Hallucinations in Computer Vision

Caption Hallucination Assessment

As previously stated, the researchers devised a new statistic called CHAIR (Caption Hallucination Assessment with Image Relevance). The ratio of words produced by models in the ground truth sentences and object segmentation is calculated here. The metric includes two types: per-instance, which determines what percentage of object instances are hallucinated, and per-sentence, which determines what percentage of phrases contain a hallucinated object.

^source

Around 80 MSCOCO objects were chosen, and each sentence and word were tokenized and singularized. "For each ground truth statement, we determine a list of MSCOCO objects in the same way," the researchers wrote in their article. By merely relying on the specified object labels, the MSCOCO segmentation annotations are used." The researchers discovered that the annotation source is crucial.

They used the sentence annotations for hunting ground truths and investigating the annotations' human biases. As a result, they discovered that employing only segmentation labels or only reference captions causes more hallucination. The researchers also develop an image and linguistic consistency concept. The image consistency section states that the consistent mistakes from captioning models are with a model that predicts objects based on an image alone. The language consistency model states how consistent errors from captioning models are with a model that predicts words based purely on previously created words.

Evaluation

Many baseline models were constructed by the researchers. They've examined a wide range of model designs. They also look at models that have and don't have attention processes. LSTM RNNs are utilized to generate text in the various baseline models. The self-critical (SC) loss and the standard cross-entropy (CE) loss are used by the majority of models to learn. The captioning models are evaluated on two MSCOCO splits, as previously stated.

Conclusion

Finally, the researchers discovered a variety of causes for hallucinations, as well as the fact that hallucinations do not necessarily match the results of typical captioning measures. One of the most noteworthy findings was that neural networks with attention mechanisms have fewer hallucinations, although the convolutional aspects of the systems deserve more credit. They also discovered that having a strong visual representation helps to decrease hallucinations. The researchers said that the design and training of captioning models should be influenced not only by cross-entropy loss or traditional sentence metrics but also by image relevance.

Also read, Sampling and Quantization

FAQs

1. What is a hallucination in machine learning? 

A computer hallucination is an artificial intelligence (AI) machine vision and machine learning technology interpretation error. Computer hallucinations are created by a variety of factors that cause AI systems to misclassify.

2. What is a hallucination in NLP?

Hallucination in NLP is a sensory experience of something that does not exist.
 

3. What is the need for image captioning?

The image caption is a key aspect of scene understanding, which combines computer vision and natural language processing knowledge to automatically generate natural language descriptions based on the content observed in an image.

Key Takeaways

In this article, we have discussed what hallucination is in neural networks and how it occurs.

Want to learn more about Machine Learning? Here is an excellent course that can guide you in learning. 

Happy Coding!