Evaluating CLIP: Towards Characterization of Broader Capabilities and Downstream Implications

By S. Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, Miles Brundage. 2021

174 citations recorded by the source corpus.

Recently, there have been breakthroughs in computer vision ("CV") models that are more generalizable with the advent of models such as CLIP and ALIGN. In this paper, we analyze CLIP and highlight some of the challenges such models pose. CLIP reduces the need for task specific training data, potentially opening up many niche tasks to automation. CLIP also allows its users to flexibly specify image classification classes in natural language, which we find can shift how biases manifest. Additionally, through some preliminary probes we find that CLIP can inherit biases found in prior computer vision systems. Given the wide and unpredictable domain of uses for such models, this raises questions regarding what sufficiently safe behaviour for such systems may look like. These results add evidence to the growing body of work calling for a change in the notion of a 'better' model--to move beyond simply looking at higher accuracy at task-oriented capability evaluations, and towards a broader 'better' that takes into account deployment-critical features such as different use contexts, and people who interact with the model when thinking about model deployment.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Multimodal Learning, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Multimodal Learning

In Evaluation & Benchmarks

Citations in the corpus

CLOOB: Modern Hopfield Networks with InfoLOOB Outperform CLIP (2021, methodology): “The Country211 (Radford et al., 2021) dataset consists of photos across 211 countries and is designed to test the geolocalization capability of visual representations.”
Text2Mesh: Text-Driven Neural Stylization for Meshes (2021, methodology): “Our framework uses a pre-trained CLIP [44] model, which has been shown to contain bias [1].”
Stable Bias: Evaluating Societal Representations in Diffusion Models (2023, unknown): “Work in the bias and fairness community has also examined how models such as CLIP [49], which are used to guide the diffusion process, encode societal biases between race and cultural identity [1, 67, 68].”
GPT-4V(ision) System Card (2023, background): “…biases such as facial recognition systems having disparate performance based on race, stereotyping members of different races, etc. [13, 14, 15, 16, 17] It is important to note that even with performance parity, differences in downstream impact and harm could still occur depending on the…”
Modular Visual Question Answering via Code Generation (2023, background): “Prior work (Ross et al., 2020; Agarwal et al., 2021) has found evidence of social biases in vision-language models trained on image-captions.”
CLIPScore: A Reference-free Evaluation Metric for Image Captioning (2021, methodology): “…using CLIP-S as an ofﬂine evaluation metric for literal caption quality accords with the recommendations of CLIP’s model card (Mitchell et al., 2019), Agarwal et al. (2021)’s study demonstrates that CLIP can make disproportionate incorrect classiﬁcations of people, e.g., “male images were…”
ReCLIP: A Strong Zero-Shot Baseline for Referring Expression Comprehension (2022, background): “Agarwal et al. (2021) provide evidence that CLIP has racial and other biases, which makes sense since CLIP was trained on data collected from the web and not necessarily curated carefully.”
Open-Vocabulary Image Segmentation (2021, result): “As a result, we share the similar concern as recent studies for evaluating image-text models like Clip [1].”