Image captioning is a task of generating descriptions of a given image in natural language. In a general encoder-decoder structure (Vinyals et al., 2015)
, image features are encoded in a CNN and decoded into a caption in a word by word manner. Based on the loss function, typical approaches addressing the problem could be divided into three categories: MLE (Maximum Likelihood Estimation), RL (Reinforcement Learning) and GAN (Generative Adversarial Network).
Early proposed methods were based on MLE function and made improvements by designing specific model structure (Xu et al., 2015)
. MLE adopts the cross-entropy loss and learns a one-hot distribution for each word in the sentence. By maximizing the probability of the ground truth word whilst suppressing other reasonable vocabularies, the probability distribution learned by MLE tends to besparse and the generated captions have limited diversity (Dai et al., 2017). On the other hand, RL has advantages in boosting the model performance by optimizing the handcrafted metrics (Rennie et al., 2017; Liu et al., 2017; Chen et al., 2019). However, due to the reward hacking problem, RL maximizes the reward in an unintended way and fails to produce human-like descriptions (Li et al., 2019a). Considering naturalness and diversity of the generated captions, GAN has raised attention in image captioning for its capability of producing descriptions that are indistinguishable from human-written ones (Dai et al., 2017; Shetty et al., 2017; Chen et al., 2019; Dognin et al., 2019). See Figure 1 for a few examples.
In image captioning, the generator of GAN learns true data distribution by maximizing the reward function learned from a discriminator, and the discriminator distinguishes the generated sample from the true data. The adversarial training converges to an equilibrium point (i.e., Nash equilibrium) at which both the generator and discriminator cannot improve (Goodfellow et al., 2014)
. GAN is less biased towards frequently occurring n-grams and learns to describe images with human-like descriptions(Shetty et al., 2017). However, previous work of adversarial networks in image captioning gives one reward function for a complete sentence consisting of words. This strategy causes the reward ambiguity problem (Ng et al., 1999) since which word(s) causes the reward to increase or decrease is not accounted for, and thus there are many optimal policies that determine the sentence can explain one reward. As shown in Figure 2, the generated two captions have the same reward () in GAN, whereas the contribution of each word to this reward remains unknown. The first caption gives the wrong verb “looking at” whilst the second caption has an incorrect object “a plate of food”. However, the ambiguous reward in GAN makes it unable to locate the inappropriate words. On the other hand, from the perspective on the system level, learning sentence-level reward from different image-caption pairs is analogous to learning reward of a trajectory from different system dynamics, which makes the discriminator unable to distinguish the true reward functions from those shaped by the environment dynamics (Fu et al., 2018).
Facing above challenges, we adopt Adversarial Inverse Reinforcement Learning (AIRL) (Fu et al., 2018) to solve the reward ambiguity problem by learning a compact reward function, where compact means the reward function should satisfy two conditions: 1) The reward is word-wise and disentangled for each word in a sentence from different image-caption pairs, as shown in Figure 2. 2) The reward difference of two words is positively correlated to their semantic difference. For instance, words with similar semantics, such as children and kids
, correspond to close reward values. A compact reward function can precisely tell the contribution of each word and thus help to locate the wrong words. It saves the effort of predefining a handcrafted reward function, and can recover the true reward up to a constant at optimality. Driven by such compact reward function from the discriminator, the generator can learn the optimal policy and thus produces qualitative descriptions. However, there are still two major problems to address: 1) AIRL is difficult to converge to Nash equilibrium using policy gradient, requiring Hessian of the gradient vector filed being positive definite(Mescheder and Geiger, 2017). We will discuss this in detail in Section 4.2. 2) As a GAN based method, AIRL has a sharp decision boundary for two disjoint distributions, which means the discriminator can be far more stronger than the generator. The consequence is a limited diversity in the generated captions, which is a commonly encountered issue in GAN called mode collapse (Mirza and Osindero, 2014).
In this paper, we propose a refined AIRL method to learn a compact reward function for each word, as well as achieve stable adversarial training by refining the loss function to shift the generator towards Nash equilibrium. In addition, a conditional term is introduced in the loss function to mitigate mode collapse and to increase the diversity of the generated descriptions. Both the caption evaluator (i.e., discriminator) (Cui et al., 2018; Sharif et al., 2018) and the generator are cast into this unified framework, where the discriminator evaluates captions using a learned compact reward function, and the generator produces qualitative image descriptions. We demonstrate the effectiveness of our method in the experiments.
