Literature
Display
← All papers

Introduction to Latent Variable Energy-Based Models: A Path Towards Autonomous Machine Intelligence

Anna Dawid · Yann LeCun

· Journal of Statistical Mechanics: Theory and Experiment ↗

World models9 min text · CondensedOriginal paper ↗
Contents

Overview

Abstract

Imagine planning a route without predicting the color of every leaf along the way. This tutorial develops that distinction: learn representations that keep what matters for a prediction, score compatible possibilities with low energy, and use hidden variables to describe uncertainty. It explains energy-based models, the danger of trivial constant representations, and ways to prevent that collapse. JEPA combines these ingredients to predict one representation from another; H-JEPA stacks predictions at different levels of detail and time. The work teaches a proposed route toward autonomous agents. It does not demonstrate a complete agent or establish that probabilistic learning cannot work.

1 Introduction

These lecture notes follow LeCun’s 2022 Les Houches lectures and explain the ideas behind his autonomous-intelligence position paper. Their motivation is the gap between successful task-specific systems and learning that transfers efficiently across new situations. Read them as a conceptual tutorial, with historical examples, rather than a benchmark comparison of modern systems.

2 Towards autonomous machine intelligence

Large neural networks and fast matrix multiplication enabled major advances in perception and language. The authors use these successes to motivate a next question: what additional machinery would allow a system to learn a world model and use it to plan?

2.1 Applications of machine learning today

Supervised learning learns from supplied answers; reinforcement learning learns through interactions and rewards. Self-supervised pretraining instead obtains a training signal from relationships inside the observations themselves. The following examples describe applications available when these lectures were prepared.

Transportation

Driver-assistance systems use learned perception for functions such as braking and lane keeping. The paper distinguishes those useful components and geographically restricted robotaxis from the much broader demand of autonomous driving in unfamiliar conditions.

Translation, transcription, accessibility

Translation and speech systems illustrate reusable internal representations: first encode the input, then decode the required output. Multilingual models can share this machinery across many language pairs, including pairs with little direct training data.

Online safety and security

Content moderation combines learned language and visual representations with human review and matching against known material. The historical figures in the paper illustrate deployment at scale; they do not resolve the separate question of who should decide what content is acceptable.

Computer vision

Visual models can identify objects, estimate human pose, and label image pixels. Semantic segmentation assigns a class to each pixel; instance segmentation additionally separates individual objects of the same class.

Medical analysis

The fastMRI example uses learned reconstruction to estimate images from fewer measurements. It illustrates how a network can exploit structure in data. The paper also points toward predicting useful outputs directly from measurements instead of always reconstructing a human-viewable image first.

Biomedical sciences

The examples range from modeling neural activity and gene regulation to predicting protein structure. They show the breadth of learned representations, while remaining distinct scientific tasks rather than evidence of a single general-purpose autonomous system.

Physics

In physics, networks can act as fast approximations to expensive simulations, represent quantum states, or recognize patterns in experimental and astronomical data.

Chemistry and material science

In chemistry and materials research, learned models can help search for compounds, approximate costly calculations, and propose materials with useful properties.

2.2 Limitations of the current machine learning

The authors argue that future agents need three abilities: learn predictive world representations from observation, use those models for deliberate reasoning, and organize long tasks into shorter subplans. Their comparisons with human learning motivate this research program. Broad claims about the limits of language models and the state of applications belong to the 2022–2023 context of the notes.

Current ML systems still miss crucial requirements for the future AI systems, such as a basic understanding of the world and humans that may be called “common sense” which we understand here as the ability to use models of the world to fill in information about the world that is unavailable from perception or memory (e.g., to predict the future) [6, 32].
Refer to caption
Figure 1: The modular structure of an autonomous AI proposed by LeCun in Ref. [6]. Drawings generated by DALL-E 2 [12].

2.3 New paradigm to autonomous intelligence

Think of the architecture as a planning loop. Perception estimates the present; a world model predicts candidate futures; an actor searches for actions with low expected cost. An immediate intrinsic cost and a learned critic score those futures, short-term memory supplies recent experience, and a configurator selects the current task. This resembles model predictive control with a learned model. Fully differentiable components are an architectural proposal, not a demonstrated complete implementation.

