arXiv subjects

Alberto Bernacchia

Publications and source records attributed to Alberto Bernacchia.

27 records · Page 2Linked to original sources

arXiv2021

Meta-learning with negative learning rates

Deep learning models require a large amount of data to perform well. When data is scarce for a target task, we can transfer the knowledge gained by training on similar tasks to quickly learn the target. A successful approach is meta-learning, or "learning to learn" a distribution of tasks, where "learning" is represented by an outer loop, and "to learn" by an inner loop of gradient descent. However, a number of recent empirical studies argue that the inner loop is unnecessary and more simple models work equally well or even better. We study the performance of MAML as a function of the learning rate of the inner loop, where zero learning rate implies that there is no inner loop. Using random matrix theory and exact solutions of linear models, we calculate an algebraic expression for the test loss of MAML applied to mixed linear regression and nonlinear regression with overparameterized models. Surprisingly, while the optimal learning rate for adaptation is positive, we find that the optimal learning rate for training is always negative, a setting that has never been considered before. Therefore, not only does the performance increase by decreasing the learning rate to zero, as suggested by recent work, but it can be increased even further by decreasing the learning rate to negative values. These results help clarify under what circumstances meta-learning performs best.

cs.LG↗

arXiv2021

Meta-Learning with MAML on Trees

In meta-learning, the knowledge learned from previous tasks is transferred to new ones, but this transfer only works if tasks are related. Sharing information between unrelated tasks might hurt performance, and it is unclear how to transfer knowledge across tasks with a hierarchical structure. Our research extends a model agnostic meta-learning model, MAML, by exploiting hierarchical task relationships. Our algorithm, TreeMAML, adapts the model to each task with a few gradient steps, but the adaptation follows the hierarchical tree structure: in each step, gradients are pooled across tasks clusters, and subsequent steps follow down the tree. We also implement a clustering algorithm that generates the tasks tree without previous knowledge of the task structure, allowing us to make use of implicit relationships between the tasks. We show that the new algorithm, which we term TreeMAML, performs better than MAML when the task structure is hierarchical for synthetic experiments. To study the performance of the method in real-world data, we apply this method to Natural Language Understanding, we use our algorithm to finetune Language Models taking advantage of the language phylogenetic tree. We show that TreeMAML improves the state of the art results for cross-lingual Natural Language Inference. This result is useful, since most languages in the world are under-resourced and the improvement on cross-lingual transfer allows the internationalization of NLP models. This results open the window to use this algorithm in other real-world hierarchical datasets.

cs.LG↗

arXiv2020

Cyclic orthogonal convolutions for long-range integration of features

In Convolutional Neural Networks (CNNs) information flows across a small neighbourhood of each pixel of an image, preventing long-range integration of features before reaching deep layers in the network. We propose a novel architecture that allows flexible information flow between features $z$ and locations $(x,y)$ across the entire image with a small number of layers. This architecture uses a cycle of three orthogonal convolutions, not only in $(x,y)$ coordinates, but also in $(x,z)$ and $(y,z)$ coordinates. We stack a sequence of such cycles to obtain our deep network, named CycleNet. As this only requires a permutation of the axes of a standard convolution, its performance can be directly compared to a CNN. Our model obtains competitive results at image classification on CIFAR-10 and ImageNet datasets, when compared to CNNs of similar size. We hypothesise that long-range integration favours recognition of objects by shape rather than texture, and we show that CycleNet transfers better than CNNs to stylised images. On the Pathfinder challenge, where integration of distant features is crucial, CycleNet outperforms CNNs by a large margin. We also show that even when employing a small convolutional kernel, the size of receptive fields of CycleNet reaches its maximum after one cycle, while conventional CNNs require a large number of layers.

cs.CV↗

arXiv2012

Decorrelation by recurrent inhibition in heterogeneous neural circuits

The activity of neurons is correlated, and this correlation affects how the brain processes information. We study the neural circuit mechanisms of correlations by analyzing a network model characterized by strong and heterogeneous interactions: excitatory input drives the fluctuations of neural activity, which are counterbalanced by inhibitory feedback. In particular, excitatory input tends to correlate neurons, while inhibitory feedback reduces correlations. We demonstrate that heterogeneity of synaptic connections is necessary for this inhibition of correlations. We calculate statistical averages over the disordered synaptic interactions, and we apply our findings to both a simple linear model and to a more realistic spiking network model. We find that correlations at zero time-lag are positive and of magnitude K^{-1/2}, where K is the number of connections to a neuron. Correlations at longer timescales are of smaller magnitude, of order K^{-1}, implying that inhibition of correlations occurs quickly, on a timescale of K^{-1/2}. The small magnitude of correlations agrees qualitatively with physiological measurements in the Cerebral Cortex and Basal Ganglia. The model could be used to study correlations in brain regions dominated by recurrent inhibition, such as the Striatum and Globus Pallidus.