2 Related Work
The development of image captioning can be summarized into two directions: model structure design (Lu et al., 2017; Yao et al., 2018) and loss function construction (Rennie et al., 2017; Ren et al., 2017). In the methods based on model structure design, attention mechanism and the fusion of visual and semantic information are the key focus. Lu et al. (2017, 2018) proposed a sentinel gate to learn adaptive attention between visual content and non-visual text. Yao et al. (2018) explored the role of visual relationship in image captioning. On the other hand, methods based on loss function construction focus on optimization of the loss function. Rennie et al. (2017)
optimized on non-differentiable evaluation metric using policy gradient, and improved scores of these metrics on various models.Ren et al. (2017) designed an embedding reward under actor-critic reinforcement learning. Similarly, we address the construction of loss functions, and thus our algorithm can be built on existing model structures.
Adversarial Methods for Image Captioning.
Adversarial methods are known for producing plausible samples by training the generator and the discriminator in an adversarial manner (Goodfellow et al., 2014)
. In image captioning, the discriminator is formed as a binary classifier that distinguishes the generated sentence from the ground truth, while the generator produces captions that can fool the discriminator. Conditional GAN was proposed in(Dai et al., 2017) to improve the naturalness and diversity of generated captions. CNN and RNN based discriminators were introduced in (Chen et al., 2019). However, existing methods estimate a reward function for the complete sentence consisting of words, where multiple optimal policies that determine the sentences can correspond to one reward (Ng et al., 1999). Thus the learned reward is ambiguous and ill-defined. We solve this problem by recovering a compact reward function for each word in the sentence under a refined AIRL framework. Although AIRL has been utilized to solve problems in other fields (Wang et al., 2018; Li et al., 2019b; Shi et al., 2018), we are the first to make algorithmic improvements to AIRL such that the model can converge to Nash equilibrium, and that diversity of the outputs can be increased.
3 Adversarial Inverse Reinforcement Learning
Due to the high variance estimate of a full sentence and the reward ambiguity problem, instead of learning reward for a complete sentence, we could learn the reward distributionat time for each word-state pair so that the true reward function can be recovered at optimality (Fu et al., 2018). In the following, we use to represent the word at time , and is the corresponding state vector at time . Note that in an LSTM based model structure, refers to the hidden state of the LSTM cell. In the following, we introduce how AIRL disentangles reward for each word-state pair .
AIRL is an adversarial reward learning algorithm based on Maximum-Entropy-IRL. Finn et al. (2016) first proved that Maximum-Entropy-IRL is mathematically equivalent to GAN under a special form of the discriminator:
where is the data distribution estimated by the discriminator at time , parameterized by . is estimated using the natural exponential function , where is the reward function. is the policy distribution produced by the generator at time , paramterized by . is the generated vocabulary distribution under the context of image captioning. is the decision boundary, which represents the probability that comes from the true word distribution rather than . The discriminator is trained to differentiate between the true words and the generated words whereas the generator tries to fool the discriminator by learning a policy to maximize the reward from .
Considering the reward ambiguity problem, Fu et al. (2018) further extended the above theory to AIRL by introducing a reward shaping term into . The reward shaping term disentangles reward from different system dynamics, which refer to different image-caption pairs under the context of image captioning.
where denotes the reward approximator that recovers the true reward up to a constant, and is the reward shaping term that preserves the optimal . is a constant in range .
algocf[!t] Then the estimated data distribution becomes
For convenience, the decision boundary
can be represented as a sigmoid function:
In the context of divergence minimization, the adversarial process between the discriminator and the generator can be represented as a two-player min-max game (Mescheder and Geiger, 2017):
where is the true word distribution and is the ground truth word sampled from the true data. is the corresponding state of word . In the two-player game, the discriminator maximizes the divergence between the true word distribution and the generated vocabulary distribution, whereas the generator minimizes the divergence. The adversarial training reaches Nash equilibrium when the generated vocabulary distribution approximates the estimated data distribution , i.e., , and both the discriminator and the generator converge. As a result, the discriminator estimates that approximates the true word distribution , and the generator learns an optimal vocabulary distribution that maximizes the reward from .
