Emergent Communication through Metropolis-Hastings Naming Game with Deep Generative Models

• Refer to the following paper: Advanced Robotics 2023 - Emergent Communication through Metropolis-Hastings Naming Game with Deep Generative Models by Taniguchi Tadahiro, Yuto Yoshida, Akira Taniguchi, Yoshinobu Hagiwara.
• The proposed model in the paper, Inter-GMM+VAE, is implemented in Pytorch.

Symbol Emergence-VAE-GMM (Inter-GMM+VAE)

Naming game with probabilistic inference between agents represented by VAE and GMM.

Model (Inter-GMM-VAE) Overview.
Each agent is represented by a VAE and a GMM.
Agents reason in terms of naming probabilistic inferences based on the Metropolis Hastings algorithm.

---

The objective of the naming game is to match the variables (w^A, w^B) of both agents
Agents play the naming game in the following sequence：

1. Estimation of latent variables (z^A, z^B) by VAE：Agents A and B estimate a latent variable (z^A, z^B) that follows a multivariate normal distribution from image observations. This is done by the VAE within the agent.
2. Agent A is the speaker：Agent A clusters variables (z^A), estimates discrete variables (w^A) following a categorical distribution, and proposes them to Agent B.
3. Agent B is the listener：Agent B decides whether to accept or reject the variable (z^B) proposed by Agent A using the Metropolis Hastings method.
4. Swap the speaker and listener：The Metropolis Hastings algorithm is executed with Agent B as the speaker and Agent A as the listener.
5. Knowledge Update：The agent updates the parameters of the multivariate normal distribution. The updated parameters are then redefined as the parameters of the VAE prior distribution and return to 1.

---

What this repo contains:

• main.py: Main code for training model. Outputs ARI and Kappa coefficients.
• cnn_vae_module_mnist.py: A training program for VAE, running in main.py.
• recall_image.py: Recall the image to the agent.
• tool.py: Various functions handled in the program.

How to run

You can train model by running main.py.

 $ python main.py # Communication (Metropolis Hastings algorithm) 

 $ python main.py --mode 0 # No Communication 

 $ python main.py --mode 1 # All Acceptance 

• Communication：Playing the naming game with probabilistic inference based on the Metropolis-Hastings algorithm. Acceptance and rejection of the speaker's utterance under the Metropolis Hastings algorithm.
• No communication：Both agents do inference independently.
• All Acceptance：Both agents accept all of each other's utterances. No rejection based on Metropolis Hastings method.

Sample output

 $ python3 main.py 

CUDA True
Dataset : MNIST
Total data:10000, Category:10
VAE_iter:50, Batch_size:10
MH_iter:50, MH_mode:-1(-1:Com 0:No-com 1:All accept)
------------------Mutual learning session 0 begins------------------
VAE_Agent A Training Start(0): Epoch:50, Batch_size:10
====> Epoch: 1 Average loss: 168.2284
====> Epoch: 25 Average loss: 98.9265
====> Epoch: 50 Average loss: 92.6812
VAE_Agent B Training Start(0): Epoch:50, Batch_size:10
====> Epoch: 1 Average loss: 160.5718
====> Epoch: 25 Average loss: 98.2610
====> Epoch: 50 Average loss: 91.9621
M-H algorithm Start(0): Epoch:50
=> Epoch: 1, ARI_A: 0.023, ARI_B: 0.015, Kappa:0.422, A2B:5523, B2A:2410
=> Epoch: 10, ARI_A: 0.41, ARI_B: 0.387, Kappa:0.853, A2B:8585, B2A:8743
=> Epoch: 20, ARI_A: 0.665, ARI_B: 0.66, Kappa:0.915, A2B:8990, B2A:9128
=> Epoch: 30, ARI_A: 0.735, ARI_B: 0.736, Kappa:0.933, A2B:9168, B2A:9289
=> Epoch: 40, ARI_A: 0.779, ARI_B: 0.783, Kappa:0.945, A2B:9212, B2A:9284
=> Epoch: 50, ARI_A: 0.803, ARI_B: 0.808, Kappa:0.95, A2B:9278, B2A:9390
Iteration:0 Done:max_ARI_A: 0.803, max_ARI_B: 0.808, max_Kappa:0.95

About the evaluation index:

• ARI_A: ARI of Agent A：The degree of agreement between agent A's sign variable w^A and the true MNIST label．
• ARI_B: ARI of Agent B：The degree of agreement between agent B's sign variable w^B and the true MNIST label．
• Kappa: Kappa coefficients：The degree of agreement of the sine variables w^A and w^B between agents．
• A2B: When speaker A and listener B, the number of times B accepted the sign proposed by A.
• B2A: When speaker B and listener A, the number of times A accepted the sign proposed by B.
  ※ARI tends to be low if batch size is increased too much

Recalled image by Agents

Agents can recall the image after the naming game is over.
Image recall uses the mean parameters estimated by the GMM for the latent variables in the VAE within the agent.
This mean parameter is input to the VAE decoder to generate the image.

