Fabio Anselmi, Joel Z. Leibo, Lorenzo Rosasco, Jim Mutch, Andrea Tacchetti, Tomaso Poggio
The present phase of Machine Learning is characterized by supervised learning algorithms relying on large sets of labeled examples ($n \to \infty$). The next phase is likely to focus on algorithms capable of learning from very few labeled examples ($n \to 1$), like humans seem able to do. We propose an approach to this problem and describe the underlying theory, based on the unsupervised, automatic learning of a ``good'' representation for supervised learning, characterized by small sample complexity ($n$). We consider the case of visual object recognition though the theory applies to other domains. The starting point is the conjecture, proved in specific cases, that image representations which are invariant to translations, scaling and other transformations can considerably reduce the sample complexity of learning. We prove that an invariant and unique (discriminative) signature can be computed for each image patch, $I$, in terms of empirical distributions of the dot-products between $I$ and a set of templates stored during unsupervised learning. A module performing filtering and pooling, like the simple and complex cells described by Hubel and Wiesel, can compute such estimates. Hierarchical architectures consisting of this basic Hubel-Wiesel moduli inherit its properties of invariance, stability, and discriminability while capturing the compositional organization of the visual world in terms of wholes and parts. The theory extends existing deep learning convolutional architectures for image and speech recognition. It also suggests that the main computational goal of the ventral stream of visual cortex is to provide a hierarchical representation of new objects/images which is invariant to transformations, stable, and discriminative for recognition---and that this representation may be continuously learned in an unsupervised way during development and visual experience.
Tomaso Poggio, Jim Mutch, Leyla Isik
We develop a sampling extension of M-theory focused on invariance to scale and translation. Quite surprisingly, the theory predicts an architecture of early vision with increasing receptive field sizes and a high resolution fovea -- in agreement with data about the cortical magnification factor, V1 and the retina. From the slope of the inverse of the magnification factor, M-theory predicts a cortical "fovea" in V1 in the order of $40$ by $40$ basic units at each receptive field size -- corresponding to a foveola of size around $26$ minutes of arc at the highest resolution, $\approx 6$ degrees at the lowest resolution. It also predicts uniform scale invariance over a fixed range of scales independently of eccentricity, while translation invariance should depend linearly on spatial frequency. Bouma's law of crowding follows in the theory as an effect of cortical area-by-cortical area pooling; the Bouma constant is the value expected if the signature responsible for recognition in the crowding experiments originates in V2. From a broader perspective, the emerging picture suggests that visual recognition under natural conditions takes place by composing information from a set of fixations, with each fixation providing recognition from a space-scale image fragment -- that is an image patch represented at a set of increasing sizes and decreasing resolutions.
Hrushikesh Mhaskar, Tomaso Poggio
This paper is motivated by an open problem around deep networks, namely, the apparent absence of over-fitting despite large over-parametrization which allows perfect fitting of the training data. In this paper, we analyze this phenomenon in the case of regression problems when each unit evaluates a periodic activation function. We argue that the minimal expected value of the square loss is inappropriate to measure the generalization error in approximation of compositional functions in order to take full advantage of the compositional structure. Instead, we measure the generalization error in the sense of maximum loss, and sometimes, as a pointwise error. We give estimates on exactly how many parameters ensure both zero training error as well as a good generalization error. We prove that a solution of a regularization problem is guaranteed to yield a good training error as well as a good generalization error and estimate how much error to expect at which test data.
Tomaso Poggio, Kenji Kawaguchi, Qianli Liao, Brando Miranda, Lorenzo Rosasco, Xavier Boix, Jack Hidary, Hrushikesh Mhaskar
A main puzzle of deep networks revolves around the absence of overfitting despite large overparametrization and despite the large capacity demonstrated by zero training error on randomly labeled data. In this note, we show that the dynamics associated to gradient descent minimization of nonlinear networks is topologically equivalent, near the asymptotically stable minima of the empirical error, to linear gradient system in a quadratic potential with a degenerate (for square loss) or almost degenerate (for logistic or crossentropy loss) Hessian. The proposition depends on the qualitative theory of dynamical systems and is supported by numerical results. Our main propositions extend to deep nonlinear networks two properties of gradient descent for linear networks, that have been recently established (1) to be key to their generalization properties: 1. Gradient descent enforces a form of implicit regularization controlled by the number of iterations, and asymptotically converges to the minimum norm solution for appropriate initial conditions of gradient descent. This implies that there is usually an optimum early stopping that avoids overfitting of the loss. This property, valid for the square loss and many other loss functions, is relevant especially for regression. 2. For classification, the asymptotic convergence to the minimum norm solution implies convergence to the maximum margin solution which guarantees good classification error for "low noise" datasets. This property holds for loss functions such as the logistic and cross-entropy loss independently of the initial conditions. The robustness to overparametrization has suggestive implications for the robustness of the architecture of deep convolutional networks with respect to the curse of dimensionality.