As mentioned before, a compact reward function can precisely tell the contribution of each word and thus help to locate the wrong words. AIRL can learn a compact reward function at optimality in that 1) it disentangles word-wise reward from different image-caption pairs; 2) the reward difference of two words is positively correlated to their semantic difference if AIRL can recover the true reward for each word. However, AIRL is difficult to converge to Nash equilibrium using policy gradient, requiring Hessian of the gradient vector filed being positive definite (see details in Section 4.2). When the adversarial training of AIRL is not convergent, apparently the true reward can not be recovered. As a result, the learned reward function is not compact. Besides, as a GAN based method, AIRL has a sharp decision boundary for two disjoint distributions, which means the discriminator is much stronger than the generator. The consequence is a limited diversity in the generated captions (see details in Section 4.2), which is called mode collapse in GAN. To solve the two major problems, we explicate in the next section about how we refine the loss function to shift the generator towards Nash equilibrium and to mitigate mode collapse in the two-player game.
4 Learning Compact Reward for Image Captioning
To address the problems discussed above, we refine the loss function to 1) find a compact reward function that can reach its optimum in the adversarial training; 2) increase diversity of the generated captions. In particular, a constant term is used to solve 1) by shifting the generator to Nash equilibrium, and a conditional term is introduced to solve 2) by utilizing mode control techniques. Our algorithm is detailed in Algorithm 1, where is the sentence length and denotes number of iterations.
In the following notations, and are the parameters of the discriminator, and represents the parameter of the generator. and denote the word and its corresponding hidden state vector, respectively. For better clarity, policy is hereinafter referred to as the generated vocabulary distribution
The objective of the discriminator is to distinguish the true caption from the generated one. At time , the discriminator maximizes the divergence in Eq. (5) by
where is the true word and is its corresponding state. The expectation disappears since it is estimated by sampling a mini-batch from the corresponding distribution. is computed as in Eq. (4), where the discriminator learns the reward function for and the generator estimates the vocabulary distribution for , respectively.
Given a word that is sampled from the vocabulary distribution , the generator maximizes by
Using REINFORCE algorithm (Sutton and Barto, 1998), the gradient becomes:
When the decision boundary reaches its optimum () as in Figure 3(a), the generator can only converge when in Eq. (8), requiring Hessian of the gradient vector filed being positive definite (Mescheder and Geiger, 2017). Otherwise, even if the generator has learned the true word distribution (, the non-zero gradient from itself still pushes it away from the true word distribution. Thus the decision boundary can discriminate between the generated vocabulary distribution and the true word distribution in the gray area of Figure 3(b). Then the discriminator tries to maximize the divergence between the two distributions and deviates from the optimum , which leads to the results in Figure 3(c). Such dynamics cause non-convergence in the adversarial training. If the generator converges at instead of in Eq. (8), then its gradient becomes at and the Nash equilibrium in Figure 3(a) can be maintained. Therefore, we introduce a constant term in the expectation in Eq. (7)
Thus, according to Eq. (8), we have
such that the generator converges at the equilibrium point where , i.e., Figure 3(a). It is noted that the constant term can also be regarded as baseline in REINFORCE, except it is utilized to centralize the stationary point instead of reducing variance of the estimation.
The expectation has been removed from the gradient using REINFORCE. Thus the expectation in the loss function disappears by taking the integral of its gradient w.r.t. , for which we have
In practice, the discriminator is usually easier to converge than the generator. If the discriminator converges too early, the generated vocabulary distribution hasn’t approximated the true word distribution, which makes them two disjoint distributions. The gradients of are thus zero almost everywhere (see Figure 4) (Peng et al., 2019), causing limited diversity of the generated captions. The problem is called mode collapse, meaning that the generator produces a single or limited modes. If the generated vocabulary distribution has some overlap with the true word distribution, then the discriminator can not easily differentiate between them, which makes the decision boundary more smooth. Therefore, we introduce ground truth data into the generator as a conditional term (Mirza and Osindero, 2014):
where is the approximated probability of the true word in the generator, and is the conditional term. To give a rough idea, the conditional term smooths the decision boundary by increase the overlap between the true word distributions and the generated vocabulary distribution (see Figure 4) by introducing ground truth data. It strengthens the generator in the following way: 1) When at time , the gradient of the true word becomes larger than that of the generated one (), and thus the generator further increases the probability of the true word (). 2) Otherwise (i.e., ), the generator prefers sampling its self-generated word to confuse the discriminator. By picking the one from the two distributions that can fool the discriminator, the generated vocabulary distribution has more overlap with the true word distribution, which smooths the decision boundary as in Figure 4 and thus helps maintain informative gradient during the adversarial training. The coefficient of is symmetrical to the coefficient of and is updated adaptively in the training process. Note that adding the conditional term does not change the model’s convergence to Nash equilibrium since at the equilibrium.
