Pascal Tikeng Notsawo

Pascal Jr. Tikeng Notsawo, Hattie Zhou, Mohammad Pezeshki et al. · mila

This paper focuses on predicting the occurrence of grokking in neural networks, a phenomenon in which perfect generalization emerges long after signs of overfitting or memorization are observed. It has been reported that grokking can only be observed with certain hyper-parameters. This makes it critical to identify the parameters that lead to grokking. However, since grokking occurs after a large number of epochs, searching for the hyper-parameters that lead to it is time-consuming. In this paper, we propose a low-cost method to predict grokking without training for a large number of epochs. In essence, by studying the learning curve of the first few epochs, we show that one can predict whether grokking will occur later on. Specifically, if certain oscillations occur in the early epochs, one can expect grokking to occur if the model is trained for a much longer period of time. We propose using the spectral signature of a learning curve derived by applying the Fourier transform to quantify the amplitude of low-frequency components to detect the presence of such oscillations. We also present additional experiments aimed at explaining the cause of these oscillations and characterizing the loss landscape.

Pascal Jr Tikeng Notsawo, Guillaume Dumas, Guillaume Rabusseau · mila

This paper investigates the grokking phenomenon, which refers to the sudden transition from a long memorization to generalization observed during neural networks training, in the context of learning multiplication in finite-dimensional algebras (FDA). While prior work on grokking has focused mainly on group operations, we extend the analysis to more general algebraic structures, including non-associative, non-commutative, and non-unital algebras. We show that learning group operations is a special case of learning FDA, and that learning multiplication in FDA amounts to learning a bilinear product specified by the algebra's structure tensor. For algebras over the reals, we connect the learning problem to matrix factorization with an implicit low-rank bias, and for algebras over finite fields, we show that grokking emerges naturally as models must learn discrete representations of algebraic elements. This leads us to experimentally investigate the following core questions: (i) how do algebraic properties such as commutativity, associativity, and unitality influence both the emergence and timing of grokking, (ii) how structural properties of the structure tensor of the FDA, such as sparsity and rank, influence generalization, and (iii) to what extent generalization correlates with the model learning latent embeddings aligned with the algebra's representation. Our work provides a unified framework for grokking across algebraic structures and new insights into how mathematical structure governs neural network generalization dynamics.

Pascal Junior Tikeng Notsawo · mila

Despite the recent growth of theoretical studies and empirical successes of neural networks, gradient backpropagation is still the most widely used algorithm for training such networks. On the one hand, we have deterministic or full gradient (FG) approaches that have a cost proportional to the amount of training data used but have a linear convergence rate, and on the other hand, stochastic gradient (SG) methods that have a cost independent of the size of the dataset, but have a less optimal convergence rate than the determinist approaches. To combine the cost of the stochastic approach with the convergence rate of the deterministic approach, a stochastic average gradient (SAG) has been proposed. SAG is a method for optimizing the sum of a finite number of smooth convex functions. Like SG methods, the SAG method's iteration cost is independent of the number of terms in the sum. In this work, we propose to compare SAG to some standard optimizers used in machine learning. SAG converges faster than other optimizers on simple toy problems and performs better than many other optimizers on simple machine learning problems. We also propose a combination of SAG with the momentum algorithm and Adam. These combinations allow empirically higher speed and obtain better performance than the other methods, especially when the landscape of the function to optimize presents obstacles or is ill-conditioned.

Tommaso Tosato, Pascal Jr Tikeng Notsawo, Saskia Helbling et al.

Language Models (LMs) have achieved impressive performance on various linguistic tasks, but their relationship to human language processing in the brain remains unclear. This paper examines the gaps and overlaps between LMs and the brain at different levels of analysis, emphasizing the importance of looking beyond input-output behavior to examine and compare the internal processes of these systems. We discuss how insights from neuroscience, such as sparsity, modularity, internal states, and interactive learning, can inform the development of more biologically plausible language models. Furthermore, we explore the role of scaling laws in bridging the gap between LMs and human cognition, highlighting the need for efficiency constraints analogous to those in biological systems. By developing LMs that more closely mimic brain function, we aim to advance both artificial intelligence and our understanding of human cognition.