Tomaso Poggio, Gil Kur, Andrzej Banburski
In solving a system of $n$ linear equations in $d$ variables $Ax=b$, the condition number of the $n,d$ matrix $A$ measures how much errors in the data $b$ affect the solution $x$. Estimates of this type are important in many inverse problems. An example is machine learning where the key task is to estimate an underlying function from a set of measurements at random points in a high dimensional space and where low sensitivity to error in the data is a requirement for good predictive performance. Here we discuss the simple observation, which is known but surprisingly little quoted (see Theorem 4.2 in \cite{Brgisser:2013:CGN:2526261}): when the columns of $A$ are random vectors, the condition number of $A$ is highest if $d=n$, that is when the inverse of $A$ exists. An overdetermined system ($n>d$) as well as an underdetermined system ($n<d$), for which the pseudoinverse must be used instead of the inverse, typically have significantly better, that is lower, condition numbers. Thus the condition number of $A$ plotted as function of $d$ shows a double descent behavior with a peak at $d=n$.
Tomaso Poggio
We show that \emph{efficient Turing computability} at any fixed input/output precision implies the existence of \emph{compositionally sparse} (bounded-fan-in, polynomial-size) DAG representations and of corresponding neural approximants achieving the target precision. Concretely: if $f:[0,1]^d\to\R^m$ is computable in time polynomial in the bit-depths, then for every pair of precisions $(n,m_{\mathrm{out}})$ there exists a bounded-fan-in Boolean circuit of size and depth $\poly(n+m_{\mathrm{out}})$ computing the discretized map; replacing each gate by a constant-size neural emulator yields a deep network of size/depth $\poly(n+m_{\mathrm{out}})$ that achieves accuracy $\varepsilon=2^{-m_{\mathrm{out}}}$. We also relate these constructions to compositional approximation rates \cite{MhaskarPoggio2016b,poggio_deep_shallow_2017,Poggio2017,Poggio2023HowDS} and to optimization viewed as hierarchical search over sparse structures.
Hrushikesh Mhaskar, Tomaso Poggio
The paper briefy reviews several recent results on hierarchical architectures for learning from examples, that may formally explain the conditions under which Deep Convolutional Neural Networks perform much better in function approximation problems than shallow, one-hidden layer architectures. The paper announces new results for a non-smooth activation function - the ReLU function - used in present-day neural networks, as well as for the Gaussian networks. We propose a new definition of relative dimension to encapsulate different notions of sparsity of a function class that can possibly be exploited by deep networks but not by shallow ones to drastically reduce the complexity required for approximation and learning.
Tomaso Poggio, Qianli Liao, Brando Miranda, Andrzej Banburski, Xavier Boix, Jack Hidary
A main puzzle of deep neural networks (DNNs) revolves around the apparent absence of "overfitting", defined in this paper as follows: the expected error does not get worse when increasing the number of neurons or of iterations of gradient descent. This is surprising because of the large capacity demonstrated by DNNs to fit randomly labeled data and the absence of explicit regularization. Recent results by Srebro et al. provide a satisfying solution of the puzzle for linear networks used in binary classification. They prove that minimization of loss functions such as the logistic, the cross-entropy and the exp-loss yields asymptotic, "slow" convergence to the maximum margin solution for linearly separable datasets, independently of the initial conditions. Here we prove a similar result for nonlinear multilayer DNNs near zero minima of the empirical loss. The result holds for exponential-type losses but not for the square loss. In particular, we prove that the weight matrix at each layer of a deep network converges to a minimum norm solution up to a scale factor (in the separable case). Our analysis of the dynamical system corresponding to gradient descent of a multilayer network suggests a simple criterion for ranking the generalization performance of different zero minimizers of the empirical loss.
Chiyuan Zhang, Qianli Liao, Alexander Rakhlin, Brando Miranda, Noah Golowich, Tomaso Poggio
In Theory IIb we characterize with a mix of theory and experiments the optimization of deep convolutional networks by Stochastic Gradient Descent. The main new result in this paper is theoretical and experimental evidence for the following conjecture about SGD: SGD concentrates in probability -- like the classical Langevin equation -- on large volume, "flat" minima, selecting flat minimizers which are with very high probability also global minimizers
Akshay Rangamani, Lorenzo Rosasco, Tomaso Poggio
We study the average $\mbox{CV}_{loo}$ stability of kernel ridge-less regression and derive corresponding risk bounds. We show that the interpolating solution with minimum norm minimizes a bound on $\mbox{CV}_{loo}$ stability, which in turn is controlled by the condition number of the empirical kernel matrix. The latter can be characterized in the asymptotic regime where both the dimension and cardinality of the data go to infinity. Under the assumption of random kernel matrices, the corresponding test error should be expected to follow a double descent curve.