4.3 Discussion on Loss Functions
We compare the formula of the proposed loss function with existing methods in Table 1, including MLE, RL and GAN. is the length of a sentence. is the handcrafted metric, such as BLEU, CIDEr and SPICE. is the probability of the generated word, and is the probability of the true word. The loss functions are rewritten using similar symbols for direct comparison. MLE maximizes the probability of the true data whist RL and GAN optimize the reward by sampling from . GAN is different from RL in that its reward is learned from the discriminator adversarially instead of being predefined. GAN is capable of mimicking human-written captions by adversarial learning, but the estimated reward function of a full trajectory can be explained by multiple optimal policies and thus is too ambiguous. The proposed rAIRL further disentangles the reward into at each time step , as well as incorporating the true data for better diversity. From the perspective of loss functions, rAIRL can be regarded as an integration of GAN and the first term of MLE using coefficients and .
In the experiments, we validate the effectiveness of the proposed algorithm by answering two questions: 1) Is the caption evaluator (i.e., discriminator) capable of learning compact reward? 2) Driven by the learned reward, is the caption generator effective to produce qualitative captions?
To answer 1), we first tested the compactness of the learned reward by observing performance of the collected top- captions. Then we explored the correlation between the learned reward and the human evaluation results. To answer 2), we built our algorithm on existing learning methods and compared their performance on conventional evaluation metrics. For a comprehensive evaluation, we also evaluated the quality of the generated caption on its content, diversity and grammar. Besides, ablation experiments were conducted to demonstrate the importance of each component of our algorithm.
5.1 Implementation Details
We conducted experiments on the well-known benchmark datasets MS COCO (Chen et al., 2015) and Flickr30K Young et al. (2014), which have 123,287 and 31,783 labeled images, respectively, and each image has at least 5 human annotated captions as reference. We use the public available split Karpathy and Fei-Fei (2015) for Flickr30K. To assess the robustness of our algorithm, we use two splits of the COCO dataset: standard split (Karpathy and Fei-Fei, 2015) which is created by randomly picking test images, and robust split (Lu et al., 2018) which is organized to maximize difference of the co-occurrence distribution between the training and test set. The robust split is recently proposed and is more challenging due to its distribution difference between the training and test set. The standard split has train/val/test images and the robust split has train/val/test images.
We implement our algorithm using Adam optimizer (Kingma and Ba, 2014) with fixed learning rate . Our vocabulary size is for MS COCO and Flickr30K, respectively, including a special start sign <BOS>and an end sign <EOS>. In the generator, the number of hidden nodes of every layer is . For simplicity, the discriminator has the same model structure as the generator except having one additional layer for estimating . For fair comparison, all the methods in ML (Up-Down), RL(Up-Down), GAN(Up-Down), AIRL(Up-Down) and rAIRL(Up-Down) were produced using the same image features and model structure (Up-Down) in (Anderson et al., 2018). Specifically, RL(Up-Down) adopts the self-critical loss in (Rennie et al., 2017). GAN(Up-Down) uses the adversarial loss in (Dai and Lin, 2017) that learns sentence-level reward. AIRL(Up-Down) is the standard adversarial inverse reinforcement learning method in (Fu et al., 2018), and rAIRL(Up-Down) is the proposed method. Note that our scores of MLE(Up-Down) are lower on the standard split but higher on the robust split than (Anderson et al., 2018) because we used fixed number of the bounding box (i.e.,
) for simplicity, and the hyperparameters were tuned to adapt to both splits and thus are not exactly the same with(Anderson et al., 2018).
5.2 Performance of the Recovered Reward
In this section, we evaluated the learned reward function on three aspects: compactness, correlation with human evaluation and performance on diagnosing captions. Compactness shows accuracy of the disentangled reward in measuring word semantic. It is evaluated by computing the correlation between the reward differences and semantic differences after replacing specific words in the caption. Correlation with human evaluation indicates how well the caption evaluator correlates with human judgments. It is based on human scores collected by Aditya et al. (2017). Diagnosing captions using the learned reward function can help improving captions, whose performance is evaluated by the relative improvement after diagnosing and re-written.
