Dated Data: Tracing Knowledge Cutoffs in Large Language Models

Released Large Language Models (LLMs) are often paired with a claimed knowledge cutoff date, or the dates at which training data was gathered. Such information is crucial for applications where the LLM must provide up to date information. However, this statement only scratches the surface: do all resources in the training data share the same knowledge cutoff date? Does the model's demonstrated knowledge for these subsets closely align to their cutoff dates? In this work, we define the notion of an effective cutoff. This is distinct from the LLM designer reported cutoff and applies separately to sub-resources and topics. We propose a simple approach to estimate effective cutoffs on the resource-level temporal alignment of an LLM by probing across versions of the data. Using this analysis, we find that effective cutoffs often differ from reported cutoffs. To understand the root cause of this observation, we conduct a direct large-scale analysis on open pre-training datasets. Our analysis reveals two reasons for these inconsistencies: (1) temporal biases of CommonCrawl data due to non-trivial amounts of old data in new dumps and (2) complications in LLM deduplication schemes involving semantic duplicates and lexical near-duplicates. Overall, our results show that knowledge cutoffs are not as simple as they have seemed and that care must be taken both by LLM dataset curators as well as practitioners who seek to use information from these models.

Augmenting Supervised Learning by Meta-learning Unsupervised Local Rules

The brain performs unsupervised learning and (perhaps) simultaneous supervised learning. This raises the question as to whether a hybrid of supervised and unsupervised methods will produce better learning. Inspired by the rich space of Hebbian learning rules, we set out to directly learn the unsupervised learning rule on local information that best augments a supervised signal. We present the Hebbian-augmented training algorithm (HAT) for combining gradient-based learning with an unsupervised rule on pre-synpatic activity, post-synaptic activities, and current weights. We test HAT's effect on a simple problem (Fashion-MNIST) and find consistently higher performance than supervised learning alone. This finding provides empirical evidence that unsupervised learning on synaptic activities provides a strong signal that can be used to augment gradient-based methods. We further find that the meta-learned update rule is a time-varying function; thus, it is difficult to pinpoint an interpretable Hebbian update rule that aids in training. We do find that the meta-learner eventually degenerates into a non-Hebbian rule that preserves important weights so as not to disturb the learner's convergence.

Bilingual is At Least Monolingual (BALM): A Novel Translation Algorithm that Encodes Monolingual Priors

State-of-the-art machine translation (MT) models do not use knowledge of any single language's structure; this is the equivalent of asking someone to translate from English to German while knowing neither language. BALM is a framework incorporates monolingual priors into an MT pipeline; by casting input and output languages into embedded space using BERT, we can solve machine translation with much simpler models. We find that English-to-German translation on the Multi30k dataset can be solved with a simple feedforward network under the BALM framework with near-SOTA BLEU scores.

AI Reasoning Systems: PAC and Applied Methods

Learning and logic are distinct and remarkable approaches to prediction. Machine learning has experienced a surge in popularity because it is robust to noise and achieves high performance; however, ML experiences many issues with knowledge transfer and extrapolation. In contrast, logic is easily intepreted, and logical rules are easy to chain and transfer between systems; however, inductive logic is brittle to noise. We then explore the premise of combining learning with inductive logic into AI Reasoning Systems. Specifically, we summarize findings from PAC learning (conceptual graphs, robust logics, knowledge infusion) and deep learning (DSRL, $\partial$ILP, DeepLogic) by reproducing proofs of tractability, presenting algorithms in pseudocode, highlighting results, and synthesizing between fields. We conclude with suggestions for integrated models by combining the modules listed above and with a list of unsolved (likely intractable) problems.