{"work":{"id":"050fb4ab-62d4-4150-866a-77bc502eca22","openalex_id":"https://openalex.org/W2547875792","doi":"10.48550/arxiv.1611.01144","arxiv_id":"1611.01144","raw_key":null,"title":"Categorical Reparameterization with Gumbel-Softmax","authors":null,"authors_text":"Eric Jang, Shixiang Gu, Ben Poole","year":2016,"venue":"stat.ML","abstract":"Categorical variables are a natural choice for representing discrete structure in the world. However, stochastic neural networks rarely use categorical latent variables due to the inability to backpropagate through samples. In this work, we present an efficient gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample from a novel Gumbel-Softmax distribution. This distribution has the essential property that it can be smoothly annealed into a categorical distribution. We show that our Gumbel-Softmax estimator outperforms state-of-the-art gradient estimators on structured output prediction and unsupervised generative modeling tasks with categorical latent variables, and enables large speedups on semi-supervised classification.","external_url":"https://arxiv.org/abs/1611.01144","cited_by_count":3241,"metadata_source":"pith","metadata_fetched_at":"2026-08-05T02:28:24.338817+00:00","pith_arxiv_id":"1611.01144","created_at":"2026-05-09T01:59:36.105650+00:00","updated_at":"2026-08-05T02:28:24.338817+00:00","title_quality_ok":true,"display_title":"Categorical Reparameterization with Gumbel-Softmax","render_title":"Categorical Reparameterization with Gumbel-Softmax"},"hub":{"state":{"work_id":"050fb4ab-62d4-4150-866a-77bc502eca22","tier":"super_hub","tier_reason":"100+ Pith inbound or 10,000+ external citations","pith_inbound_count":113,"external_cited_by_count":3241,"distinct_field_count":23,"first_pith_cited_at":"2019-06-19T20:29:31+00:00","last_pith_cited_at":"2026-07-07T15:49:04+00:00","author_build_status":"needed","summary_status":"needed","contexts_status":"needed","graph_status":"needed","ask_index_status":"needed","reader_status":"not_needed","recognition_status":"not_needed","updated_at":"2026-08-21T19:39:22.489941+00:00","tier_text":"super_hub"},"tier":"super_hub","role_counts":[{"context_role":"method","n":6},{"context_role":"background","n":2}],"polarity_counts":[{"context_polarity":"use_method","n":6},{"context_polarity":"background","n":2}],"runs":{"ask_index":{"job_type":"ask_index","status":"succeeded","result":{"title":"Categorical Reparameterization with Gumbel-Softmax","claims":[{"claim_text":"Categorical variables are a natural choice for representing discrete structure in the world. However, stochastic neural networks rarely use categorical latent variables due to the inability to backpropagate through samples. In this work, we present an efficient gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample from a novel Gumbel-Softmax distribution. This distribution has the essential property that it can be smoothly annealed into a categorical distribution. We show that our Gumbel-Softmax estimator outperforms state-o","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"metric Constraints in Inverse-Designed Nanophotonic Devices.ACS Photonics2020, 7, 3141-3151. 20 (27) Wang, C.; Chen, N.; Heidrich, W. dO: A Differentiable Engine for Deep Lens Design of Computational Imaging Systems.IEEE Transactions on Computational Imaging2022, 8, 905-916. (28) Goodman, J.Introduction to Fourier Optics, fourth edition ed.; W. H. Freeman, 2017. (29) Taflove, A.; Hagness, S. C.Computational Electrodynamics: The Finite-Difference Time-Domain Method, 3rd ed.; Artech House, 2005. (","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"demand, infra features, time encodings) are processed as 𝐻 (1) =𝜎 (̂𝐴Φ𝑊 0 +𝟏𝑏 ⊤ 0 ), 𝐸=𝐻 (2) =𝜎 (̂𝐴 𝐻(1) 𝑊1 +𝟏𝑏 ⊤ 1 ),(24) yielding hex embeddings𝐸that inform both the actor and critics. In our implementation,𝜎is the SiLU activation. 