|Method||Standard Split||Robust Split|
|RP S||RP D||RP S||RP D|
Compactness means that the reward values should be close for similar words and different for distinct words. For example, kid can also be referred to as little boy or little girl, and thus their reward values should be close to each other in the discriminator. To see the correlation between the reward differences and semantic differences, we replace a given word in the generated caption with a similar word and a distinct word , respectively. Specifically, in a generated caption, the first word that belongs to the COCO class111https://github.com/jiasenlu/NeuralBabyTalk/blob/master/data/coco/coco_class_name.txt (Lu et al., 2018) is replaced. A sentence is discarded if no word can be replaced. The words within the same class are considered to be similar (such as bike and bicycle), and the words that belong to difference classes are distinct (such as man and bike). For RL, since it maximizes a handcrafted reward (SPICE (Anderson et al., 2016) in our experiment) instead of learning a reward function, the reward difference is the variation of SPICE before and after replacement. For reward-learning methods, the reward difference is the variation of the learned reward. The semantic difference is the Euclidean distance between the Glove embedding vectors of two words (Pennington et al., 2014). The results are reported in Table 2. Higher correlation indicates better compactness. RL serves as a baseline in that the handcrafted reward SPICE compares -gram overlapping without considering the semantic difference. The reward differences of rAIRL correlate the best with the semantic differences for both similar words and distinct words, proving the compactness of the learned reward. It’s also noted that due to the reward ambiguity problem, the reward differences of GAN poorly correlate with the semantic differences, especially for similar words.
Figure 5 shows the top- generated captions of an given image. Driven by the learned compact reward function, the top- captions of an image have the same format whilst some words are replaced with their synonyms. For example, the model uses blue and white bus, city bus and public transit bus to describe the bus in the second picture, while the format of its generated sentence remains the same.
Correlation with human evaluation.
As a caption evaluator, the discriminator learns that recovers the true reward up to a constant at optimality (Fu et al., 2018). To explore the correlation between the recovered reward and the human evaluation scores, we used the human scores in the COMPOSITE222https://imagesdg.wordpress.com/image-to-scene-description-graph/ dataset (Aditya et al., 2017), whose images are subsets from Flickr8k, Flickr30k and MS COCO. The descriptions from this dataset are either ground truth captions or generated sentences by (Aditya et al., 2017; Johnson et al., 2015). In the human evaluation process, the AMT worker was asked to give a score at range of 1-5 to evaluate the correctness and throughness of each sentence. Captions with length exceeding were discarded, resulting a total of sentences. Full results of the correlation is shown in Table 3. The correlation is evaluated using Pearson , Kendall’s and Spearman’s correlation coefficients.
In Table 3, the reward of AIRL/rAIRL is the sum of the word-wise reward , and the reward of rAIRL+SPICE is a linear combination of and the SPICE score. Among the reward-learning methods, AIRL poorly correlates with human, whereas the proposed rAIRL improves AIRL on all the correlation metrics, especially on the Pearson correlation (from to ). Furthermore, a simple combination of SPICE and the recovered reward leads to an increased correlation with the human scores, which proves the capacity of the discriminator as a caption evaluator. We also found that conventional metrics, especially BLEU, do not correlate well with human evaluation in terms of linear correlation. Therefore, in the experiments of evaluating the captions in the next section, we directly adopt human studies as the evaluation method, along with other objective evaluation metrics that have proven to correlate well with human, including SPICE (Anderson et al., 2016), CHAIR and CHAIR (Rohrbach et al., 2018). Results on the conventional metrics are also reported for comparison with existing methods, but they are not addressed.
|Learning Method||Model||Standard Split||Robust Split||Flickr30K|
Diagnose and improve captions.
Since the proposed rAIRL learns a word-wise reward, it’s also applicable to diagnose a given caption by finding the wrong word (e.g., the word whose reward decreases sharply compared with that of the previous word) and rewriting the caption to improve its quality. For example, improving a man is playing soccer to a man and a kid are playing soccer can be done by rewriting the caption from is (see Figure 6 for more examples). Therefore, we choose to rewrite a given caption (source caption) from the word whose reward has a decrease rate larger than . However, we found that even rewriting the source caption from a random position using rAIRL can also improve the evaluation scores. Thus, rewriting from a random position is selected as the baseline to compare with rewriting from the located position. Table 4 shows results of rewriting from the located position, where the source captions are given by MLE, RL, GAN and AIRL. Beside each score we report its improvement relative to rewriting from a random position, whose values are mostly positive. This demonstrates that the proposed rAIRL can diagnose the caption at a word level, and further improves the caption quality by rewriting from the located position.
|Method||Standard Split||Robust Split|
5.3 Evaluation on the Generated Captions
In this section, we evaluated the generated captions mainly on three aspects: content correctness, diversity and grammar. Firstly, the results of the caption generator are compared with existing methods on the handcrafted evaluation metrics: BLEU4, CIDEr and SIPCE. However, since BLEU and CIDEr do not correlate well with human (Anderson et al., 2016), we choose other metrics to evaluate the captions in the following experiments. For a comprehensive evaluation, diversity and grammar are also considered as representation of the caption quality. Finally, results of the ablation studies are reported to show importance of each component of our algorithm in caption generation.