q-bio.NC↗

arXiv2012

On the equivalence of Hopfield Networks and Boltzmann Machines

A specific type of neural network, the Restricted Boltzmann Machine (RBM), is implemented for classification and feature detection in machine learning. RBM is characterized by separate layers of visible and hidden units, which are able to learn efficiently a generative model of the observed data. We study a "hybrid" version of RBM's, in which hidden units are analog and visible units are binary, and we show that thermodynamics of visible units are equivalent to those of a Hopfield network, in which the N visible units are the neurons and the P hidden units are the learned patterns. We apply the method of stochastic stability to derive the thermodynamics of the model, by considering a formal extension of this technique to the case of multiple sets of stored patterns, which may act as a benchmark for the study of correlated sets. Our results imply that simulating the dynamics of a Hopfield network, requiring the update of N neurons and the storage of N(N-1)/2 synapses, can be accomplished by a hybrid Boltzmann Machine, requiring the update of N+P neurons but the storage of only NP synapses. In addition, the well known glass transition of the Hopfield network has a counterpart in the Boltzmann Machine: It corresponds to an optimum criterion for selecting the relative sizes of the hidden and visible layers, resolving the trade-off between flexibility and generality of the model. The low storage phase of the Hopfield model corresponds to few hidden units and hence a overly constrained RBM, while the spin-glass phase (too many hidden units) corresponds to unconstrained RBM prone to overfitting of the observed data.

cond-mat.dis-nn↗

arXiv2010

Self-consistent method for density estimation

The estimation of a density profile from experimental data points is a challenging problem, usually tackled by plotting a histogram. Prior assumptions on the nature of the density, from its smoothness to the specification of its form, allow the design of more accurate estimation procedures, such as Maximum Likelihood. Our aim is to construct a procedure that makes no explicit assumptions, but still providing an accurate estimate of the density. We introduce the self-consistent estimate: the power spectrum of a candidate density is given, and an estimation procedure is constructed on the assumption, to be released \emph{a posteriori}, that the candidate is correct. The self-consistent estimate is defined as a prior candidate density that precisely reproduces itself. Our main result is to derive the exact expression of the self-consistent estimate for any given dataset, and to study its properties. Applications of the method require neither priors on the form of the density nor the subjective choice of parameters. A cutoff frequency, akin to a bin size or a kernel bandwidth, emerges naturally from the derivation. We apply the self-consistent estimate to artificial data generated from various distributions and show that it reaches the theoretical limit for the scaling of the square error with the dataset size.

stat.ME↗

arXiv2007

Continuous or discrete attractors in neural circuits? A self-organized switch at maximal entropy

A recent experiment suggests that neural circuits may alternatively implement continuous or discrete attractors, depending on the training set up. In recurrent neural network models, continuous and discrete attractors are separately modeled by distinct forms of synaptic prescriptions (learning rules). Here, we report a solvable network model, endowed with Hebbian synaptic plasticity, which is able to learn either discrete or continuous attractors, depending on the frequency of presentation of stimuli and on the structure of sensory coding. A continuous attractor is learned when experience matches sensory coding, i.e. when the distribution of experienced stimuli matches the distribution of preferred stimuli of neurons. In that case, there is no processing of sensory information and neural activity displays maximal entropy. If experience goes beyond sensory coding, processing is initiated and the continuous attractor is destabilized into a set of discrete attractors.

physics.bio-ph↗

arXiv2007

Detecting spatial patterns with the cumulant function. Part II: An application to El Nino

The spatial coherence of a measured variable (e.g. temperature or pressure) is often studied to determine the regions where this variable varies the most or to find teleconnections, i.e. correlations between specific regions. While usual methods to find spatial patterns, such as Principal Components Analysis (PCA), are constrained by linear symmetries, the dependence of variables such as temperature or pressure at different locations is generally nonlinear. In particular, large deviations from the sample mean are expected to be strongly affected by such nonlinearities. Here we apply a newly developed nonlinear technique (Maxima of Cumulant Function, MCF) for the detection of typical spatial patterns that largely deviate from the mean. In order to test the technique and to introduce the methodology, we focus on the El Nino/Southern Oscillation and its spatial patterns. We find nonsymmetric temperature patterns corresponding to El Nino and La Nina, and we compare the results of MCF with other techniques, such as the symmetric solutions of PCA, and the nonsymmetric solutions of Nonlinear PCA (NLPCA). We found that MCF solutions are more reliable than the NLPCA fits, and can capture mixtures of principal components. Finally, we apply Extreme Value Theory on the temporal variations extracted from our methodology. We find that the tails of the distribution of extreme temperatures during La Nina episodes is bounded, while the tail during El Ninos is less likely to be bounded. This implies that the mean spatial patterns of the two phases are asymmetric, as well as the behaviour of their extremes.

physics.data-an↗

arXiv2007

Detecting spatial patterns with the cumulant function. Part I: The theory

In climate studies, detecting spatial patterns that largely deviate from the sample mean still remains a statistical challenge. Although a Principal Component Analysis (PCA), or equivalently a Empirical Orthogonal Functions (EOF) decomposition, is often applied on this purpose, it can only provide meaningful results if the underlying multivariate distribution is Gaussian. Indeed, PCA is based on optimizing second order moments quantities and the covariance matrix can only capture the full dependence structure for multivariate Gaussian vectors. Whenever the application at hand can not satisfy this normality hypothesis (e.g. precipitation data), alternatives and/or improvements to PCA have to be developed and studied. To go beyond this second order statistics constraint that limits the applicability of the PCA, we take advantage of the cumulant function that can produce higher order moments information. This cumulant function, well-known in the statistical literature, allows us to propose a new, simple and fast procedure to identify spatial patterns for non-Gaussian data. Our algorithm consists in maximizing the cumulant function. To illustrate our approach, its implementation for which explicit computations are obtained is performed on three family of of multivariate random vectors. In addition, we show that our algorithm corresponds to selecting the directions along which projected data display the largest spread over the marginal probability density tails.

math.ST↗