ClaimsPeople

Yann LeCun

Yann LeCun is a computer scientist who shared the 2018 ACM Turing Award, is a Jacob T. Schwartz Professor at NYU's Courant Institute, was Meta's Chief AI Scientist and founding director of FAIR, and is Executive Chairman of AMI Labs.

1,175 claims collected, 197 researched, 84 resolved. Report of 4 November 2026, method 0.1.0+c876030c.

Scores by kind of claim

Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.

  • Factual claims

    Statements about the present or the past, checked against the record: did they hold up?

    78 72 83

    High confidence, 79 factual claims resolved. From 16 documents, 2006 to 2026; 19 documents sampled from 33 candidates found.

    Not scored: 43 unresolvable.

  • Predictions

    Statements about events outside the person's control, checked once they fell due: did they come true?

    64 43 83

    Low confidence, 5 predictions resolved. From 4 documents, 2010 to 2024; 19 documents sampled from 33 candidates found.

    Not scored: 66 unresolvable. 3 more fall due later.

    Does not yet allow for how likely each prediction was when made.

  • Promises

    Commitments to the person's or organisation's own future action, with something to deliver: was it delivered?

    Insufficient record: 0 promises resolved, 5 needed for a score.

    Not scored: 1 unresolvable.

All types combined: 77 (likely 71 to 82, high confidence), from 84 claims: 79 factual claims, 5 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.