6.2. Actor with Gumbel-Softmax and squashed Gaussian To allow backpropagation through the discrete action selection, we use the Gumbel-Softmax reparameterization [26]. For a set of logits(𝑢 𝑘), the discrete intentioñ 𝑧is sampled as: ̃ 𝑧𝑘 = exp ((𝑢𝑘 +𝑔 𝑘)∕𝜏) ∑ 𝑗 exp","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":", computing counterfactual features xCF u from zu and the recommended action A=do(x u,i :=a ii∈I ), where I denotes the set of globally actionable features chosen for intervention. Recourse recommendations are generated via two jointly trained networks. TheMask Networklearns a Bernoulli distribution over feature interventions, producing a binary mask I via the Gumbel trick [18], while enforcing global actionability through a fixed actionability mask. TheAction Networkpredicts intervention magnit","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"To further eliminate redundant Gaussians in over-populated areas and suppress residual floating artifacts, we integrate MaskGaussian [ 45]. Instead of relying on heuristic hard-pruning, MaskGaussian models the existence of each Gaussian as a probabilistic entity. Concretely, for the k-th Gaussian, a binary mask Mk ∈ {0,1} is sampled via Gumbel-Softmax [25] from learnable mask logits. This mask is then incorporated into the tile-based rasterizer through amasked-rasterizationscheme. For a given pi","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"permutations during optimization could converge to solutions far from any single permutation (e.g., a uniform matrix), creating a significant gap between the loss measured during optimization and the attack's true effectiveness. Thus, we must encourage the soft matrices to be \"sharp\" and structurally similar to a hard permutation. A classic strategy is to follow Gumbel-Softmax [20, 31] to introduce temperature τ into the Sinkhorn algorithm, as S(θ/τ) . Lowering τ is analogous to pushing the opti","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"on more general generation tasks [7, 17, 20, 53, 119]. Discrete diffusion language models remove explicit left-to-right factorization [25, 35, 36, 110], yet they still typically perform observation recovery in discrete token space, leading to costly multi-step sampling, while intermediate discrete states are not well suited to stably represent global semantic structure [40, 62, 86, 90, 94, 115, 116]. Continuous diffusion methods further introduce continuous representation spaces [28, 81, 89], bu","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Categorical Reparameterization with Gumbel-Softmax because it crossed a citation-hub threshold. Current citing contexts most often use it as method evidence (6 contexts).","role_counts":[{"n":6,"context_role":"method"},{"n":2,"context_role":"background"}]},"error":null,"updated_at":"2026-07-03T10:33:58.377252+00:00"},"author_expand":{"job_type":"author_expand","status":"succeeded","result":{"authors_linked":[{"id":"a54699ed-6350-46a2-922b-41bf21df87fb","orcid":null,"display_name":"Eric Jang"},{"id":"0226913a-35be-4685-b429-9d56ad1a5fe0","orcid":null,"display_name":"Shixiang Gu"},{"id":"eb2faec5-1621-4088-bc26-f7fa3009e416","orcid":null,"display_name":"Ben Poole"}]},"error":null,"updated_at":"2026-07-03T10:33:58.914991+00:00"},"context_extract":{"job_type":"context_extract","status":"succeeded","result":{"enqueued_papers":25},"error":null,"updated_at":"2026-05-14T18:20:07.554213+00:00"},"graph_features":{"job_type":"graph_features","status":"succeeded","result":{"co_cited":[{"title":"Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation","work_id":"1fe8c7c8-aff7-4b94-9096-e549d7e60789","shared_citers":6},{"title":"The concrete distribution: A continuous relaxation of discrete random variables","work_id":"03a7ba77-5b42-48c0-9d8b-c3b64720ad12","shared_citers":6},{"title":"Adam: A Method for Stochastic Optimization","work_id":"1910796d-9b52-4683-bf5c-de9632c1028b","shared_citers":5},{"title":"Auto-Encoding Variational Bayes","work_id":"97d95295-30e1-42b4-bbf6-85f0fa4edb44","shared_citers":4},{"title":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","work_id":"e96730e3-129b-4db6-b981-15ab7932e297","shared_citers":3},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":3},{"title":"LLaMA: Open and Efficient Foundation Language