Refer to caption
Figure 2: SSL (SSL). (a) In SSL, the system is trained to predict hidden parts of the input (in orange) from visible parts of the input (in blue). (b) SSL will play a central role in the future AI systems. SSL is the cake (provides millions of information bits per sample), SL is the icing (10-10,000 bits per sample), RL is the cherry on top (a few bits of information for some samples). A cake image was generated by DALL-E 2 [12].

2.4 Self-supervised learning and representing uncertainty

Predicting a single exact future is difficult because several outcomes may be possible. A pixelwise squared-error predictor can average them into a blurry result. The proposed alternative is to retain predictable, useful information in a representation and represent remaining uncertainty separately. Avoiding explicit normalization can remove one expensive computation; it does not make high-dimensional inference or learning automatically easy.

Therefore, a great challenge of SSL for the future AI systems is (1) to represent uncertainty in the prediction and (2) to allow for multiple predictions being equally probable for a single input, i.e., multimodal predictions. The proposed solution is, therefore, to replace probabilistic models with EBM (EBM) and combine them with latent variables for handling uncertainty and task-dependent information redundancy as with the hierarchical structure for complex planning within the architecture called H-JEPA (H-JEPA).

2.5 Structure of the manuscript

The tutorial first defines energy and latent variables, then explains training and collapse. Classical models provide concrete examples before the final sections assemble the ingredients into JEPA and H-JEPA.

3 Introduction to energy-based models

An energy function is a compatibility score: lower means that an input and a candidate answer fit together better. Inference searches for a low-energy answer while keeping the input fixed. Training is a different optimization problem: change the model’s parameters so observed relationships receive useful scores. Multiple low-energy answers can express several compatible outcomes.

\Acfp EBM capture dependencies between variables xx and yy by associating a scalar energy F⁡(x,y)F(x,y) to each configuration of the variables [37, 38]. Low energies should be assigned to compatible pairs of xx and yy (e.g., slow driving when snow is present) and high energies to incompatible pairs. The inference consists then in finding values of yy that minimize F𝒘​(x,y)F_{\bm{w}}(x,y): yˇ=argminy​F𝒘​(x,y)\check{y}=\mathrm{argmin}_{y}F_{\bm{w}}(x,y). Therefore, the energy function is not the objective function to minimize within the learning! The energy is used only for inference.
Refer to caption
Figure 3: Towards EBM (EBM). (a) To achieve multimodal predictions in high dimensions, we can replace probabilistic models with EBM. Then, instead of minimizing the divergence measure between the prediction and the target, we look for yy’s that satisfy a set of constraints posed by xx, expressed as the energy function, F⁡(x,y)F(x,y). A trained EBM should assign low energies to yy’s that are a good continuation of xx (in case of video or text) or that are compatible or similar (in case of images of an object taken from different angles). (b) Exemplary energy function capturing the dependence of xx and yy (which is y=x2y=x^{2}) from the training set, represented as blue points. The applied architecture is presented in Fig. 5(a). Note that the energy function is not unique given only the training data set!

3.1 Energy-based models vs. probabilistic models

An energy model can rank answers without assigning probabilities that sum to one. To turn its energies into probabilities, exponentiate negative energy and divide by a partition function that sums or integrates over all answers. That denominator can be costly. Unnormalized scores provide flexibility, but their scales are not automatically comparable across models and their uncertainty is not calibrated. These are tradeoffs; intractability depends on the model and inference method.

Equations 1–2
−logP𝒘(y1,…,yn∣𝐱1,…,𝐱n)=−log∏iP𝒘(yi∣𝐱i)=−∑ilogP𝒘(yi∣𝐱i),-\log P_{\bm{w}}(y_{1},\ldots,y_{n}\mid\mathbf{x}_{1},\ldots,\mathbf{x}_{n})=-\log\prod_{i}P_{\bm{w}}(y_{i}\mid\mathbf{x}_{i})=-\sum_{i}\log P_{\bm{w}}(y_{i}\mid\mathbf{x}_{i})\,, (1)
P⁡(y∣x)=exp⁡[−β​F​(x,y)]∫⁡d​y′​exp⁡[−β​F​(x,y′)],P(y\mid x)=\frac{\exp{\left[-\beta F(x,y)\right]}}{\int\mathop{}\!\mathrm{d}y^{\prime}\exp{\left[-\beta F(x,y^{\prime})\right]}}\,, (2)

3.2 Latent variable energy-based models

A latent variable describes an unobserved factor, such as an object’s pose or another driver’s possible behavior. For the ellipse example, the latent angle selects a point on the curve. Minimizing over angles finds the closest point; marginalizing combines contributions from all angles. Limiting latent capacity matters: otherwise the latent variable can carry the entire answer and let the model ignore its input. In the source prose, an argmin is sometimes written where an energy value would require a min; the distinction is the optimizing angle versus the resulting score.

