problems we were able to solve with the naive graph bijection solver (that uses our object-specific DSL).