Models","work_id":"c018fc23-6f3f-4035-9d02-28a2173b2b9d","shared_citers":3},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":3},{"title":"RoBERTa: A Robustly Optimized BERT Pretraining Approach","work_id":"41fe12c4-e538-4890-a244-480650ed3078","shared_citers":3},{"title":"ALBERT: A Lite BERT for Self-supervised Learning of Language Representations","work_id":"aedf7950-7c35-4e28-a32d-bec290f51669","shared_citers":2},{"title":"arXiv preprint arXiv:2209.14860 , year=","work_id":"27679e01-8a7c-4326-8289-1c72cb93fece","shared_citers":2},{"title":"Bowman, Luke Vilnis, Oriol Vinyals, Andrew M","work_id":"1cfd59d5-37b6-4939-878b-01bfdd27c2ff","shared_citers":2},{"title":"Decoupled Weight Decay Regularization","work_id":"07ef7360-d385-4033-83f7-8384a6325204","shared_citers":2},{"title":"DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter","work_id":"756f9764-ecd6-4672-8043-b37c698c7ad2","shared_citers":2},{"title":"Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities","work_id":"008df105-2fdd-45d8-857a-8e35868aecb6","shared_citers":2},{"title":"Kimi K2.5: Visual Agentic Intelligence","work_id":"d690be8f-5d53-49b0-b1e7-79668eb8fcdb","shared_citers":2},{"title":"Learning latent permuta- tions with gumbel-sinkhorn networks","work_id":"f24fe268-ea6b-408a-b42f-3598cbb8400c","shared_citers":2},{"title":"Longformer: The Long-Document Transformer","work_id":"abea7a44-6668-4de7-aab6-f53a6e5aa088","shared_citers":2},{"title":"SAM 2: Segment Anything in Images and Videos","work_id":"acc13f66-d814-44f9-9688-375688bf2d4a","shared_citers":2},{"title":"SAM 3: Segment Anything with Concepts","work_id":"4a72a006-2592-4554-aad0-a9c41a9f952d","shared_citers":2},{"title":"Semi-Supervised Classification with Graph Convolutional Networks","work_id":"21fff118-807d-49cd-8229-f7087ba57b5d","shared_citers":2},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":2},{"title":"Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge","work_id":"28ea1282-d657-4c61-a83c-f1249be6d6b1","shared_citers":2},{"title":"Transformer-xl: Attentive language models beyond a fixed-length context","work_id":"eb970d64-41ff-4e44-afd0-e3bb975e0dc4","shared_citers":2}],"time_series":[{"n":1,"year":2019},{"n":31,"year":2026}],"dependency_candidates":[]},"error":null,"updated_at":"2026-05-14T18:20:22.675928+00:00"},"identity_refresh":{"job_type":"identity_refresh","status":"succeeded","result":{"items":[{"title":"Qwen3 Technical Report","outcome":"unchanged","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","resolver":"local_arxiv","confidence":0.98,"old_work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e"}],"counts":{"fixed":0,"merged":0,"unchanged":1,"quarantined":0,"needs_external_resolution":0},"errors":[],"attempted":1},"error":null,"updated_at":"2026-05-14T18:19:30.624644+00:00"},"role_polarity":{"job_type":"role_polarity","status":"succeeded","result":{"title":"Categorical Reparameterization with Gumbel-Softmax","claims":[{"claim_text":"Categorical variables are a natural choice for representing discrete structure in the world. However, stochastic neural networks rarely use categorical latent variables due to the inability to backpropagate through samples. In this work, we present an efficient gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample from a novel Gumbel-Softmax distribution. This distribution has the essential property that it can be smoothly annealed into a categorical distribution. We show that our Gumbel-Softmax estimator outperforms state-o","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"metric Constraints in Inverse-Designed Nanophotonic Devices.ACS Photonics2020, 7, 3141-3151. 20 (27) Wang, C.; Chen, N.; Heidrich, W. dO: A Differentiable Engine for Deep Lens Design of Computational Imaging Systems.IEEE Transactions on Computational Imaging2022, 8, 905-916. (28) Goodman, J.Introduction to Fourier Optics, fourth edition ed.; W. H. Freeman, 2017. (29) Taflove, A.; Hagness, S. C.Computational Electrodynamics: The Finite-Difference Time-Domain Method, 3rd ed.; Artech House, 2005. (","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"demand, infra features, time encodings) are processed as 𝐻 (1) =𝜎 (̂𝐴Φ𝑊 0 +𝟏𝑏 ⊤ 0 ), 𝐸=𝐻 (2) =𝜎 (̂𝐴 𝐻(1) 𝑊1 +𝟏𝑏 ⊤ 1 ),(24) yielding hex embeddings𝐸that inform both the actor and critics. In our implementation,𝜎is the SiLU activation. 