Yann LeCun's claims

1,175 claims, sorted by newest said. Choose a claim to read it in full.
VerdictKindSaidClaimTopic
Not checkable Not checkable 2006 Non-linear factors allow the modeling of highly complex dependencies between inputs and labels Not checkable · 2006
Not researched Factual 2006 The dependency between X and Y sometimes cannot be expressed as a function mapping X to Y, e.g. the constraint X^2+Y^2=1. Factual · 2006 AI capabilities
Not checkable Not checkable 2006 Probabilistic models have two major disadvantages. Not checkable · 2006
Not checkable Not checkable 2006 In the energybased framework, there is no need to make excuses for breaking the rules. Not checkable · 2006
Not checkable Not checkable 2006 This loss function can be optimized with a variety of techniques. Not checkable · 2006
Not researched Factual 2006 With the NLL loss the minimum is deeper than with the square-square loss because incorrect answers' energies are pushed up to infinity, with decreasing force. Factual · 2006 AI capabilities
Not researched Factual 2006 Inference by marginalization over latent variables reduces to inference by minimization when β→∞ (zero temperature). Factual · 2006 AI capabilities
Resolved true Factual 2006 Each NLL iteration involves considerably more work because pushing up every incorrect answer is expensive when the second term's derivative has no analytical form. Factual · 2006 AI capabilities
Not researched Factual 2006 In 2006, the authors said that, to their knowledge, no systematic experimental study had been published on whether second-order methods bring a significant speed improvement over well-tuned stochastic gradient methods for training these models. Factual · 2006 AI capabilities
Not checkable Not checkable 2006 Energy-based learning can be seen as an alternative to probabilistic estimation for prediction, classification, or decision-making tasks. Not checkable · 2006
Not researched Factual 2006 Because energies are uncalibrated, there is no a priori guarantee that separately trained energy-based models have commensurate energy scales. Factual · 2006 Science and research methods
Not checkable Not checkable 2006 In this way, the regression on X1 and X2 is implicitly learned through the constraint Y rather than explicitly learned through supervision. Not checkable · 2006
Not checkable Not checkable 2006 Many existing learning models can be be expressed simply in the framework of energy-based learning. Not checkable · 2006
Not checkable Not checkable 2006 What we propose here is to view the probabilistic approach as a special case of a much larger family of energy-based methods. Not checkable · 2006
Not researched Factual 2006 One of the alleged advantages of CRFs is that their loss function is convex with respect to the parameters W. Factual · 2006 AI capabilities
Not researched Factual 2006 The main problem with the perceptron loss is the absence of a margin. Factual · 2006 AI capabilities
Not researched Factual 2006 Because of its zero margin, the generalized perceptron loss could lead to a collapsed energy surface. Factual · 2006 AI capabilities
Not researched Factual 2006 The log loss can be seen as a soft version of the hinge loss with an infinite margin. Factual · 2006 Science and research methods
Not checkable Not checkable 2006 Properly designing the architecture and the loss function is critical. Not checkable · 2006
Not checkable Not checkable 2006 Such hidden variables are called latent variables, by analogy with a similar concept in probabilistic modeling. Not checkable · 2006
Not researched Factual 2006 Learning and inference with EBMs involve minimizing the energy over the set of answers Y and latent variables Z. Factual · 2006 AI capabilities
Not researched Factual 2006 When the energy is linear in W, the hinge loss with L2 regularizer takes the simple form of Eq. 81. Factual · 2006 AI capabilities
Not researched Factual 2006 The energy and loss functions must satisfy the stated sufficient conditions to be guaranteed to work in an energy-based setting. Factual · 2006 AI capabilities
Not researched Factual 2006 The NLL loss reduces to the generalized perceptron loss as β→∞, and to the log loss when Y has two elements. Factual · 2006 AI capabilities
Not researched Factual 2006 The square-exponential loss has an infinite margin and pushes incorrect answers' energies to infinity with exponentially decreasing force. Factual · 2006 AI capabilities
Not researched Factual 2006 Siamese architectures are used to learn similarity metrics with labeled examples. Factual · 2006 AI capabilities
Not checkable Not checkable 2006 Graph transformer networks are hierarchical sequence modeling systems in which the objects that are manipulated are trellises containing… Not checkable · 2006
Not checkable Not checkable 2006 For other models, one cannot guarantee anything. Not checkable · 2006
Not researched Factual 2006 Maximizing the conditional likelihood of training samples is equivalent to minimizing the negative log-likelihood loss. Factual · 2006 AI capabilities
Unresolvable Factual 2006 There exists at least one point in HP1 at which the loss is less than at all points in HP2. Factual · 2006 AI capabilities
Not researched Factual 2006 With a linear parameterization, training with the energy loss reduces to a convex least-squares problem. Factual · 2006 AI capabilities
Not checkable Not checkable 2006 Energy-based methods are equally well justified as probabilistic methods. Not checkable · 2006
Not researched Factual 2006 The energy of the most offending incorrect answer is bounded above by the energy of the next most offending incorrect answer. Factual · 2006 AI capabilities
Not checkable Not checkable 2006 The architectures described in the previous section are simple functions of Y with a single minimum within the set Y. Not checkable · 2006
Not checkable Not checkable 2006 This could be construed as a thinly disguised way of removing the normalization constraint. Not checkable · 2006
Not researched Factual 2006 Normalization is only necessary for generative models. Factual · 2006 AI capabilities
Not checkable Not checkable 2006 These models can be easily described in an EBM setting. Not checkable · 2006