Prediction problem Examples of latent variables
Face recognition
the gender of the person, the orientation of their face,
the color of their eyes/hair
Object recognition
or scene parsing
the pose parameters of the object (location, orientation, scale),
lighting conditions, segmentation of the image into component
objects, assignment of labels to objects
Speech tagging sentence segmentation into syntactic units, the parse tree
Speech recognition sentence segmentation into phonemes or phones
Handwriting recognition the segmentation of the line into characters
Table 1: Examples of prediction problems and potentially relevant latent variables
In a structured prediction problem, we assume that the data has some unknown structure that the learner must resolve to make an accurate prediction.
Original figure
Figure 4: Latent variable EBM (EBM). (a) Inference in latent variable EBM additionally includes the minimization (or marginalization) with respect to the latent variable. (b) An example of the latent variable EBM in the problem of finding the distance of a green point yy from an ellipse learned from the training points depicted as blue dots. The latent variable here encoded the angle at which the closest to yy point on the ellipse lies.
An important feature of latent variable models is that while they are composed of deterministic functions, they are non-deterministic as the latent variable they depend on is determined outside the model.

4 Training an energy-based model

Simply reducing the mismatch of observed pairs can have a useless shortcut: both encoders output the same constant, so every pair appears compatible. A generous latent code can create a similar shortcut by memorizing the answer. The training problem must therefore make correct pairs fit while limiting the set of other inputs that also receive low energy. The following sections present alternative ways of doing this.

The main principle of training an EBM is to sculpt the energy landscape such that data points that are probable or compatible have low energy, while unlikely data points have high energy. There are two main approaches: contrastive methods and regularized and architectural methods.
Original figure
Figure 5: \AcfpEBM can collapse. (a) Standard deterministic architecture for prediction or regression, where energy function F𝒘​(x,y)F_{\bm{w}}(x,y) is the distance between the NN prediction for xx and the yy itself, is immune to collapse. (b) An example of an EBM that can collapse.
To prevent the energy collapse, i.e., flat energy landscape, we need to constrain our energy surface so that data points have low energies but also that points outside the regions of high data density have higher energies.
Refer to caption
Figure 6: Training of an EBM (EBM). (a) Proper training minimizes training samples’ energy while preventing energy collapse. (b) Contrastive methods minimize the energy of training samples while increasing the energy of samples outside the training set. (c) Regularized methods limit the volume of space that the low-energy region can take. Adapted from Ref. [6].

4.1 Contrastive methods

Contrastive learning supplies incompatible alternatives as well as compatible pairs. A hinge loss requires the real pair to beat a negative pair by a margin. The margin prevents a constant energy function from satisfying the objective. Batch-based objectives such as InfoNCE compare a larger group of alternatives.

Contrastive methods aim to push down the energy of training samples and to pull up the energy of suitably-generated contrastive samples that by definition are out of the training set. They work for any architecture but scale badly with data dimension.
Equation 3
Lhinge​(x,y,y^,𝒘)=[F𝒘​(x,y)−F𝒘​(x,y^)+m⁡(y,y^)]+.L_{\rm hinge}(x,y,\hat{y},\bm{w})=[F_{\bm{w}}(x,y)-F_{\bm{w}}(x,\hat{y})+m(y,\hat{y})]^{+}\,. (3)

4.1.1 How to generate contrastive points

Negative examples may be enumerated, sampled, selected because the model currently confuses them, or generated by another network. In a Siamese model, views of different objects can supply negatives. The challenge is finding informative alternatives without spending excessive compute searching a huge output space.

The need for generating contrastive samples is the source of bad (even exponential!) scaling of contrastive methods with the data dimension. As real-world data can be extremely complex and high-dimensional, this unfavorable scaling makes contrastive methods unlikely to lead to autonomous intelligence of the future.

4.1.2 Maximum likelihood as a special case of a contrastive method

For a Gibbs model, the likelihood gradient has two effects: lower the energy of observed data and raise energy in regions the current model favors. The second part can be estimated using samples from that model. Sampling and regularization can be demanding, especially in high dimensions; the discussion motivates alternatives rather than proving that every likelihood-based model must fail.