6.2. Actor with Gumbel-Softmax and squashed Gaussian To allow backpropagation through the discrete action selection, we use the Gumbel-Softmax reparameterization [26]. For a set of logits(𝑢 𝑘), the discrete intentioñ 𝑧is sampled as: ̃ 𝑧𝑘 = exp ((𝑢𝑘 +𝑔 𝑘)∕𝜏) ∑ 𝑗 exp","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":", computing counterfactual features xCF u from zu and the recommended action A=do(x u,i :=a ii∈I ), where I denotes the set of globally actionable features chosen for intervention. Recourse recommendations are generated via two jointly trained networks. TheMask Networklearns a Bernoulli distribution over feature interventions, producing a binary mask I via the Gumbel trick [18], while enforcing global actionability through a fixed actionability mask. TheAction Networkpredicts intervention magnit","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"To further eliminate redundant Gaussians in over-populated areas and suppress residual floating artifacts, we integrate MaskGaussian [ 45]. Instead of relying on heuristic hard-pruning, MaskGaussian models the existence of each Gaussian as a probabilistic entity. Concretely, for the k-th Gaussian, a binary mask Mk ∈ {0,1} is sampled via Gumbel-Softmax [25] from learnable mask logits. This mask is then incorporated into the tile-based rasterizer through amasked-rasterizationscheme. For a given pi","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"permutations during optimization could converge to solutions far from any single permutation (e.g., a uniform matrix), creating a significant gap between the loss measured during optimization and the attack's true effectiveness. Thus, we must encourage the soft matrices to be \"sharp\" and structurally similar to a hard permutation. A classic strategy is to follow Gumbel-Softmax [20, 31] to introduce temperature τ into the Sinkhorn algorithm, as S(θ/τ) . Lowering τ is analogous to pushing the opti","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"on more general generation tasks [7, 17, 20, 53, 119]. Discrete diffusion language models remove explicit left-to-right factorization [25, 35, 36, 110], yet they still typically perform observation recovery in discrete token space, leading to costly multi-step sampling, while intermediate discrete states are not well suited to stably represent global semantic structure [40, 62, 86, 90, 94, 115, 116]. Continuous diffusion methods further introduce continuous representation spaces [28, 81, 89], bu","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Categorical Reparameterization with Gumbel-Softmax because it crossed a citation-hub threshold. Current citing contexts most often use it as method evidence (6 contexts).","role_counts":[{"n":6,"context_role":"method"},{"n":2,"context_role":"background"}]},"error":null,"updated_at":"2026-07-03T10:33:58.381353+00:00"},"summary_claims":{"job_type":"summary_claims","status":"succeeded","result":{"title":"Categorical Reparameterization with Gumbel-Softmax","claims":[{"claim_text":"Categorical variables are a natural choice for representing discrete structure in the world. However, stochastic neural networks rarely use categorical latent variables due to the inability to backpropagate through samples. In this work, we present an efficient gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample from a novel Gumbel-Softmax distribution. This distribution has the essential property that it can be smoothly annealed into a categorical distribution. We show that our Gumbel-Softmax estimator outperforms state-o","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Categorical Reparameterization with Gumbel-Softmax because it crossed a citation-hub threshold.","role_counts":[]},"error":null,"updated_at":"2026-05-14T18:20:22.681704+00:00"}},"summary":{"title":"Categorical Reparameterization with Gumbel-Softmax","claims":[{"claim_text":"Categorical variables are a natural