After the naming game by main.py is finished, run recall_image.py.
Recall image of Agent A in /model/debug/reconA/：

Recall image of Agent B in /model/debug/reconB/：

Communication between agents shows that the objects in the recalled image are shared.

---

変分オートエンコーダを活用した実画像からの記号創発

VAEとGMMによって表現されるエージェント間の確率的推論によるネーミングゲームにより，両エージェントの記号を共有することを目的とします．

リポジトリのプログラムについて:

• main.py: メインプログラム．エージェント間でネーミングゲームを行います．
• cnn_vae_module_mnist.py: main.py内でVAEの学習を行わせるプログラム．
• recall_image.py: 学習後のエージェントに画像の想起を行わせるプログラム．
• tool.py: 様々な関数が格納されたプログラム.

本リポジトリは2022年度人工知能学会全国大会で発表したものとなります

実行方法

main.pyを実行することで確率的推論によるネーミングゲームを行います．

 $ python main.py # コミュニケーションモデル：メトロポリスヘイスティングス法によるネーミングゲームを行うモデル

 $ python main.py --mode 0 # ノンコミュニケーションモデル：個々のエージェントが独立して推論を行うモデル

 $ python main.py --mode 1 # All Acceptance model：コミュニケーションを行いますが発話者の提案を聞き手が全て受容してしまうモデル

出力例

 $ python3 main.py 

CUDA True
Dataset : MNIST
Total data:10000, Category:10
VAE_iter:50, Batch_size:10
MH_iter:50, MH_mode:-1(-1:Com 0:No-com 1:All accept)
------------------Mutual learning session 0 begins------------------
VAE_Agent A Training Start(0): Epoch:50, Batch_size:10
====> Epoch: 1 Average loss: 168.2284
====> Epoch: 25 Average loss: 98.9265
====> Epoch: 50 Average loss: 92.6812
VAE_Agent B Training Start(0): Epoch:50, Batch_size:10
====> Epoch: 1 Average loss: 160.5718
====> Epoch: 25 Average loss: 98.2610
====> Epoch: 50 Average loss: 91.9621
M-H algorithm Start(0): Epoch:50
=> Epoch: 1, ARI_A: 0.023, ARI_B: 0.015, Kappa:0.422, A2B:5523, B2A:2410
=> Epoch: 10, ARI_A: 0.41, ARI_B: 0.387, Kappa:0.853, A2B:8585, B2A:8743
=> Epoch: 20, ARI_A: 0.665, ARI_B: 0.66, Kappa:0.915, A2B:8990, B2A:9128
=> Epoch: 30, ARI_A: 0.735, ARI_B: 0.736, Kappa:0.933, A2B:9168, B2A:9289
=> Epoch: 40, ARI_A: 0.779, ARI_B: 0.783, Kappa:0.945, A2B:9212, B2A:9284
=> Epoch: 50, ARI_A: 0.803, ARI_B: 0.808, Kappa:0.95, A2B:9278, B2A:9390
Iteration:0 Done:max_ARI_A: 0.803, max_ARI_B: 0.808, max_Kappa:0.95

評価値について:

• ARI_A: エージェントAのARI：エージェントAのサイン変数 w^A と真のMNISTラベルとの一致度合いを表す．
• ARI_B: エージェントBのARI：エージェントBのサイン変数 w^B と真のMNISTラベルとの一致度合いを表す．
• Kappa: カッパ係数：エージェント間のサイン変数 w^A と w^B の一致度合いを表す．
• A2B: 発話者A・聞き手Bのとき，Aが提案したサインをBが受容した回数.
• B2A: 発話者B・聞き手Aのとき，Bが提案したサインをAが受容した回数.

エージェントによる画像の想起

ネーミングゲームの終了後にエージェントは画像の想起を行うことができます.
エージェントによる画像の想起では，エージェント内のVAEが推論した潜在変数に対してGMMが推定した平均パラメータを用います.
この平均パラメータをVAEデコーダに入力することで画像を再構成しエージェントに画像を想起させます.

main.pyによるネーミングゲーム終了後，recall_image.pyを実行してください．

Recall image of Agent A in /model/debug/reconA/：

Recall image of Agent B in /model/debug/reconB/：

コミュニケーションを行ったモデルではエージェントの想起画像が共有されていることがわかります.