Equations 4–5
LNLL​(x,y,𝒘)=−1β​log​P𝒘​(y∣x)=F𝒘​(x,y)+1β​log​[∫⁡d​y′​exp​[−β​F𝒘​(x,y′)]],L_{\rm NLL}(x,y,\bm{w})=-\frac{1}{\beta}\log P_{\bm{w}}(y\mid x)=F_{\bm{w}}(x,y)+\frac{1}{\beta}\log\left[\int\mathop{}\!\mathrm{d}y^{\prime}\exp{\left[-\beta F_{\bm{w}}\left(x,y^{\prime}\right)\right]}\right]\,, (4)
∂[1β​log​[∫⁡d​y′​exp​[−β​F𝒘​(x,y′)]]]∂𝒘=−∫⁡d​y′​P𝒘​(y′∣x)​∂F𝒘​(x,y′)∂𝒘.\frac{\partial\left[\frac{1}{\beta}\log\left[\int\mathop{}\!\mathrm{d}y^{\prime}\exp{\left[-\beta F_{\bm{w}}\left(x,y^{\prime}\right)\right]}\right]\right]}{\partial\bm{w}}=-\int\mathop{}\!\mathrm{d}y^{\prime}P_{\bm{w}}\left(y^{\prime}\mid x\right)\frac{\partial F_{\bm{w}}\left(x,y^{\prime}\right)}{\partial\bm{w}}\,. (5)
It means that the minimization of LNLLL_{\rm NLL} consists in minimizing the energy of training samples while pulling up the energy of Monte Carlo-generated samples. Negative log-likelihood is then a contrastive method!

4.2 Architectural and regularized methods

Another strategy limits how much of the input space can receive low energy. A small bottleneck, a finite set of codes, sparsity, or controlled noise can restrict reconstruction capacity. Regularizers can enforce related constraints. The point is to remove the all-inputs-fit shortcut without explicitly generating many negatives.

Another option to prevent collapse is to use regularized methods. These methods aim to minimize the volume of space that the low-energy points can take. When the energy of input data is lowered, it is required to increase the energy of some parts of the space. Because regularized methods are less likely than contrastive methods to fall victim to the curse of dimensionality, they seem more promising to train EBM of the future.

4.2.1 Regularized autoencoders

Autoencoding compresses an input and reconstructs it. Principal component analysis restricts the representation to a low-dimensional linear space; a bottleneck autoencoder permits a nonlinear space; k-means permits only a finite collection of prototypes; sparse coding discourages using many code components at once. All balance fitting the data against the capacity available to explain it.

Equation
Lreg​(y,𝒘)=D⁡(y,Dec​(Enc​(y)))+R⁡(Enc​(y)),\displaystyle L_{\rm reg}(y,\bm{w})=D(y,\text{Dec}(\text{Enc}(y)))+R(\text{Enc}(y))\,,

4.2.2 Regularization through the variational marginalization of a latent variable

A stochastic latent code can be regularized by comparing its input-dependent distribution with a reference distribution using KL divergence. If the two stay close, the code cannot freely transmit every detail of the answer. The expected reconstruction energy rewards a useful code, while the information constraint prevents unrestricted memorization.

Equation
⟨F⁡(y)⟩z=∫⁡d​z​q​(z∣y)​E​(y,z)​𝑑z.\displaystyle\langle F(y)\rangle_{z}=\int\mathop{}\!\mathrm{d}zq(z\mid y)E(y,z)\,dz.

5 Examples of energy-based models

Hopfield networks, Boltzmann machines and denoising autoencoders show how an energy interpretation applies to familiar models. Their different training rules make the contrast between fitting observations and controlling unwanted low-energy states concrete.

5.1 Hopfield networks

A Hopfield network stores patterns in symmetric connections. Repeatedly updating one neuron according to the others descends toward a locally stable state, which can recover a stored pattern from partial information. A simple Hebbian rule strengthens connections between units active together. Unwanted stable states can also appear, so low energy need not identify a true stored memory.

