Introduction It's now well known that simple neural network models often "grok" algorithmic tasks. That is, when trained for many epochs on a subset of the full input space, the model quickly attains perfect train accuracy and then, much later, near-perfect test accuracy. In the former phase, the model memorizes...
The slingshot effect is a late-stage training anomaly found in various adaptive gradient optimization methods. In particular, slingshots are present with AdamW, the optimizer most widely used for modern transformer training. The original slingshot paper observes that slingshots tend to occur alongside grokking, a phenomenon in which neural networks trained...
It's now well known that simple neural network models often "grok" algorithmic tasks. That is, when trained for many epochs on a subset of the full input space, the model quickly attains perfect train accuracy and then, much later, near-perfect test accuracy. In the former phase, the model memorizes the training set; in the latter, it generalizes out-of-distribution to the test set.
In the algorithmic grokking literature, there is typically exactly one natural generalization from the training set to the test set. What if, however, the training set were instead under-specified in such a way that there were multiple possible generalizations? Would the model grok at all? If so, which of the generalizing... (read 3076 more words →)
The slingshot effect is a late-stage training anomaly found in various adaptive gradient optimization methods. In particular, slingshots are present with AdamW, the optimizer most widely used for modern transformer training. The original slingshot paper observes that slingshots tend to occur alongside grokking, a phenomenon in which neural networks trained on algorithmic tasks generalize to the test set long after perfectly fitting the training set. In this post we take a closer look at slingshots and their effect on generalization in the setting of 1-hidden-layer MLPs trained on -sparse parity, a specific algorithmic task. The main results are 1) an explanation of why slingshots occur in models trained with hinge loss that partially transfers to... (read 2348 more words →)
For that earlier section, we used smaller models trained on intersect (4,000 parameters) instead of intersect (80,000 parameters) -- the only reason for this was to allow for a larger sample size of 10,000 models with our compute budget. All subsequent sections use the models.