Tomaso Poggio, Qianli Liao
Deep ReLU networks trained with the square loss have been observed to perform well in classification tasks. We provide here a theoretical justification based on analysis of the associated gradient flow. We show that convergence to a solution with the absolute minimum norm is expected when normalization techniques such as Batch Normalization (BN) or Weight Normalization (WN) are used together with Weight Decay (WD). The main property of the minimizers that bounds their expected error is the norm: we prove that among all the close-to-interpolating solutions, the ones associated with smaller Frobenius norms of the unnormalized weight matrices have better margin and better bounds on the expected classification error. With BN but in the absence of WD, the dynamical system is singular. Implicit dynamical regularization -- that is zero-initial conditions biasing the dynamics towards high margin solutions -- is also possible in the no-BN and no-WD case. The theory yields several predictions, including the role of BN and weight decay, aspects of Papyan, Han and Donoho's Neural Collapse and the constraints induced by BN on the network weights.
Andrzej Banburski, Qianli Liao, Brando Miranda, Lorenzo Rosasco, Fernanda De La Torre, Jack Hidary, Tomaso Poggio
The key to generalization is controlling the complexity of the network. However, there is no obvious control of complexity -- such as an explicit regularization term -- in the training of deep networks for classification. We will show that a classical form of norm control -- but kind of hidden -- is present in deep networks trained with gradient descent techniques on exponential-type losses. In particular, gradient descent induces a dynamics of the normalized weights which converge for $t \to \infty$ to an equilibrium which corresponds to a minimum norm (or maximum margin) solution. For sufficiently large but finite $ρ$ -- and thus finite $t$ -- the dynamics converges to one of several margin maximizers, with the margin monotonically increasing towards a limit stationary point of the flow. In the usual case of stochastic gradient descent, most of the stationary points are likely to be convex minima corresponding to a constrained minimizer -- the network with normalized weights-- which corresponds to vanishing regularization. The solution has zero generalization gap, for fixed architecture, asymptotically for $N \to \infty$, where $N$ is the number of training examples. Our approach extends some of the original results of Srebro from linear networks to deep networks and provides a new perspective on the implicit bias of gradient descent. We believe that the elusive complexity control we describe is responsible for the puzzling empirical finding of good predictive performance by deep networks, despite overparametrization.
Tomaso Poggio, Stephen Voinea, Lorenzo Rosasco
In batch learning, stability together with existence and uniqueness of the solution corresponds to well-posedness of Empirical Risk Minimization (ERM) methods; recently, it was proved that CV_loo stability is necessary and sufficient for generalization and consistency of ERM. In this note, we introduce CV_on stability, which plays a similar note in online learning. We show that stochastic gradient descent (SDG) with the usual hypotheses is CVon stable and we then discuss the implications of CV_on stability for convergence of SGD.
Guillermo D. Canas, Tomaso Poggio, Lorenzo Rosasco
We study the problem of estimating a manifold from random samples. In particular, we consider piecewise constant and piecewise linear estimators induced by k-means and k-flats, and analyze their performance. We extend previous results for k-means in two separate directions. First, we provide new results for k-means reconstruction on manifolds and, secondly, we prove reconstruction bounds for higher-order approximation (k-flats), for which no known results were previously available. While the results for k-means are novel, some of the technical tools are well-established in the literature. In the case of k-flats, both the results and the mathematical tools are new.
Anna Volokitin, Gemma Roig, Tomaso Poggio
Crowding is a visual effect suffered by humans, in which an object that can be recognized in isolation can no longer be recognized when other objects, called flankers, are placed close to it. In this work, we study the effect of crowding in artificial Deep Neural Networks for object recognition. We analyze both standard deep convolutional neural networks (DCNNs) as well as a new version of DCNNs which is 1) multi-scale and 2) with size of the convolution filters change depending on the eccentricity wrt to the center of fixation. Such networks, that we call eccentricity-dependent, are a computational model of the feedforward path of the primate visual cortex. Our results reveal that the eccentricity-dependent model, trained on target objects in isolation, can recognize such targets in the presence of flankers, if the targets are near the center of the image, whereas DCNNs cannot. Also, for all tested networks, when trained on targets in isolation, we find that recognition accuracy of the networks decreases the closer the flankers are to the target and the more flankers there are. We find that visual similarity between the target and flankers also plays a role and that pooling in early layers of the network leads to more crowding. Additionally, we show that incorporating the flankers into the images of the training set does not improve performance with crowding.