Equations 6–7
F𝒘(𝒚)=−∑i​jyiwi​jyjF_{\bm{w}}(\bm{y})=-\sum_{ij}y_{i}w_{ij}y_{j} (6)
wi​j←wi​j−∂L⁡(𝒚train,𝒘)∂wi​j=wi​j+yitrain​yjtrain.w_{ij}\leftarrow w_{ij}-\frac{\partial L(\bm{y}^{\rm train},\bm{w})}{\partial w_{ij}}=w_{ij}+y^{\rm train}_{i}y^{\rm train}_{j}\,. (7)
Original figure
Figure 7: Examples of EBM (EBM). (a) Hopfield network. (b) Boltzmann machine, which is the Hopfield network with hidden units. (c) Restricted Boltzmann machine has only connections between visible and hidden units.

5.2 Boltzmann machines

Boltzmann machines add hidden units and stochastic states. Training compares correlations under observed data with correlations produced by the model: strengthen relationships supported by the data and weaken those the model overproduces. Estimating those model correlations requires sampling, which can be costly. Restricting the connection pattern simplifies some computations without removing every sampling challenge.

Equations 8–12
E(𝒚,𝒛)=−∑i​jyiwi​jy​yyj−∑i​jziwi​jz​zzj−∑i​jyiwi​jy​zzj,E(\bm{y},\bm{z})=-\sum_{ij}y_{i}w_{ij}^{yy}y_{j}-\sum_{ij}z_{i}w_{ij}^{zz}z_{j}-\sum_{ij}y_{i}w_{ij}^{yz}z_{j}\,,\\ (8)
F(𝒚)=−log∑𝒛exp[−E(𝒚,𝒛)]F(\bm{y})=-\log\sum_{\bm{z}}\exp{\left[-E(\bm{y},\bm{z})\right]} (9)
LBoltzmann(y,𝒘)=F𝒘(y)+log∑y′exp[F𝒘(y′)],L_{\rm Boltzmann}(y,\bm{w})=F_{\bm{w}}(y)+\log\sum_{y^{\prime}}\exp{\left[F_{\bm{w}}\left(y^{\prime}\right)\right]}\,, (10)
∂LBoltzmann​(y,𝒘)∂wi​j=∂F𝒘​(y)∂𝒘−∑y′P⁡(y′)​∂F𝒘​(y′)∂𝒘withP⁡(y)=exp⁡[−F𝒘​(y)]∑y′exp⁡[−F𝒘​(y′)]\frac{\partial L_{\rm Boltzmann}(y,\bm{w})}{\partial w_{ij}}=\frac{\partial F_{\bm{w}}(y)}{\partial\bm{w}}-\sum_{y^{\prime}}P\left(y^{\prime}\right)\frac{\partial F_{\bm{w}}\left(y^{\prime}\right)}{\partial\bm{w}}\quad\text{with}\quad P(y)=\frac{\exp\left[-F_{\bm{w}}(y)\right]}{\sum_{y^{\prime}}\exp\left[-F_{\bm{w}}\left(y^{\prime}\right)\right]} (11)
∂LBoltzmann​(y,𝒘)∂wi​j=\displaystyle\frac{\partial L_{\rm Boltzmann}(y,\bm{w})}{\partial w_{ij}}= ∑z′P⁡(z′∣y)​∂E𝒘​(y,z′)∂𝒘−∑y′,z′P⁡(y′,z′)​∂E𝒘​(y′,z)∂𝒘 where\displaystyle\sum_{z^{\prime}}P\left(z^{\prime}\mid y\right)\frac{\partial E_{\bm{w}}\left(y,z^{\prime}\right)}{\partial\bm{w}}-\sum_{y^{\prime},z^{\prime}}P\left(y^{\prime},z^{\prime}\right)\frac{\partial E_{\bm{w}}\left(y^{\prime},z\right)}{\partial\bm{w}}\quad\text{ where } (12)
P⁡(z∣y)=\displaystyle P(z\mid y)= exp⁡[−E𝒘​(y,z)]∑z′exp⁡[−E𝒘​(y,z′)] and P⁡(y,z)=exp⁡[−E𝒘​(y,z)]∑y′,z′exp⁡[−E𝒘​(y′,z′)].\displaystyle\frac{\exp\left[-E_{\bm{w}}(y,z)\right]}{\sum_{z^{\prime}}\exp\left[-E_{\bm{w}}\left(y,z^{\prime}\right)\right]}\quad\text{ and }\quad P(y,z)=\frac{\exp\left[-E_{\bm{w}}(y,z)\right]}{\sum_{y^{\prime},z^{\prime}}\exp\left[-E_{\bm{w}}\left(y^{\prime},z^{\prime}\right)\right]}\,.
Refer to caption
Figure 8: Denoising and masked AE (AE). (a) A scheme of a denoising (and masked) AE. (b) An exemplary denoising problem.

