Elan Barenholtz

Elan Barenholtz

Human language comprehension unfolds sequentially: each word is processed in the context of those that came before, and the interpretation builds incrementally over time. Surprisal, the negative log probability of a word given its context, has been the dominant predictor of incremental processing cost. But surprisal reduces rich sequential representations to a single scalar at each word, discarding information about the direction in which the interpretation has been evolving. Dynamical-systems approaches suggest that the trajectory of the evolving interpretive state, not just its position at each moment,should shape processing, and language itself may have local momentum, since speakers plan utterances a few words at a time. We introduce trajectory extrapolation error: at each word, we fit a linear trajectory to the preceding hidden states of a transformer language model and measure deviation from the extrapolated path. On the Natural Stories corpus, this measure is nearly orthogonal to surprisal (r = .044) and independently predicts self-paced reading times. The effect is especially pronounced in garden-path sentences, strengthens with model scale (GPT-2 Small to Large), and replicates across architectures with different positional encoding schemes (GPT-2 vs. Pythia/RoPE). A displacement control shows the effect is not reducible to representational change magnitude: displacement and extrapolation error predict in opposite directions. These findings reveal two dissociable components of processing cost: word-level prediction error (surprisal) and sensitivity to the local momentum of the unfolding interpretation (trajectory extrapolation error).

Elan Barenholtz

Recent work interprets the linear recoverability of geographic and temporal variables from large language model (LLM) hidden states as evidence for world-like internal representations. We test a simpler possibility: that much of the relevant structure is already latent in text itself. Applying the same class of ridge regression probes to static co-occurrence-based embeddings (GloVe and Word2Vec), we find substantial recoverable geographic signal and weaker but reliable temporal signal, with held-out R^2 values of 0.71-0.87 for city coordinates and 0.48-0.52 for historical birth years. Semantic-neighbor analyses and targeted subspace ablations show that these signals depend strongly on interpretable lexical gradients, especially country names and climate-related vocabulary. These findings suggest that ordinary word co-occurrence preserves richer spatial, temporal, and environmental structure than is often assumed, revealing a remarkable and underappreciated capacity of simple static embeddings to preserve world-shaped structure from text alone. Linear probe recoverability alone therefore does not establish a representational move beyond text.

Rachel A. StClair, William Edward Hahn, Elan Barenholtz

One goal of general intelligence is to learn novel information without overwriting prior learning. The utility of learning without forgetting (CF) is twofold: first, the system can return to previously learned tasks after learning something new. In addition, bootstrapping previous knowledge may allow for faster learning of a novel task. Previous approaches to CF and bootstrapping are primarily based on modifying learning in the form of changing weights to tune the model to the current task, overwriting previously tuned weights from previous tasks.However, another critical factor that has been largely overlooked is the initial network topology, or architecture. Here, we argue that the topology of biological brains likely evolved certain features that are designed to achieve this kind of informational conservation. In particular, we consider that the highly conserved property of modularity may offer a solution to weight-update learning methods that adheres to the learning without catastrophic forgetting and bootstrapping constraints. Final considerations are then made on how to combine these two learning objectives in a dynamical, general learning system.

Michael Teti, William Edward Hahn, Shawn Martin et al.

Self-driving technology is advancing rapidly --- albeit with significant challenges and limitations. This progress is largely due to recent developments in deep learning algorithms. To date, however, there has been no systematic comparison of how different deep learning architectures perform at such tasks, or an attempt to determine a correlation between classification performance and performance in an actual vehicle, a potentially critical factor in developing self-driving systems. Here, we introduce the first controlled comparison of multiple deep-learning architectures in an end-to-end autonomous driving task across multiple testing conditions. We compared performance, under identical driving conditions, across seven architectures including a fully-connected network, a simple 2 layer CNN, AlexNet, VGG-16, Inception-V3, ResNet, and an LSTM by assessing the number of laps each model was able to successfully complete without crashing while traversing an indoor racetrack. We compared performance across models when the conditions exactly matched those in training as well as when the local environment and track were configured differently and objects that were not included in the training dataset were placed on the track in various positions. In addition, we considered performance using several different data types for training and testing including single grayscale and color frames, and multiple grayscale frames stacked together in sequence. With the exception of a fully-connected network, all models performed reasonably well (around or above 80\%) and most very well (~95\%) on at least one input type but with considerable variation across models and inputs. Overall, AlexNet, operating on single color frames as input, achieved the best level of performance (100\% success rate in phase one and 55\% in phase two) while VGG-16 performed well most consistently across image types.