choice for representing discrete structure in the world. However, stochastic neural networks rarely use categorical latent variables due to the inability to backpropagate through samples. In this work, we present an efficient gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample from a novel Gumbel-Softmax distribution. This distribution has the essential property that it can be smoothly annealed into a categorical distribution. We show that our Gumbel-Softmax estimator outperforms state-o","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Categorical Reparameterization with Gumbel-Softmax because it crossed a citation-hub threshold.","role_counts":[]},"graph":{"co_cited":[{"title":"Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation","work_id":"1fe8c7c8-aff7-4b94-9096-e549d7e60789","shared_citers":6},{"title":"The concrete distribution: A continuous relaxation of discrete random variables","work_id":"03a7ba77-5b42-48c0-9d8b-c3b64720ad12","shared_citers":6},{"title":"Adam: A Method for Stochastic Optimization","work_id":"1910796d-9b52-4683-bf5c-de9632c1028b","shared_citers":5},{"title":"Auto-Encoding Variational Bayes","work_id":"97d95295-30e1-42b4-bbf6-85f0fa4edb44","shared_citers":4},{"title":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","work_id":"e96730e3-129b-4db6-b981-15ab7932e297","shared_citers":3},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":3},{"title":"LLaMA: Open and Efficient Foundation Language Models","work_id":"c018fc23-6f3f-4035-9d02-28a2173b2b9d","shared_citers":3},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":3},{"title":"RoBERTa: A Robustly Optimized BERT Pretraining Approach","work_id":"41fe12c4-e538-4890-a244-480650ed3078","shared_citers":3},{"title":"ALBERT: A Lite BERT for Self-supervised Learning of Language Representations","work_id":"aedf7950-7c35-4e28-a32d-bec290f51669","shared_citers":2},{"title":"arXiv preprint arXiv:2209.14860 , year=","work_id":"27679e01-8a7c-4326-8289-1c72cb93fece","shared_citers":2},{"title":"Bowman, Luke Vilnis, Oriol Vinyals, Andrew M","work_id":"1cfd59d5-37b6-4939-878b-01bfdd27c2ff","shared_citers":2},{"title":"Decoupled Weight Decay Regularization","work_id":"07ef7360-d385-4033-83f7-8384a6325204","shared_citers":2},{"title":"DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter","work_id":"756f9764-ecd6-4672-8043-b37c698c7ad2","shared_citers":2},{"title":"Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities","work_id":"008df105-2fdd-45d8-857a-8e35868aecb6","shared_citers":2},{"title":"Kimi K2.5: Visual Agentic Intelligence","work_id":"d690be8f-5d53-49b0-b1e7-79668eb8fcdb","shared_citers":2},{"title":"Learning latent permuta- tions with gumbel-sinkhorn networks","work_id":"f24fe268-ea6b-408a-b42f-3598cbb8400c","shared_citers":2},{"title":"Longformer: The Long-Document Transformer","work_id":"abea7a44-6668-4de7-aab6-f53a6e5aa088","shared_citers":2},{"title":"SAM 2: Segment Anything in Images and Videos","work_id":"acc13f66-d814-44f9-9688-375688bf2d4a","shared_citers":2},{"title":"SAM 3: Segment Anything with Concepts","work_id":"4a72a006-2592-4554-aad0-a9c41a9f952d","shared_citers":2},{"title":"Semi-Supervised Classification with Graph Convolutional Networks","work_id":"21fff118-807d-49cd-8229-f7087ba57b5d","shared_citers":2},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":2},{"title":"Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge","work_id":"28ea1282-d657-4c61-a83c-f1249be6d6b1","shared_citers":2},{"title":"Transformer-xl: Attentive language models beyond a fixed-length context","work_id":"eb970d64-41ff-4e44-afd0-e3bb975e0dc4","shared_citers":2}],"time_series":[{"n":1,"year":2019},{"n":31,"year":2026}],"dependency_candidates":[]},"authors":[{"id":"eb2faec5-1621-4088-bc26-f7fa3009e416","orcid":null,"display_name":"Ben Poole","source":"manual","import_confidence":0.72},{"id":"a54699ed-6350-46a2-922b-41bf21df87fb","orcid":null,"display_name":"Eric Jang","source":"manual","import_confidence":0.72},{"id":"0226913a-35be-4685-b429-9d56ad1a5fe0","orcid":null,"display_name":"Shixiang Gu","source":"manual","import_confidence":0.72}]}}