The hippocampus plays a critical role in learning, memory, and spatial representation, processes that depend on the NMDA receptor (NMDAR).

The stability of language model pre-training and its effects on downstream performance are still understudied. Prior work shows that the training process can yield significantly different results in response to slight variations in initial conditions, e.g., the random seed. Crucially, the research community still lacks sufficient resources and tools to systematically investigate pre-training stability, particularly for decoder-only language models. We introduce the PolyPythias, a set of 45 new training runs for the Pythia model suite: 9 new seeds across 5 model sizes, from 14M to 410M parameters, resulting in about 7k new checkpoints that we release. Using these new 45 training runs, in addition to the 5 already available, we study the effects of different initial conditions determined by the seed -- i.e., parameters' initialisation and data order -- on (i) downstream performance, (ii) learned linguistic representations, and (iii) emergence of training phases. In addition to common scaling behaviours, our analyses generally reveal highly consistent training dynamics across both model sizes and initial conditions. Further, the new seeds for each model allow us to identify outlier training runs and delineate their characteristics. Our findings show the potential of using these methods to predict training stability.

The impact of randomness on model training is poorly understood. How do differences in data order and initialization actually manifest in the model, such that some training runs outperform others or converge faster? Furthermore, how can we interpret the resulting training dynamics and the phase transitions that characterize different trajectories? To understand the effect of randomness on the dynamics and outcomes of neural network training, we train models multiple times with different random seeds and compute a variety of metrics throughout training, such as the $L_2$ norm, mean, and variance of the neural network's weights. We then fit a hidden Markov model (HMM) over the resulting sequences of metrics. The HMM represents training as a stochastic process of transitions between latent states, providing an intuitive overview of significant changes during training. Using our method, we produce a low-dimensional, discrete representation of training dynamics on grokking tasks, image classification, and masked language modeling. We use the HMM representation to study phase transitions and identify latent "detour" states that slow down convergence.

Robots can now learn how to make decisions and control themselves, generalizing learned behaviors to unseen scenarios. In particular, AI powered robots show promise in rough environments like the lunar surface, due to the environmental uncertainties. We address this critical generalization aspect for robot locomotion in rough terrain through a training algorithm we have created called the Path Planning and Motion Control (PPMC) Training Algorithm. This algorithm is coupled with any generic reinforcement learning algorithm to teach robots how to respond to user commands and to travel to designated locations on a single neural network. In this paper, we show that the algorithm works independent of the robot structure, demonstrating that it works on a wheeled rover in addition the past results on a quadruped walking robot. Further, we take several big steps towards real world practicality by introducing a rough highly uneven terrain. Critically, we show through experiments that the robot learns to generalize to new rough terrain maps, retaining a 100% success rate. To the best of our knowledge, this is the first paper to introduce a generic training algorithm teaching generalized PPMC in rough environments to any robot, with just the use of reinforcement learning.

In this paper we describe a hierarchical method for procedurally generating 2D game maps using multi-dimensional Markov chains (MdMCs). Our method takes a collection of 2D game maps, breaks them into small chunks and performs clustering to find a set of chunks that correspond to high-level structures (high-level tiles) in the training maps. This set of high-level tiles is then used to re-represent the training maps, and to fit two sets of MdMC models: a high-level model captures the distribution of high-level tiles in the map, and a set of low-level models capture the internal structure of each high-level tile. These two sets of models can then be used to hierarchically generate new maps. We test our approach using two classic games, Super Mario Bros. and Loderunner, and compare the results against other existing map generators.