​​Neural Algorithmic Reasoning Without Intermediate Supervision

Guest post by Gleb Rodionov

📝 Paper: https://openreview.net/forum?id=vBwSACOB3x (NeurIPS 2023)
🛠️ Code: in the supplementary on OpenReview

Algorithmic reasoning aims to capture computations with neural networks, imitating the execution of classical algorithms. Typically, the generalization abilities of such models are improved through various forms of intermediate supervision, which demonstrate a particular execution trajectory (a sequence of intermediate steps, called hints) that the model needs to follow.

However, progress can also be made on the other side of the spectrum, where models are trained only with input-output pairs. Such models are not tied to any particular execution trajectory and are free to converge to the optimal execution flow for their own architecture. We demonstrate that models without hints can be competitive with hint-based models or even outperform them:

1️⃣ We propose several architectural modifications for models trained without intermediate supervision, that are aimed at making the comparison versus hint-based models clearer and fairer.

2️⃣ We build a self-supervised objective that can regularize intermediate computations of the model without access to the algorithm trajectory.

We hope our work will encourage further investigation of neural algorithmic reasoners without intermediate supervision. For more details, see the blog post.