Comparison with existing methods.
Categorized by the loss functions, existing models are divided into three categories in Table 5, and we chose recent proposed methods for comparison: Att2in (Rennie et al., 2017), G-GAN (Dai and Lin, 2017), NBT (Lu et al., 2018), Up-Down (Anderson et al., 2018) and GAN, GAN (Dognin et al., 2019). Although some metrics based on -gram overlapping (BLEU4, CIDEr) do not correlate well with human, their results are also reported in Table 5 for fair comparison. Among the adversarial methods (GAN category), our rAIRL performs the best on all metrics.
To test the generalizing ability of our algorithm, we also built our algorithm on the non-adversarial based models. The composite models are denoted with rAIRL +MLE and rAIRL+RL. In rAIRL+MLE, the conditional term is replaced by the cross-entropy loss of MLE; in rAIRL+RL, the RL loss is added into the loss function of the generator. In Table 5, our rAIRL+MLE further improves the MLE baseline (i.e., Up-Down using MLE loss) on SPICE, whereas rAIRL+RL does not improve the RL baseline (i.e., Up-Down using RL loss) on these metrics. This is caused by the difficulty of normalizing the learned reward and the handcrafted reward to the same order of magnitude (Shelton Christian, 2001). Although RL shows better performance on MS COCO by directly optimizing the handcrafted metric (CIDEr), we show in the following experiments that the overall quality of its generated descriptions is not as satisfying as it seems, especially on human evaluations and grammar.
For a comprehensive evaluation of the content correctness, the results of both the handcrafted metrics and human studies are shown in Table 6. For the handcrafted metrics, we report scores of SPICE and the recently proposed CHAIR and CHAIR since they correlate well with human (Anderson et al., 2016; Rohrbach et al., 2018). SPICE computes similarity with the ground truth captions based on scene graph whilst CHIAIR and CHIAR indicate ratio of hallucinated objects. Compared with non-adversarial methods (i.e., MLE, RL), existing adversarial net (GAN) does not perform well on SPICE due to the reward ambiguity problem, whereas our rAIRL improves GAN (from to ) by disentangling reward for each word, and even outperforms RL (from to ) on the robust split. The lowest scores on CHIAIR and CHIAR suggest that object hallucination is less likely in rAIRL.
As for the human evaluation, we randomly selected test images from the standard split and robust split of MS COCO, respectively. The worker was asked “which caption is the best” by given an image with five sentences generated from the adversarial and non-adversarial methods, as shown in Figure 7. The worker was allowed but not encouraged to make multiple choices if he/she thinks these captions are equally correct. The order of captions produced by different methods was randomized. Following (Shetty et al., 2017), each image in the test set was evaluated by workers. Human in Table 6 indicates the percentage of captions that are considered the best among the five methods. The descriptions generated by our rAIRL are considered the best for over images, whilst RL has the lowest scores that are less than . The results of RL on human studies are almost contrary to its performance on the handcrafted metrics in Table 5. This suggests that RL may optimize these metrics in an unintended way such that the scores are improved but the quality of caption is not. On the other hand, by self-learning a reward function, the proposed rAIRL has consistent performance on the human studies and the handcrafted metrics.
|Method||Standard Split||Robust Split|
|Method||Standard Split||Robust Split|
|Method||Standard Split||Robust Split|
The diversity of captions is evaluated on a corpus level, indicating to what extent the generated captions of different images have diverse expressions. The results are presented in Table 7. Vocabulary Coverage is the number of vocabularies of the generated captions over number of vocabularies of the ground truth captions. Novel Sentence indicates the ratio of sentences that do not appear in the training set. The fact that RL has high ratio of novel sentence (/) but low vocabulary coverage (/) suggests that it uses high-frequency words (such as “in a”, “of a”) to reconstruct captions that appear to be different from the training set (Li et al., 2019a). Our rAIRL improves AIRL on the diversity metrics and outperforms other learning methods on vocabulary coverage, indicating its capability of generating diverse descriptions on a corpus level. Figure 8 gives two examples showing diversity of the generated caption. The proposed rAIRL recognizes notable differences between two similar images and give diverse descriptions for each image.