Olivier Morère, Jie Lin, Antoine Veillard, Vijay Chandrasekhar, Tomaso Poggio
The goal of this work is the computation of very compact binary hashes for image instance retrieval. Our approach has two novel contributions. The first one is Nested Invariance Pooling (NIP), a method inspired from i-theory, a mathematical theory for computing group invariant transformations with feed-forward neural networks. NIP is able to produce compact and well-performing descriptors with visual representations extracted from convolutional neural networks. We specifically incorporate scale, translation and rotation invariances but the scheme can be extended to any arbitrary sets of transformations. We also show that using moments of increasing order throughout nesting is important. The NIP descriptors are then hashed to the target code size (32-256 bits) with a Restricted Boltzmann Machine with a novel batch-level regularization scheme specifically designed for the purpose of hashing (RBMH). A thorough empirical evaluation with state-of-the-art shows that the results obtained both with the NIP descriptors and the NIP+RBMH hashes are consistently outstanding across a wide range of datasets.
Andrea Tacchetti, Leyla Isik, Tomaso Poggio
Jun 15, 2016·q-bio.NC·PDF Recognizing the actions of others from visual stimuli is a crucial aspect of human visual perception that allows individuals to respond to social cues. Humans are able to identify similar behaviors and discriminate between distinct actions despite transformations, like changes in viewpoint or actor, that substantially alter the visual appearance of a scene. This ability to generalize across complex transformations is a hallmark of human visual intelligence. Advances in understanding motion perception at the neural level have not always translated in precise accounts of the computational principles underlying what representation our visual cortex evolved or learned to compute. Here we test the hypothesis that invariant action discrimination might fill this gap. Recently, the study of artificial systems for static object perception has produced models, CNNs, that achieve human level performance in complex discriminative tasks. Within this class of models, architectures that better support invariant object recognition also produce image representations that match those implied by human and primate neural data. However, whether these models produce representations of action sequences that support recognition across complex transformations and closely follow neural representations remains unknown. Here we show that spatiotemporal CNNs appropriately categorize video stimuli into actions, and that deliberate model modifications that improve performance on an invariant action recognition task lead to data representations that better match human neural recordings. Our results support our hypothesis that performance on invariant discrimination dictates the neural representations of actions computed by human visual cortex.
Liu Ziyin, Isaac Chuang, Tomer Galanti, Tomaso Poggio
Understanding neural representations will help open the black box of neural networks and advance our scientific understanding of modern AI systems. However, how complex, structured, and transferable representations emerge in modern neural networks has remained a mystery. Building on previous results, we propose the Canonical Representation Hypothesis (CRH), which posits a set of six alignment relations to universally govern the formation of representations in most hidden layers of a neural network. Under the CRH, the latent representations (R), weights (W), and neuron gradients (G) become mutually aligned during training. This alignment implies that neural networks naturally learn compact representations, where neurons and weights are invariant to task-irrelevant transformations. We then show that the breaking of CRH leads to the emergence of reciprocal power-law relations between R, W, and G, which we refer to as the Polynomial Alignment Hypothesis (PAH). We present a minimal-assumption theory proving that the balance between gradient noise and regularization is crucial for the emergence of the canonical representation. The CRH and PAH lead to an exciting possibility of unifying major key deep learning phenomena, including neural collapse and the neural feature ansatz, in a single framework.
Silvia Villa, Lorenzo Rosasco, Tomaso Poggio
We consider the fundamental question of learnability of a hypotheses class in the supervised learning setting and in the general learning setting introduced by Vladimir Vapnik. We survey classic results characterizing learnability in term of suitable notions of complexity, as well as more recent results that establish the connection between learnability and stability of a learning algorithm.
Youssef Mroueh, Tomaso Poggio, Lorenzo Rosasco, Jean-Jacques Slotine
In this paper we discuss a novel framework for multiclass learning, defined by a suitable coding/decoding strategy, namely the simplex coding, that allows to generalize to multiple classes a relaxation approach commonly used in binary classification. In this framework, a relaxation error analysis can be developed avoiding constraints on the considered hypotheses class. Moreover, we show that in this setting it is possible to derive the first provably consistent regularized method with training/tuning complexity which is independent to the number of classes. Tools from convex analysis are introduced that can be used beyond the scope of this paper.