5.3 Denoising autoencoders

A denoising autoencoder learns to map a corrupted observation back toward clean data. The spiral example makes the goal visible: noisy points should return to the curve. If a point lies equally near two branches, a single deterministic output cannot express both possibilities. Masking creates a related prediction task, while latent variables and representation-space prediction offer ways to handle ambiguity.

6 Building block of proposed autonomous systems of the future

JEPA combines learned representations, prediction and a mechanism for handling uncertainty. Its aim is to compare the information needed for a prediction without requiring a complete reconstruction of the raw observation.

6.1 Joint embedding predictive architecture

Two encoders map related inputs into representations, and a predictor estimates the target representation using the context and, when needed, a latent variable. Constant outputs must be prevented, and the latent channel must not carry the whole answer. VICReg illustrates explicit regularization: keep each feature’s batch variance above a threshold, reduce cross-feature covariance, and match paired representations. These constraints encourage informative, nonredundant features; decorrelation alone is not a guarantee of statistical independence or complete retained information.

Original figure
Figure 9: A JEPA (JEPA). It consists of two parallel encoders, which can be different in general, learning representations of xx and yy, i.e., sxs_{x} and sys_{y}. Those encoders can learn to filter irrelevant information about xx and yy. The aim of JEPA is to predict sys_{y} out of sxs_{x} with help of a latent variable zz that can represent relevant information about yy that may be missing from xx. The energy is the distance between those representations.
Original figure
Figure 10: Regularized loss function for training a JEPA (JEPA). Apart from the prediction error that needs to be minimized, there are also additional terms that maximize the information content of representations about their respective input data points and minimize the information content of a latent variable. Adapted from Ref. [6].

6.2 Hierarchical joint embedding predictive architecture

A hierarchy uses detailed representations for short-horizon predictions and more abstract representations for longer horizons. For example, a high level might plan to enter a room while a lower level chooses the next movement. Stacking JEPAs is proposed as a way to learn these levels and support hierarchical planning. The notes present the construction and motivation, rather than an evaluated end-to-end autonomous agent.

Original figure
Figure 11: H-JEPA (H-JEPA) can be used for multiscale planning where a lower-level JEPA-1 is used for short-time predictions based on many details, and a higher-level JEPA-2 makes predictions for the longer term based on fewer details. Adapted from Ref. [6].

7 Conclusion

The tutorial’s main contribution is a connected vocabulary for a research direction: energy scores compatibility, latent variables represent unresolved factors, regularization prevents shortcuts, and JEPA predicts in a learned representation space. H-JEPA adds abstraction and multiple timescales. Which representations, constraints and planning procedures make the full proposal work remains an empirical and algorithmic question.

Acknowledgements

The notes are based on LeCun’s 2022 Les Houches lectures and acknowledge discussions with Alfredo Canziani, Lucas Clarte and Max Daniels.

Funding information

The authors acknowledge support from Polish, Spanish, Catalan and European research programs, including support for the ICFO group. The views expressed are the authors’ own.

List of acronyms

Overview

The source acronym list is retained below; the Terms tab also supplies definitions beside the reading text.

Appendix A How to turn energies into probabilities

The appendix gives the probability dictionary for energy models. A Gibbs distribution converts low energies into high probabilities by exponentiating and normalizing. Conditioning holds an observed variable fixed; marginalizing integrates out a hidden variable. The resulting free energy is a logarithm of a sum or integral of exponentiated energies. Minimization keeps the best latent explanation, whereas marginalization combines all explanations with energy-dependent weights.