|Method||Standard Split||Robust Split|
|rAIRL(Up-Down, w/o term1)||18.8||10.5||8.2||12.8||73.5||1.07||17.0||19.9||17.5||14.1||71.6||0.95|
|rAIRL(Up-Down, w/o term2)||19.3||9.4||7.4||12.2||71.3||0.83||17.9||18.9||15.8||13.7||62.4||0.72|
We used LanguageTool 333https://languagetool.org to check grammar of the generated captions. Table 8 shows percentage of sentences that have grammar errors found by LanguageTool: 1) Redundancy means repeated phrases in a sentence; 2) Agreement Error means subject-verb agreement error, such as “people is”; 3) Article Misuse denotes inappropriate usage of indefinite articles, such as using “a” before uncountable nouns or plural words; 4) Incomplete Sentence refers to incomplete sentence that lacks a subject. We found captions produced by RL have the most grammar errors ( on the standard split and on the robust split), especially the Article Misuse. On the other hand, by approximating the true data distribution of each word in the sentence, rAIRL and MLE have the least grammar errors among all learning methods ( on the standard split and on the robust split)). We also noticed that each method except rAIRL is biased towards a particular type of grammar error: agreement error in MLE, article misuse in RL, redundancy in GAN, article misuse in AIRL. On both splits, our rAIRL does not appear to be biased towards a specific type of these grammar errors.
Theoretically, our algorithm is model-agnostic since it is independent of the design of model architecture. Therefore, we compare the results of using Att2in (Rennie et al., 2017) and Up-Down (Anderson et al., 2018) model architectures in Table 9, respectively. We report the metrics used in the original paper for fair comparison. The proposed rAIRL mainly improves SPICE on both architectures.
We conducted another ablation experiment to understand the importance of each component of our algorithm in caption generation. Specifically, the constant term in Eq. (9) and the conditional term in Eq. (12) is removed, respectively. Scores of all the evaluation techniques mentioned above are presented in Table 10. All the scores have a drop after removing either one of the terms. Comparing these two terms, the constant term seems more important in recognizing objects and relations in the image since removing it has larger drop on SPICE. The lager drop on vocabulary coverage and ratio of novel sentence in the second row indicates that the conditional term plays a significant role in increasing the diversity of the generated captions.
A few examples of the generated captions produced by different methods are shown in Figures 9 and 10. We compare the captioning results of rAIRL with three other methods: MLE, RL and GAN. Figure 9 gives successful examples, especially on captioning relations between objects. Figure 10 shows failed examples, where objects and relations are not correctly recognized by the captioning model.
Through extensive experiments on caption generation, we proved that the proposed rAIRL constantly performs well on both splits of MS COCO. Compared with RL, rAIRL optimizes the learned reward instead of the handcrafted metrics, and is capable of producing qualitative captions with few grammar errors. As an adversarial algorithm, rAIRL enhances GAN by disentangling compact reward for each word in the caption and improves AIRL by shifting the generator towards Nash equilibrium.
In this paper, we address the reward ambiguity problem in image captioning and propose a refined Adversarial Inverse Reinforcement Learning (rAIRL) method that solves the problem by disentangling reward for each word in a sentence. Moreover, it achieves stable adversarial training by refining the loss function to shift the generator towards Nash equilibrium, and mode control technique is incorporated to mitigate mode collapse. It is demonstrated that our method can learn compact reward through extensive experiments on MS COCO and Flickr30K.
- Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding 173, pp. 33–45. Cited by: §5.2, §5.2.
- SPICE: semantic propositional image caption evaluation. In European conference on computer vision, pp. 382–398. Cited by: §5.2, §5.2, §5.3, §5.3.
Bottom-up and top-down attention for image captioning and visual question answering.
IEEE conference on computer vision and pattern recognition, pp. 6077–6086. Cited by: §5.1, §5.3, §5.3.
Improving image captioning with conditional generative adversarial nets.
AAAI conference on artificial intelligence, pp. 8142–8150. Cited by: §1, §2.
- Microsoft COCO captions: data collection and evaluation server. Cited by: §5.1.