Not researched Factual 2006 When the RBF centers are learned rather than fixed, there is no guarantee that d1 + d2 ≥ d. Factual · 2006 AI capabilities
Not researched Factual 2006 Section 5 of this tutorial concluded that the generalized perceptron loss is not a good loss function. Factual · 2006 AI capabilities
Not checkable Not checkable 2006 The EBM approach provides a common theoretical framework for many learning models, including traditional discriminative and generative… Not checkable · 2006
Not researched Factual 2006 Energy-based models have no requirement for proper normalization, so they avoid the problem of evaluating intractable normalizing integrals. Factual · 2006 Science and research methods
Not researched Factual 2006 The cost of shortest-path inference is proportional to the number of edges, which is exponentially smaller than the number of paths in general. Factual · 2006 AI capabilities
Not researched Factual 2006 Any probability distribution over Y can be approximated arbitrarily closely by a Gibbs distribution of the form e^(-βE)/Σe^(-βE). Factual · 2006 AI capabilities
Not researched Factual 2006 The EBM approach eliminates the problem of having to perform a normalization step after each gradient parameter update (citing LeCun et al., 1998a). 1998a. Factual · 2006 Their own work or organisation
Not checkable Not checkable 2005 Face recognition: the gender of the person, the orientation of the face. Not checkable · 2005
Not checkable Not checkable 2005 MAP Inference: Given an input X, find the value of Y that minimizes the energy: Not checkable · 2005
Not checkable Not checkable 2005 Variables that would make the task easier if they were known: Not checkable · 2005
Not checkable Not checkable 2005 Do not answer a more complex question than necessary Not checkable · 2005
Not checkable Not checkable 2005 Minimize the Regularization term: Make W as short as possible. Not checkable · 2005
Not checkable Not checkable 2005 This is equivalent to keeping ||W|| constant, while maximizing m. Not checkable · 2005
Not checkable Not checkable 2005 Minimize the hinge loss: make the energy of all the “good” answers smaller that the energy of any “bad” answer by at least m (the margin). Not checkable · 2005
Not checkable Not checkable 2005 Probabilistic approaches compute the distribution P(Y|X), and choose the Y that maximizes it. Not checkable · 2005
Not checkable Not checkable 2005 Computing a distribution of steering angles is of little use. Not checkable · 2005
Not checkable Not checkable 2005 W is the parameter to be learned. Not checkable · 2005
Not checkable Not checkable 2005 For each question, a different learning strategy is required Not checkable · 2005
Not checkable Not checkable 2005 Pick an optimization method. Not checkable · 2005
Not checkable Not checkable 2005 Linear Model + Margin Loss + Regularization = SVM Not checkable · 2005
Not checkable Not checkable 2005 E(W,Y,X): is a scalar energy function (a.k.a. Contrast function) that measures the “compatibility” between Y and X. Not checkable · 2005
Not checkable Not checkable 2005 Object recognition: the pose parameters of the object (location, orientation, scale), the lighting conditions. Not checkable · 2005
Not checkable Not checkable 2005 Handwriting Recognition: the segmentation of the line into characters. Not checkable · 2005
Not checkable Not checkable 2005 Design an architecture: a particular form for E(W,Y,X). Not checkable · 2005
Not checkable Not checkable 2005 Probabilistic Inference: simply marginalize over Z. Not checkable · 2005
Not checkable Not checkable 2005 A factor graph is a general way to represent a graphical model. Not checkable · 2005
Not checkable Not checkable 2005 For decision making, we need no normalization. Not checkable · 2005
Not checkable Not checkable 2005 Sometimes, we do not need probabilities. Not checkable · 2005
Not checkable Not checkable 2005 The internal structure of the box is called the architecture of the model. Not checkable · 2005
Not checkable Not checkable 2005 A model measures the goodness of a combination of observed variables (X) and variables to be predicted (Y). Not checkable · 2005
Not checkable Not checkable 2005 The questions are in increasing order of complexity. Not checkable · 2005
Not checkable Not checkable 2005 We can put anything we want in the box. Not checkable · 2005
Not checkable Not checkable 2005 Architecture + Inference Algo + Loss Function = Model Not checkable · 2005
Not checkable Not checkable 2005 In general, we will search for the value of the latent variable that allows us to get an answer (Y) of smallest energy. Not checkable · 2005
Not checkable Not checkable 2005 The machine should be designed and trained to answer the simplest question possible. Not checkable · 2005
Not checkable Not checkable 2005 Z: is the latent variable (never observed). Not checkable · 2005
Not checkable Not checkable 2005 Pick a loss function: in such a way that minimizing it with respect to W over a training set will make the inference algorithm find the… Not checkable · 2005
Not checkable Not checkable 2005 When the robot faces an obstacle, it MUST turn left of right. Not checkable · 2005
Back to the list

Yann LeCun

Open its page

Not checkable

“Minimize the hinge loss: make the energy of all the “good” answers smaller that the energy of any “bad” answer by at least m (the margin).”

Yann LeCun 2005 Energy-Based Models Part 1 Introduction (IPAM slides) (cs.nyu.edu)

Minimize the hinge loss: make the energy of all the “good” answers smaller that the energy of any “bad” answer by at least m (the margin).

Not checkable An opinion, value judgement or statement with nothing to check.

Cite this claim

Yann LeCun, “Minimize the hinge loss: make the energy of all the “good” answers smaller that the energy of any “bad” answer by at least m (the margin).” (said 2005). Verdict: not checkable. Reliability API, claim c_3841a8f4a68a, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_3841a8f4a68a

Claims without a verdict

Is a verdict here wrong, or do you speak for Yann LeCun? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.