Equations A1–A5
P𝒘​(y)=exp⁡[−β​F𝒘​(y)]∑y′exp⁡[−β​F𝒘​(y′)]andP𝒘​(y)=exp⁡[−β​F𝒘​(y)]∫⁡d​y′​exp⁡[−β​F𝒘​(y′)]P_{\bm{w}}(y)=\frac{\exp{\left[-\beta F_{\bm{w}}(y)\right]}}{\sum_{y^{\prime}}\exp{\left[-\beta F_{\bm{w}}\left(y^{\prime}\right)\right]}}\quad\mathrm{and}\quad P_{\bm{w}}(y)=\frac{\exp{\left[-\beta F_{\bm{w}}(y)\right]}}{\int\mathop{}\!\mathrm{d}y^{\prime}\exp{\left[-\beta F_{\bm{w}}\left(y^{\prime}\right)\right]}} (A1)
P𝒘​(y,z)=exp⁡[−β​E𝒘​(y,z)]∫⁡d​y′​∫⁡d​z′​exp⁡[−β​E𝒘​(y′,z′)]P_{\bm{w}}(y,z)=\frac{\exp{\left[-\beta E_{\bm{w}}(y,z)\right]}}{\int\mathop{}\!\mathrm{d}{y^{\prime}}\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(y^{\prime},z^{\prime}\right)\right]}} (A2)
P𝒘​(y,z∣x)=exp⁡[−β​E𝒘​(x,y,z)]∫⁡d​y′​∫⁡d​z′​exp⁡[−β​E𝒘​(x,y′,z′)]P_{\bm{w}}(y,z\mid x)=\frac{\exp{\left[-\beta E_{\bm{w}}(x,y,z)\right]}}{\int\mathop{}\!\mathrm{d}{y^{\prime}}\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(x,y^{\prime},z^{\prime}\right)\right]}} (A3)
P𝒘​(y∣x)=∫⁡d​z′​P𝒘​(y,z′∣x)=∫⁡d​z′​exp⁡[−β​E𝒘​(x,y,z′)]∫⁡d​y′​∫⁡d​z′​exp⁡[−β​E𝒘​(x,y′,z′)]P_{\bm{w}}(y\mid x)=\int\mathop{}\!\mathrm{d}{z^{\prime}}P_{\bm{w}}\left(y,z^{\prime}\mid x\right)=\frac{\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(x,y,z^{\prime}\right)\right]}}{\int\mathop{}\!\mathrm{d}{y^{\prime}}\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(x,y^{\prime},z^{\prime}\right)\right]}} (A4)
P𝒘​(y∣x)=∫⁡d​z′​exp⁡[−β​E𝒘​(x,y,z′)]∫⁡d​y′​∫⁡d​z′​exp⁡[−β​E𝒘​(x,y′,z′)]=exp⁡[log⁡(a)]=aexp⁡[−β​(−1β)​log⁡∫⁡d​z′​exp⁡[−β​E𝒘​(x,y,z′)]]∫⁡d​y′​exp​[−β​(−1β)​log⁡∫⁡d​z′​exp⁡[−β​E𝒘​(x,y′,z′)]]=exp⁡[−β​[−1β​log⁡∫⁡d​z′​exp⁡[−β​E𝒘​(x,y,z′)]]]∫⁡d​y′​exp​[−β​[−1β​log⁡∫⁡d​z′​exp⁡[−β​E𝒘​(x,y′,z′)]]]=exp⁡[−β​F𝒘​(x,y)]∫⁡d​y′​exp⁡[−β​F𝒘​(x,y′)]\begin{gathered}P_{\bm{w}}(y\mid x)=\frac{\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(x,y,z^{\prime}\right)\right]}}{\int\mathop{}\!\mathrm{d}{y^{\prime}}\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(x,y^{\prime},z^{\prime}\right)\right]}}\\ \mathrel{\overset{\makebox[0.0pt]{\mbox{\tiny$\exp{\left[\log(a)\right]}=a$}}}{=}}\,\,\,\,\,\,\,\frac{\exp{\left[-\beta(-\frac{1}{\beta})\log\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(x,y,z^{\prime}\right)\right]}\right]}}{\int\mathop{}\!\mathrm{d}{y^{\prime}}\exp{\left[-\beta(-\frac{1}{\beta})\log\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}\left(x,y^{\prime},z^{\prime}\right)\right]}\right]}}\\ =\frac{\exp{\left[-\beta\left[-\frac{1}{\beta}\log\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}(x,y,z^{\prime})\right]}\right]\right]}}{\int\mathop{}\!\mathrm{d}{y^{\prime}}\exp{\left[-\beta\left[-\frac{1}{\beta}\log\int\mathop{}\!\mathrm{d}{z^{\prime}}\exp{\left[-\beta E_{\bm{w}}(x,y^{\prime},z^{\prime})\right]}\right]\right]}}\\ =\frac{\exp{\left[-\beta F_{\bm{w}}(x,y)\right]}}{\int\mathop{}\!\mathrm{d}{y^{\prime}}\exp{\left[-\beta F_{\bm{w}}(x,y^{\prime})\right]}}\end{gathered} (A5)
Concept demonstration

Open image ↗