- Learning to evaluate image captioning. In IEEE conference on computer vision and pattern recognition, pp. 5804–5812. Cited by: §1.
- Towards diverse and natural image descriptions via a conditional GAN. In International conference on computer vision, pp. 2989–2998. Cited by: §1, §2.
- Contrastive learning for image captioning. In Advances in neural Information processing systems, pp. 898–907. Cited by: §5.1, §5.3.
- Adversarial semantic alignment for improved image captions. In IEEE conference on computer vision and pattern recognition, pp. 10463–10471. Cited by: §1, §5.3.
A connection between generative adversarial networks, inverse reinforcement learning, and energy-based models. In arXiv preprint arXiv:1611.03852, Cited by: §3.
- Learning robust rewards with adversarial inverse reinforcement learning. In International conference on learning representations, Cited by: §1, §3, §5.1, §5.2.
- Generative adversarial nets. In Advances in neural information processing systems , pp. 2672–2680. Cited by: §1, §2, Figure 3.
- Image retrieval using scene graphs. In IEEE conference on computer vision and pattern recognition, pp. 3668–3678. Cited by: §5.2.
- Deep visual-semantic alignments for generating image descriptions. In IEEE conference on computer vision and pattern recognition, pp. 3128–3137. Cited by: §5.1.
- Adam: a method for stochastic optimization. In International conference on learning representations, Cited by: §5.1.
- Meta learning for image captioning. In AAAI conference on artificial intelligence, pp. 8626–8633. Cited by: §1, §5.3.
Dialogue generation: from imitation learning to inverse reinforcement learning. In AAAI conference on artificial intelligence, pp. 6722–6729. Cited by: §2.
- Improved image captioning via policy gradient optimization of SPIDEr. In International conference on computer vision, pp. 873–881. Cited by: §1.
- Knowing when to look: adaptive attention via a visual sentinel for image captioning. In IEEE conference on computer vision and pattern recognition, pp. 3242–3250. Cited by: §2.
- Neural baby talk. In IEEE conference on computer vision and pattern recognition, pp. 7219–7228. Cited by: §2, §5.1, §5.2, §5.3.
- The numerics of gans. In Advances in neural information processing systems , pp. 1825–1835. Cited by: §1, §3, §4.2.
- Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784. Cited by: §1, §4.2.
Policy invariance under reward transformations: theory and application to reward shaping..
International Conference on Machine Learning, pp. 278–287. Cited by: §1, §2.
- Variational discriminator bottleneck: Improving imitation learning, inverse RL, and GANs by constraining information flow. In International conference on learning representations, Cited by: §4.2.
Glove: global vectors for word representation.
Empirical methods in natural language processing, pp. 1532–1543. Cited by: §5.2.
- Deep reinforcement learning-based image captioning with embedding reward. In IEEE conference on computer vision and pattern recognition, pp. 1151–1159. Cited by: §2.
- Self-critical sequence training for image captioning. In IEEE conference on computer vision and pattern recognition, pp. 1179–1195. Cited by: §1, §2, §5.1, §5.3, §5.3.
- Object hallucination in image captioning. In Empirical methods in natural language processing, pp. 4035–4045. Cited by: §5.2, §5.3.
- Learning-based composite metrics for improved caption evaluation.. In ACL, pp. 14–20. Cited by: §1.
- Balancing multiple sources of reward in reinforcement learning. In Advances in neural information processing systems , pp. 1082–1088. Cited by: §5.3.
- Speaking the same language: matching machine to human captions by adversarial training. In International conference on computer vision, pp. 4155–4164. Cited by: §1, §1, §5.3.
Towards diverse text generation with inverse reinforcement learning. In International joint conference on artficial intelligence, pp. 4361–4367. Cited by: §2.
- Reinforcement learning: An introduction. MIT press Cambridge. Cited by: §4.2.
- Show and tell: a neural image caption generator. In IEEE conference on computer vision and pattern recognition, pp. 3156–3164. Cited by: §1.
- No metrics are perfect: adversarial reward learning for visual storytelling. In ACL, pp. 899–909. Cited by: §2.
- Show, attend and tell: neural image caption generation with visual attention. In International conference on machine learning, pp. 2048–2057. Cited by: §1.
- Exploring visual relationship for image captioning. In European conference on computer vision, pp. 711–727. Cited by: §2.
- From image descriptions to visual denotations: new similarity metrics for semantic inference over event descriptions. In ACL, Cited by: §5.1.